diff --git a/parse/train/0lz69oI5iZP/images/174bccaa8505a067b442ea0dd0c60c559002705bb990065a45c5b35210a720b1.jpg b/parse/train/0lz69oI5iZP/images/174bccaa8505a067b442ea0dd0c60c559002705bb990065a45c5b35210a720b1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8f91423c911c5842bf4a4354d7065dfc348167c9 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/174bccaa8505a067b442ea0dd0c60c559002705bb990065a45c5b35210a720b1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee9bceac2e16c6ac1478edae48fd1ecdab702ea9b1dd710fe79d5e066329ab52 +size 92912 diff --git a/parse/train/0lz69oI5iZP/images/18e39f0d0be8dab4f49c7f6b81d89bf2c303aee95792ceac4ae92dbc76d5d6f1.jpg b/parse/train/0lz69oI5iZP/images/18e39f0d0be8dab4f49c7f6b81d89bf2c303aee95792ceac4ae92dbc76d5d6f1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c0e54bf20bd9d9ea1eebd5a1c7ce44304430eab3 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/18e39f0d0be8dab4f49c7f6b81d89bf2c303aee95792ceac4ae92dbc76d5d6f1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c5267b69aa73063def97601508c70c0d8eec961a7b2af20b122e479151ad2d9 +size 9614 diff --git a/parse/train/0lz69oI5iZP/images/3c723224322082212b38b2906f1fc2118d2308b6aa8a817be3dd1d4e4db92538.jpg b/parse/train/0lz69oI5iZP/images/3c723224322082212b38b2906f1fc2118d2308b6aa8a817be3dd1d4e4db92538.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c00acba394c0d06fcce919a6a679f0980789b832 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/3c723224322082212b38b2906f1fc2118d2308b6aa8a817be3dd1d4e4db92538.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9e3db172fffaeac51728b40e5619e9058cffdc0bcc7d59c1cb8885bff5f16b7 +size 28065 diff --git a/parse/train/0lz69oI5iZP/images/5aff2c131b80e2b82ff62860a36cbceb898b6e8d0abd975382b2ff19c3908fe6.jpg b/parse/train/0lz69oI5iZP/images/5aff2c131b80e2b82ff62860a36cbceb898b6e8d0abd975382b2ff19c3908fe6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..86fc76d6f9c90b53d3798bcf90ce00b58da5f195 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/5aff2c131b80e2b82ff62860a36cbceb898b6e8d0abd975382b2ff19c3908fe6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00fd8a9fd44f1f52fd584afdf6e537f473ef28f484d0e1f38944edf1c6fccdbd +size 8365 diff --git a/parse/train/0lz69oI5iZP/images/6c865ff89e1ebcc297bb1742e7ca52ffcee661225a78ea093cb69da114208a3b.jpg b/parse/train/0lz69oI5iZP/images/6c865ff89e1ebcc297bb1742e7ca52ffcee661225a78ea093cb69da114208a3b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..84e9a691e4026fa6b267d81297ead8a903cfb8e9 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/6c865ff89e1ebcc297bb1742e7ca52ffcee661225a78ea093cb69da114208a3b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8432b54264917de3125c95c26226ee7af6ed74565e9e8967ec5efdf3d1af0ad9 +size 41077 diff --git a/parse/train/0lz69oI5iZP/images/6e7db6f4718dfab5ac8852ebd6bd0bf33faf11664099d470869774f22066b597.jpg b/parse/train/0lz69oI5iZP/images/6e7db6f4718dfab5ac8852ebd6bd0bf33faf11664099d470869774f22066b597.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b336c54f53214f875460bb0bb0deafd8845418c6 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/6e7db6f4718dfab5ac8852ebd6bd0bf33faf11664099d470869774f22066b597.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0ed06e858ca1cdd591509ac472e62cf049aebf5b8f16abdb6b716a11a790bbe +size 73246 diff --git a/parse/train/0lz69oI5iZP/images/730498b062293d8f3037ac0e3381910d3d23eda32f6d04173f2bc62dac34b3a3.jpg b/parse/train/0lz69oI5iZP/images/730498b062293d8f3037ac0e3381910d3d23eda32f6d04173f2bc62dac34b3a3.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0ecb7ebdc106f2082cafc5bc1496e4c1ca340492 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/730498b062293d8f3037ac0e3381910d3d23eda32f6d04173f2bc62dac34b3a3.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70da892b341c5182f7483a4548a521c5c3b4bfc5960516857862e7944a9f750f +size 31591 diff --git a/parse/train/0lz69oI5iZP/images/80d5cbaf1d83eaec5aa6fc1a91131a3b51bc0a3d769578d338e903483120c65b.jpg b/parse/train/0lz69oI5iZP/images/80d5cbaf1d83eaec5aa6fc1a91131a3b51bc0a3d769578d338e903483120c65b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e0d465e51e96827e29ba903cdba7484a526e5dd0 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/80d5cbaf1d83eaec5aa6fc1a91131a3b51bc0a3d769578d338e903483120c65b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5459d0f98300c3c84859ab1300dd360d97b9478da50d21b5cd7333347be0349 +size 60474 diff --git a/parse/train/0lz69oI5iZP/images/ae89dfd9e70e3e0680496b42fc9d272c5369941d71040c6cb96f485bff205fd8.jpg b/parse/train/0lz69oI5iZP/images/ae89dfd9e70e3e0680496b42fc9d272c5369941d71040c6cb96f485bff205fd8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c768de2e51caa8799f2cebb5936830c06614c8a9 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/ae89dfd9e70e3e0680496b42fc9d272c5369941d71040c6cb96f485bff205fd8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b7fcacbfa88cf6b604d6e1f0a2413251deb801bde5f3ca75870dbd107a7df0c +size 46250 diff --git a/parse/train/0lz69oI5iZP/images/b57ad823f8fc97519af6d2b349e90c0278692fb8d303fd19f3e9b2c5cf677508.jpg b/parse/train/0lz69oI5iZP/images/b57ad823f8fc97519af6d2b349e90c0278692fb8d303fd19f3e9b2c5cf677508.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8793ce7e83386ea83052da6c7f3a07583a258660 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/b57ad823f8fc97519af6d2b349e90c0278692fb8d303fd19f3e9b2c5cf677508.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d9477e17ee79e74963fc2cad713965ec803617ea5eef540e4eec1621e865aed +size 24250 diff --git a/parse/train/0lz69oI5iZP/images/bf413290950fed2721c61eef88c71c41bb2a33319997ddede11164d04ee23480.jpg b/parse/train/0lz69oI5iZP/images/bf413290950fed2721c61eef88c71c41bb2a33319997ddede11164d04ee23480.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d3990afdf34a02a4f4e10b2a2abdd970f00707e5 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/bf413290950fed2721c61eef88c71c41bb2a33319997ddede11164d04ee23480.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c17d59f70cf328ba3a46d070ba7934a9ca3a95edbe2425323a2c0fde1eb4c58c +size 18023 diff --git a/parse/train/0lz69oI5iZP/images/eba2ac1841e35a34475de92a35f48c8abec10d0f339482086803e788c26f4f34.jpg b/parse/train/0lz69oI5iZP/images/eba2ac1841e35a34475de92a35f48c8abec10d0f339482086803e788c26f4f34.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b831841ab6ec8dc9b7807d0571498423c8eb35a9 --- /dev/null +++ b/parse/train/0lz69oI5iZP/images/eba2ac1841e35a34475de92a35f48c8abec10d0f339482086803e788c26f4f34.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c11340b8181d8a6f3c31158ad638c334ec6c845ac3028bce8e099df855a7f19 +size 8790 diff --git a/parse/train/B1IzH7cxl/images/086ed829f9d9b2f61b6fa621b7244cb8991d3a562abc2117d2cb784f92238c84.jpg b/parse/train/B1IzH7cxl/images/086ed829f9d9b2f61b6fa621b7244cb8991d3a562abc2117d2cb784f92238c84.jpg new file mode 100644 index 0000000000000000000000000000000000000000..58a82116410d1da8553894b588c24773d80ab54d --- /dev/null +++ b/parse/train/B1IzH7cxl/images/086ed829f9d9b2f61b6fa621b7244cb8991d3a562abc2117d2cb784f92238c84.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:838e4da22cb6dff7055c0aeb49d11ec150996959265f2c889efce09fd3613b6c +size 137881 diff --git a/parse/train/B1IzH7cxl/images/0ddf293af63fe45c278f714e054e435e79db14abfff0dc6c77d2e948f407f094.jpg b/parse/train/B1IzH7cxl/images/0ddf293af63fe45c278f714e054e435e79db14abfff0dc6c77d2e948f407f094.jpg new file mode 100644 index 0000000000000000000000000000000000000000..62f4282d6780498df0fac250b4cf2b078a154508 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/0ddf293af63fe45c278f714e054e435e79db14abfff0dc6c77d2e948f407f094.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30440fe448476b6ab7f74112073f28929a3325ddd2b9e3a35b834425fd512b3a +size 4494 diff --git a/parse/train/B1IzH7cxl/images/0f4267086df87d704e935cf7b8baa2f7349e39f858a7ba0e54f10a6f3ad78001.jpg b/parse/train/B1IzH7cxl/images/0f4267086df87d704e935cf7b8baa2f7349e39f858a7ba0e54f10a6f3ad78001.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c990144cc33f1a1be69dd1fb445b9a1fe77f4c34 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/0f4267086df87d704e935cf7b8baa2f7349e39f858a7ba0e54f10a6f3ad78001.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:849d2a18ede6a45d1f82a915f4c7a1bec6aaf539f333574a5657b222c63c8baf +size 92911 diff --git a/parse/train/B1IzH7cxl/images/1b651a01073939c302698d3354c49a9ab7c7b3941aa09e36a8f018842b56c5a5.jpg b/parse/train/B1IzH7cxl/images/1b651a01073939c302698d3354c49a9ab7c7b3941aa09e36a8f018842b56c5a5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2634c387021e6a98e4dea58d3a81ff2085733ee5 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/1b651a01073939c302698d3354c49a9ab7c7b3941aa09e36a8f018842b56c5a5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:565387d663a4707600f1c95ac69dacfc99aa152019b44747c56344387e8878eb +size 89155 diff --git a/parse/train/B1IzH7cxl/images/221bc6bf416c50c87a1bdc8805959b8f828ce5401c11fd4ce0398c88f6e67ff1.jpg b/parse/train/B1IzH7cxl/images/221bc6bf416c50c87a1bdc8805959b8f828ce5401c11fd4ce0398c88f6e67ff1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7c2e1a60de36a8afe3dfbcb5ff38e535b80662b7 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/221bc6bf416c50c87a1bdc8805959b8f828ce5401c11fd4ce0398c88f6e67ff1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7b5ab96d407d2b52495415725c9c20b9bcf6d992e096776805fd11c2d64c07e +size 18734 diff --git a/parse/train/B1IzH7cxl/images/26676b7f5dd2ffc265d2be16a2b3c493fe30527b31713a247322a40765ef9c68.jpg b/parse/train/B1IzH7cxl/images/26676b7f5dd2ffc265d2be16a2b3c493fe30527b31713a247322a40765ef9c68.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e6550002ab2f63cb43920b853038080f5327d4ab --- /dev/null +++ b/parse/train/B1IzH7cxl/images/26676b7f5dd2ffc265d2be16a2b3c493fe30527b31713a247322a40765ef9c68.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72d12d2ccd247fb3bd9fae355fb945c141396c608a2ce99c9a567b4432eb66d8 +size 10176 diff --git a/parse/train/B1IzH7cxl/images/2bbd46498dcd86730802dfd59bf260c9548b10f7b46d4425482500d659f32112.jpg b/parse/train/B1IzH7cxl/images/2bbd46498dcd86730802dfd59bf260c9548b10f7b46d4425482500d659f32112.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fb4d9367e8a59e07e8c4e0815990c2e3516eda7b --- /dev/null +++ b/parse/train/B1IzH7cxl/images/2bbd46498dcd86730802dfd59bf260c9548b10f7b46d4425482500d659f32112.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dca16ce3e32c88e0f20d15e33914b8113fcb40770fbdc3245d53f0cfab099981 +size 2460 diff --git a/parse/train/B1IzH7cxl/images/4019d43352f97311d7c9fcf91d932c937cf2eef3ef065be2dfb5d2655247be46.jpg b/parse/train/B1IzH7cxl/images/4019d43352f97311d7c9fcf91d932c937cf2eef3ef065be2dfb5d2655247be46.jpg new file mode 100644 index 0000000000000000000000000000000000000000..67e21391d5878785c9ea38c976ba81468cbfccbd --- /dev/null +++ b/parse/train/B1IzH7cxl/images/4019d43352f97311d7c9fcf91d932c937cf2eef3ef065be2dfb5d2655247be46.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43543a1358752113d4e75a77f5d35edfa493b9cf82d8152ce2a0046330c27ee2 +size 93746 diff --git a/parse/train/B1IzH7cxl/images/46f7009a0008fc0bd4fa4ae194e7be22c0741a3aa26e4bb26134c9ca751b6511.jpg b/parse/train/B1IzH7cxl/images/46f7009a0008fc0bd4fa4ae194e7be22c0741a3aa26e4bb26134c9ca751b6511.jpg new file mode 100644 index 0000000000000000000000000000000000000000..86bb663aeac8a36bc20759cc212c406bb4e4443d --- /dev/null +++ b/parse/train/B1IzH7cxl/images/46f7009a0008fc0bd4fa4ae194e7be22c0741a3aa26e4bb26134c9ca751b6511.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4562e485817b19d395f7057eed8da81bdc3c5e8d37615c95f741fba6872c73bf +size 19729 diff --git a/parse/train/B1IzH7cxl/images/4c909370ee63e094a15852b4484907f2ef7f166cbbb5e59bdb94f16e84d6f566.jpg b/parse/train/B1IzH7cxl/images/4c909370ee63e094a15852b4484907f2ef7f166cbbb5e59bdb94f16e84d6f566.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4196e3d71cd9dafd22ab50af20a5ab0917928696 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/4c909370ee63e094a15852b4484907f2ef7f166cbbb5e59bdb94f16e84d6f566.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfe19e42e65242a5eca2fec6951ee040e9a06e647692db329f01e27f9c3a8b46 +size 9677 diff --git a/parse/train/B1IzH7cxl/images/54e32404495620f0fef075866c0409537164d7042eaf71d07f57b3fb673ac4ff.jpg b/parse/train/B1IzH7cxl/images/54e32404495620f0fef075866c0409537164d7042eaf71d07f57b3fb673ac4ff.jpg new file mode 100644 index 0000000000000000000000000000000000000000..00d3de4d41e4d669db8c64fdab3174143f7a70ec --- /dev/null +++ b/parse/train/B1IzH7cxl/images/54e32404495620f0fef075866c0409537164d7042eaf71d07f57b3fb673ac4ff.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbb914735c6534ea9f49e5ebe40a70c6fccc1c259c977d79eab8ffe88cc4689e +size 4620 diff --git a/parse/train/B1IzH7cxl/images/6a2591cbfdc4d495105d2fbc606fee9433066047f64b991a60ca7c8e6649b14d.jpg b/parse/train/B1IzH7cxl/images/6a2591cbfdc4d495105d2fbc606fee9433066047f64b991a60ca7c8e6649b14d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b7fee9128873c4ebb4b8922ef6eb2c036700fc94 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/6a2591cbfdc4d495105d2fbc606fee9433066047f64b991a60ca7c8e6649b14d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7182c9b3cc5f35feae954bab99f3ec053b852f3e94a4e583b27f9176b38f05e +size 9833 diff --git a/parse/train/B1IzH7cxl/images/6c59a9568d5de3de16d22ae651e0b90c39f62322077353e6665f93557889c043.jpg b/parse/train/B1IzH7cxl/images/6c59a9568d5de3de16d22ae651e0b90c39f62322077353e6665f93557889c043.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b7b1994c8d69ca9d2dd378cefd1bce8958e21c9e --- /dev/null +++ b/parse/train/B1IzH7cxl/images/6c59a9568d5de3de16d22ae651e0b90c39f62322077353e6665f93557889c043.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fbfa5181c026a9222d189e989ae8b259214fc01050a85567f4250ab3119801e +size 95622 diff --git a/parse/train/B1IzH7cxl/images/852dde43a7bb5c25766ead25232f56522d381c9a08a4e2765edea88184ee2287.jpg b/parse/train/B1IzH7cxl/images/852dde43a7bb5c25766ead25232f56522d381c9a08a4e2765edea88184ee2287.jpg new file mode 100644 index 0000000000000000000000000000000000000000..118310f17689227eb84187ff71d77d542fb9e194 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/852dde43a7bb5c25766ead25232f56522d381c9a08a4e2765edea88184ee2287.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ec732b55831d17ab2a39d371249316bf74b118921d572dec48c1d9f90e25978 +size 61190 diff --git a/parse/train/B1IzH7cxl/images/93c26443bde1fed1a8cc8c8162d217377c1006d2d059441edb4f984f10ec47ce.jpg b/parse/train/B1IzH7cxl/images/93c26443bde1fed1a8cc8c8162d217377c1006d2d059441edb4f984f10ec47ce.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c8fbad5265ce8ec46848ade6598279c28cea78cb --- /dev/null +++ b/parse/train/B1IzH7cxl/images/93c26443bde1fed1a8cc8c8162d217377c1006d2d059441edb4f984f10ec47ce.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a40e7fe9245bb3ee2caa88258cbeac864545bd79c88eab6be943fa946f8ce667 +size 17760 diff --git a/parse/train/B1IzH7cxl/images/9609a1ca97ff6fb8b13d79e53c0c3cabc4748d700d773ac3f13669153dfe0026.jpg b/parse/train/B1IzH7cxl/images/9609a1ca97ff6fb8b13d79e53c0c3cabc4748d700d773ac3f13669153dfe0026.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5510ab072e5bbffe55ac8a028fd77c74d7c32d48 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/9609a1ca97ff6fb8b13d79e53c0c3cabc4748d700d773ac3f13669153dfe0026.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6dba5b64e2ef2d538a6301a23daae07f30e018aefdfe97a4982452cd262d033 +size 8180 diff --git a/parse/train/B1IzH7cxl/images/963369eaeaa2c273d404f54396d921595f2831b06d7b2d041691259db29b9631.jpg b/parse/train/B1IzH7cxl/images/963369eaeaa2c273d404f54396d921595f2831b06d7b2d041691259db29b9631.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d7177bbfba0b9ec5fa140bdd484cc9a17db167bf --- /dev/null +++ b/parse/train/B1IzH7cxl/images/963369eaeaa2c273d404f54396d921595f2831b06d7b2d041691259db29b9631.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:129f14390dd00bf4059a1e8a168501a9f547e8934eb40c5c86ee4fd3eaa4d62b +size 6968 diff --git a/parse/train/B1IzH7cxl/images/a3f74a20713481241f644c42cbf4b0433d086c05a45c23a0a5a3f2f16a215fef.jpg b/parse/train/B1IzH7cxl/images/a3f74a20713481241f644c42cbf4b0433d086c05a45c23a0a5a3f2f16a215fef.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8d0964b6faba8fd2d8cb49795e0082c4ef79bb95 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/a3f74a20713481241f644c42cbf4b0433d086c05a45c23a0a5a3f2f16a215fef.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df5cbf9bb5f199f8e02661dcc176de627995ab668aaa67f9db41e693543e800b +size 29572 diff --git a/parse/train/B1IzH7cxl/images/aaa3fb36f229f2934bd102fd6afebcd7437777c69d894aead7542c2c4c4d34e3.jpg b/parse/train/B1IzH7cxl/images/aaa3fb36f229f2934bd102fd6afebcd7437777c69d894aead7542c2c4c4d34e3.jpg new file mode 100644 index 0000000000000000000000000000000000000000..503bce1cd8580d904801045c2a1166116d5914fe --- /dev/null +++ b/parse/train/B1IzH7cxl/images/aaa3fb36f229f2934bd102fd6afebcd7437777c69d894aead7542c2c4c4d34e3.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87ceeb355c2eb89de1d738f85516a35369bbc30afc8a12bfa556ba772eb6eeda +size 95549 diff --git a/parse/train/B1IzH7cxl/images/b0fe530281b1acc1169c13d45e343fa8a7c1e248898be60bf2cd72f86cf67b7b.jpg b/parse/train/B1IzH7cxl/images/b0fe530281b1acc1169c13d45e343fa8a7c1e248898be60bf2cd72f86cf67b7b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..06c0ab2ecaefdd87b333d0de9c239703f6004800 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/b0fe530281b1acc1169c13d45e343fa8a7c1e248898be60bf2cd72f86cf67b7b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:770af5bce5456a3833a91cd927d037c289ba141c30d67a6a27afe913e9ab8fa3 +size 100132 diff --git a/parse/train/B1IzH7cxl/images/b60ff8f7498af4964775f859c472e39fce1cb58bed1ce91af895a32d8800903a.jpg b/parse/train/B1IzH7cxl/images/b60ff8f7498af4964775f859c472e39fce1cb58bed1ce91af895a32d8800903a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..08ccaf9f9cab01830dcb62254a1680ec70323c46 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/b60ff8f7498af4964775f859c472e39fce1cb58bed1ce91af895a32d8800903a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04bd799a6e2fc96376181f2cb99b31b15a21b6a558a5dbe8e37f06ba8ad0f8b0 +size 17935 diff --git a/parse/train/B1IzH7cxl/images/b7e2123c9e25b1db35117089c3a606ea2b4e705f3bd6dcbb3fd7978cf7ce2698.jpg b/parse/train/B1IzH7cxl/images/b7e2123c9e25b1db35117089c3a606ea2b4e705f3bd6dcbb3fd7978cf7ce2698.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fca944d1d22541ef658ae13d95885ddc1a409caf --- /dev/null +++ b/parse/train/B1IzH7cxl/images/b7e2123c9e25b1db35117089c3a606ea2b4e705f3bd6dcbb3fd7978cf7ce2698.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0219af622f662918d8af3f5562ff017038c0d419dfe488999edef572e4a5f903 +size 22909 diff --git a/parse/train/B1IzH7cxl/images/bf6f974490e3363d3ab41f3ef8a0b24954a785a0310a64009bb57c1c3d5c2ac6.jpg b/parse/train/B1IzH7cxl/images/bf6f974490e3363d3ab41f3ef8a0b24954a785a0310a64009bb57c1c3d5c2ac6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..aea7e2465a2e57c85547998131fb9fe52595615b --- /dev/null +++ b/parse/train/B1IzH7cxl/images/bf6f974490e3363d3ab41f3ef8a0b24954a785a0310a64009bb57c1c3d5c2ac6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ae4ad186d84381bdd2baa1f52443c866dbed5d12fc727fe4ace94a71b30437f +size 15822 diff --git a/parse/train/B1IzH7cxl/images/d8dfd67798e677bc994cd4c42fba7d3d6a04db75b79a1c7f34f517d1c131b47f.jpg b/parse/train/B1IzH7cxl/images/d8dfd67798e677bc994cd4c42fba7d3d6a04db75b79a1c7f34f517d1c131b47f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9532842b87eceec3ffd8a44fcaeab21e0e1e29e4 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/d8dfd67798e677bc994cd4c42fba7d3d6a04db75b79a1c7f34f517d1c131b47f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00e4951244751c5902940fe3e4629e7efd1d33a842b75af5488f42ed45133faf +size 4448 diff --git a/parse/train/B1IzH7cxl/images/ddc9e5049e321cddf9bd232dc9c1980ef69375de2ae5073ba41fb06eb1c56389.jpg b/parse/train/B1IzH7cxl/images/ddc9e5049e321cddf9bd232dc9c1980ef69375de2ae5073ba41fb06eb1c56389.jpg new file mode 100644 index 0000000000000000000000000000000000000000..61376b4f22e27c00ca024eec02692358091680a9 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/ddc9e5049e321cddf9bd232dc9c1980ef69375de2ae5073ba41fb06eb1c56389.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0684a385dec2cd114ba76c7219b4fbeb86747d8f1479b01b4fd44b8ee76223e0 +size 26018 diff --git a/parse/train/B1IzH7cxl/images/e0558ecd54d5f10e249f4e8a456302324b162ce0bc48f3e0a079f0a62558409c.jpg b/parse/train/B1IzH7cxl/images/e0558ecd54d5f10e249f4e8a456302324b162ce0bc48f3e0a079f0a62558409c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ca00d3805be29c105904dbea1c775ab07e0338a0 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/e0558ecd54d5f10e249f4e8a456302324b162ce0bc48f3e0a079f0a62558409c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0133b652cbda849f24337ea682464386e31f4395139e6873c7d7eb13a9369f4d +size 51476 diff --git a/parse/train/B1IzH7cxl/images/e7e5fc8100ed65282b25d01a72ae783d310027766506e1b65e750c82b5531d37.jpg b/parse/train/B1IzH7cxl/images/e7e5fc8100ed65282b25d01a72ae783d310027766506e1b65e750c82b5531d37.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c10dd759ae2fcd9ec35f3d63238576bc18ceef7e --- /dev/null +++ b/parse/train/B1IzH7cxl/images/e7e5fc8100ed65282b25d01a72ae783d310027766506e1b65e750c82b5531d37.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e4b85ad5b789b05f64b97348bbf71079732e006b48b785c02bb9214831b9136 +size 11891 diff --git a/parse/train/B1IzH7cxl/images/f43ff1bc495c169e73aa911c03a878aa3a393f55577ca834838d3698e86186d6.jpg b/parse/train/B1IzH7cxl/images/f43ff1bc495c169e73aa911c03a878aa3a393f55577ca834838d3698e86186d6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..08828c968a4a780beb1aaa3dcc90ce5bedd22875 --- /dev/null +++ b/parse/train/B1IzH7cxl/images/f43ff1bc495c169e73aa911c03a878aa3a393f55577ca834838d3698e86186d6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c369a2f645cf1cc7aacc4b2e1e89e2fa4a02e806ad534d6fdeca8781fc9f41c +size 88035 diff --git a/parse/train/FsLTUzZlsgT/images/0cae170cea9a9c74311bcdd0da5a169077434377e52599f0e2bd2419a89d9e59.jpg b/parse/train/FsLTUzZlsgT/images/0cae170cea9a9c74311bcdd0da5a169077434377e52599f0e2bd2419a89d9e59.jpg new file mode 100644 index 0000000000000000000000000000000000000000..68dac5aa60293923c5552ba4e8de9ffd4dfa972c --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/0cae170cea9a9c74311bcdd0da5a169077434377e52599f0e2bd2419a89d9e59.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28f4e0d6c396d3d016fb085861d231ae4d1798013fd8f91d5dca228227861415 +size 3510 diff --git a/parse/train/FsLTUzZlsgT/images/109df8a6a3f5df70adf77989c3f32c6fcd9087d7f5f5ce2613423e96d28c19c1.jpg b/parse/train/FsLTUzZlsgT/images/109df8a6a3f5df70adf77989c3f32c6fcd9087d7f5f5ce2613423e96d28c19c1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d99b00e4b83b41bc67f0d39b50947be79d981ffa --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/109df8a6a3f5df70adf77989c3f32c6fcd9087d7f5f5ce2613423e96d28c19c1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b211994165529b9fa0fc9a96a329c1af89d2cf5e58d57abb870ffd716ab77778 +size 4401 diff --git a/parse/train/FsLTUzZlsgT/images/127d6c997d6b33031f3726ecd06e71d71a44e9e9b5345b9a7337ab78e64e30d5.jpg b/parse/train/FsLTUzZlsgT/images/127d6c997d6b33031f3726ecd06e71d71a44e9e9b5345b9a7337ab78e64e30d5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6c6bf88d8fde1918eac0e30807078148c1724a5c --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/127d6c997d6b33031f3726ecd06e71d71a44e9e9b5345b9a7337ab78e64e30d5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:889da59349c9410f4bffd061b651c14be160f75d22fee384bd96d491ba1acd79 +size 43430 diff --git a/parse/train/FsLTUzZlsgT/images/275b006b6297c100d7203fb757b91d6fd4fcc7f5db5e49f97a824a1fb9f48a30.jpg b/parse/train/FsLTUzZlsgT/images/275b006b6297c100d7203fb757b91d6fd4fcc7f5db5e49f97a824a1fb9f48a30.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b94dbb01c13428f853257f1980695752a5d6d4af --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/275b006b6297c100d7203fb757b91d6fd4fcc7f5db5e49f97a824a1fb9f48a30.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdc3cf9a2ba9132eae088a980f2a527a8f8e913b6c3fbf2a015f310ce6ef50f8 +size 37429 diff --git a/parse/train/FsLTUzZlsgT/images/47d77477960d18d6032e3140630d6b3c650ba359e97e9bfaabdce4c1b624e1ec.jpg b/parse/train/FsLTUzZlsgT/images/47d77477960d18d6032e3140630d6b3c650ba359e97e9bfaabdce4c1b624e1ec.jpg new file mode 100644 index 0000000000000000000000000000000000000000..65780591be16be46163ffa5116ca0e86e2cc55ca --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/47d77477960d18d6032e3140630d6b3c650ba359e97e9bfaabdce4c1b624e1ec.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93fbf9e4ed582e7d6d49f18476bf67bf73f240f1022f98f3363a3f4bef85c0b0 +size 23893 diff --git a/parse/train/FsLTUzZlsgT/images/4ceef80b65a6e92513d70ef3b84cdf47bb9c2291c8ba8623844122d702f5bb1b.jpg b/parse/train/FsLTUzZlsgT/images/4ceef80b65a6e92513d70ef3b84cdf47bb9c2291c8ba8623844122d702f5bb1b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4561308fd4fc924aeb8037685afdda9f2cccfb17 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/4ceef80b65a6e92513d70ef3b84cdf47bb9c2291c8ba8623844122d702f5bb1b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13cb8350cb5f42b3fddef41bc41f95f30d2345df64e8ee3d01c41ae1c9d1a2f8 +size 54798 diff --git a/parse/train/FsLTUzZlsgT/images/5415613ac0f90424d681b44f6179e7876e02fbc2098abf408ff8c47b564655aa.jpg b/parse/train/FsLTUzZlsgT/images/5415613ac0f90424d681b44f6179e7876e02fbc2098abf408ff8c47b564655aa.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a7e2104c52c953205b879daa006b6c1af33f07a6 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/5415613ac0f90424d681b44f6179e7876e02fbc2098abf408ff8c47b564655aa.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a025ed8a94aeeebebf192a9ca096debd19a6a4151255ff1c687cb0f26e69a75 +size 2344 diff --git a/parse/train/FsLTUzZlsgT/images/6de27bef9eab970e4e0450d1a438756567355f76ed1bbd59d8e1d198f6aeadb9.jpg b/parse/train/FsLTUzZlsgT/images/6de27bef9eab970e4e0450d1a438756567355f76ed1bbd59d8e1d198f6aeadb9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f36ad4c84754be0a2bfa371651daea8a2ad285e3 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/6de27bef9eab970e4e0450d1a438756567355f76ed1bbd59d8e1d198f6aeadb9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70dbea75d8aacc33b989e5c8ac4ca52d05c1fb936a3aecf38c1ea54785928894 +size 10400 diff --git a/parse/train/FsLTUzZlsgT/images/718bb55f20e691974afe9722e25d167d714b562149d6761633705bb05fe044bf.jpg b/parse/train/FsLTUzZlsgT/images/718bb55f20e691974afe9722e25d167d714b562149d6761633705bb05fe044bf.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2b0e7b3312cf6cc08598fb6db688451d4e4ddbbd --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/718bb55f20e691974afe9722e25d167d714b562149d6761633705bb05fe044bf.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acb28634777177f597308be3566f6c727fa3cb59423920f769ae2a697ff3fa73 +size 3885 diff --git a/parse/train/FsLTUzZlsgT/images/71bc00f37af6f6fadc6bf231cc61a63db5b20fd0f4a4fac621b2dc5afd909dff.jpg b/parse/train/FsLTUzZlsgT/images/71bc00f37af6f6fadc6bf231cc61a63db5b20fd0f4a4fac621b2dc5afd909dff.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b87576921b5c2003fca330919a3352bc4632f580 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/71bc00f37af6f6fadc6bf231cc61a63db5b20fd0f4a4fac621b2dc5afd909dff.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9abeab8a0c9291a63e73dcf91998e5c25c4e8e92a33ffea9a727f771f561012 +size 94864 diff --git a/parse/train/FsLTUzZlsgT/images/89dc0dbfcc7066be63c89da91847b0fd180ba62457c254ed6522bf2367fb6122.jpg b/parse/train/FsLTUzZlsgT/images/89dc0dbfcc7066be63c89da91847b0fd180ba62457c254ed6522bf2367fb6122.jpg new file mode 100644 index 0000000000000000000000000000000000000000..22a74b59c079dda5b8877117bb707665d329e025 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/89dc0dbfcc7066be63c89da91847b0fd180ba62457c254ed6522bf2367fb6122.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8aad1709a3c230bafebb0b3a6511d7e7f2d2f9ddedf5faa2c899f0e87f0c565 +size 22841 diff --git a/parse/train/FsLTUzZlsgT/images/8ab8cbca7e382753c8c7234efbde9aadc1fb7a2db864e76c2178dba160191ffb.jpg b/parse/train/FsLTUzZlsgT/images/8ab8cbca7e382753c8c7234efbde9aadc1fb7a2db864e76c2178dba160191ffb.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4700a18018cbbceaa6039c2555a25a071bc3b0b0 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/8ab8cbca7e382753c8c7234efbde9aadc1fb7a2db864e76c2178dba160191ffb.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7343c50c75e1b7692e6ed85d3201e50e60d74ae960f66840bdc1ea519dc8bb90 +size 3760 diff --git a/parse/train/FsLTUzZlsgT/images/95c163ce7a2adf0dd0e0d1c78fe5f113ee774c59e8eaa74f74df08e4aa2fd338.jpg b/parse/train/FsLTUzZlsgT/images/95c163ce7a2adf0dd0e0d1c78fe5f113ee774c59e8eaa74f74df08e4aa2fd338.jpg new file mode 100644 index 0000000000000000000000000000000000000000..646f15a57d360a47ff9ae22c8a8805ef3822428a --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/95c163ce7a2adf0dd0e0d1c78fe5f113ee774c59e8eaa74f74df08e4aa2fd338.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bab24b9afd4c710c2a4f39635554e8fe78c92d409841d29cb96b0299f9be694d +size 20972 diff --git a/parse/train/FsLTUzZlsgT/images/a1eba89ee7418c92a3e508b0bff964b252a61d8c56cec11d9a53048d21b9bc77.jpg b/parse/train/FsLTUzZlsgT/images/a1eba89ee7418c92a3e508b0bff964b252a61d8c56cec11d9a53048d21b9bc77.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7cf3c23987dcfd9c2979302b42e0ae293c366ae9 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/a1eba89ee7418c92a3e508b0bff964b252a61d8c56cec11d9a53048d21b9bc77.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96c6394a5ffa6f6c702e18eda9de971aae2bb00ede63fe2cc124aaf4031f2a94 +size 60512 diff --git a/parse/train/FsLTUzZlsgT/images/a2dbb4cf3f09eb14ea629c00b0a515e540f2e5d6ab85e3695cb7c41dcb0d1c11.jpg b/parse/train/FsLTUzZlsgT/images/a2dbb4cf3f09eb14ea629c00b0a515e540f2e5d6ab85e3695cb7c41dcb0d1c11.jpg new file mode 100644 index 0000000000000000000000000000000000000000..074c5b7580c1020e21adfc59a96316b9bc06bf18 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/a2dbb4cf3f09eb14ea629c00b0a515e540f2e5d6ab85e3695cb7c41dcb0d1c11.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bec264fb4908f8a194e9d89c1213402696f03857c0e6bf15e239b8e3ecf0d241 +size 3331 diff --git a/parse/train/FsLTUzZlsgT/images/a31c32ca5b343ccd589f5d5f55ff5693deef4785c72b7ddde171ee03ebf39045.jpg b/parse/train/FsLTUzZlsgT/images/a31c32ca5b343ccd589f5d5f55ff5693deef4785c72b7ddde171ee03ebf39045.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e53e9701eca095fbf6586ba1801543e557df6404 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/a31c32ca5b343ccd589f5d5f55ff5693deef4785c72b7ddde171ee03ebf39045.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e53a91fcae01d2d622bb205da736f9b564d7632ae0ca74d88a2bf4baf14dc6e +size 13102 diff --git a/parse/train/FsLTUzZlsgT/images/b331ba9fcbeaca172316403cb0b4ebdfbe11471aa090f32df610a297d3f1fc0d.jpg b/parse/train/FsLTUzZlsgT/images/b331ba9fcbeaca172316403cb0b4ebdfbe11471aa090f32df610a297d3f1fc0d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b13b0909b9461d3e1aaaa283506ad8cd39e9e321 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/b331ba9fcbeaca172316403cb0b4ebdfbe11471aa090f32df610a297d3f1fc0d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32feecc111fd34242d79d120c59146be0a28a21a9f40598ca4e44b0b6ad62887 +size 64906 diff --git a/parse/train/FsLTUzZlsgT/images/bc2900aa2714a04b426a0f668760e8df1aa17303478513a097f17d2937642116.jpg b/parse/train/FsLTUzZlsgT/images/bc2900aa2714a04b426a0f668760e8df1aa17303478513a097f17d2937642116.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2b8ea2bcb9209673dd4a5dc2198a0a31d4e195cf --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/bc2900aa2714a04b426a0f668760e8df1aa17303478513a097f17d2937642116.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f46428fc9892f2858a2dab0f4ab584da000faa27c2ba10f21e167dedb98e045 +size 38234 diff --git a/parse/train/FsLTUzZlsgT/images/e24cb5c4d040a659a68786e3fd48877b4126b0eaaba0feca3d55d6c4809d5c93.jpg b/parse/train/FsLTUzZlsgT/images/e24cb5c4d040a659a68786e3fd48877b4126b0eaaba0feca3d55d6c4809d5c93.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f8682d98a9952c8b9f95e6ea0320b4760fecfefd --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/e24cb5c4d040a659a68786e3fd48877b4126b0eaaba0feca3d55d6c4809d5c93.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1eff001d903a32c12de9dbb3695ea4bfed1aa2dc99c44eacd21c942abbce5771 +size 22588 diff --git a/parse/train/FsLTUzZlsgT/images/e7cb8456771bac77487166284c59f6dc01986402959a0e6c8f11ec5d5a96eb1b.jpg b/parse/train/FsLTUzZlsgT/images/e7cb8456771bac77487166284c59f6dc01986402959a0e6c8f11ec5d5a96eb1b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..292603495e4acde0b407d99fb8ba0f9ba6abdd57 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/e7cb8456771bac77487166284c59f6dc01986402959a0e6c8f11ec5d5a96eb1b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba19a1456228997c07df1cc19b553330d7cc669032cd047f0d81b954a280b9fb +size 6631 diff --git a/parse/train/FsLTUzZlsgT/images/f3a3d16b9f78ef8dded3fb6fd3ca52c43a7c7c25bf77598070329f9c1d97c797.jpg b/parse/train/FsLTUzZlsgT/images/f3a3d16b9f78ef8dded3fb6fd3ca52c43a7c7c25bf77598070329f9c1d97c797.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8a7ed3dd0945a0e92a36beccdf71fce4df4d2551 --- /dev/null +++ b/parse/train/FsLTUzZlsgT/images/f3a3d16b9f78ef8dded3fb6fd3ca52c43a7c7c25bf77598070329f9c1d97c797.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f48c1952c9646521d908a14e47d9eb8e3e6bab13516efe541cb317cacb5222e0 +size 230005 diff --git a/parse/train/HJem3yHKwH/images/0ee838ecb563cab90ca269bab4ce5fc5e86dde4716863af006716b4e1eddd702.jpg b/parse/train/HJem3yHKwH/images/0ee838ecb563cab90ca269bab4ce5fc5e86dde4716863af006716b4e1eddd702.jpg new file mode 100644 index 0000000000000000000000000000000000000000..757faec9273661766796894ec1282d5ccd0f2238 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/0ee838ecb563cab90ca269bab4ce5fc5e86dde4716863af006716b4e1eddd702.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c253f1f87e2f097d48f2b2f75f7fb4317fc97b7992d02318a3b8701c1ad822dd +size 37728 diff --git a/parse/train/HJem3yHKwH/images/18754a4ce117c70babbf960321a959b269b81064e868bb873f8af411c7c5f232.jpg b/parse/train/HJem3yHKwH/images/18754a4ce117c70babbf960321a959b269b81064e868bb873f8af411c7c5f232.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8d08f7ecaf9453827c01ddde4e5fc634e8ae0cec --- /dev/null +++ b/parse/train/HJem3yHKwH/images/18754a4ce117c70babbf960321a959b269b81064e868bb873f8af411c7c5f232.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b4757815f6b7c8cfadbf6699a78322ceb663df87abb281c85335f1a0df9ceca +size 115350 diff --git a/parse/train/HJem3yHKwH/images/21454f0db3d0cd8d81ca85888ad061435440c2c54480229ef00563e27818ce6f.jpg b/parse/train/HJem3yHKwH/images/21454f0db3d0cd8d81ca85888ad061435440c2c54480229ef00563e27818ce6f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..df24caa434859ec9543f26614e4a92a5685187e6 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/21454f0db3d0cd8d81ca85888ad061435440c2c54480229ef00563e27818ce6f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afce69e26cbcb226ba315176496034584472ad4d4293d7c438db4bb3ee758581 +size 19728 diff --git a/parse/train/HJem3yHKwH/images/260a3fc456edea24e53617ad01bd3f8de7823ff4a0a079c675e68c44d7a25e8f.jpg b/parse/train/HJem3yHKwH/images/260a3fc456edea24e53617ad01bd3f8de7823ff4a0a079c675e68c44d7a25e8f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1b84059f99e24c7e235d5522eb47edecf1869fe2 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/260a3fc456edea24e53617ad01bd3f8de7823ff4a0a079c675e68c44d7a25e8f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b69eef8f80fd430283ae45a0288d7c0c22fee7030aad5da6111395c980bb55cb +size 8271 diff --git a/parse/train/HJem3yHKwH/images/33d7667b2ae7aa54f11201a9e0461fb9e4d3c9cbe8c9cbe7211665a96999dfa2.jpg b/parse/train/HJem3yHKwH/images/33d7667b2ae7aa54f11201a9e0461fb9e4d3c9cbe8c9cbe7211665a96999dfa2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..570d3593f05c4f0a80ae8b49da6ecc00a5d95dde --- /dev/null +++ b/parse/train/HJem3yHKwH/images/33d7667b2ae7aa54f11201a9e0461fb9e4d3c9cbe8c9cbe7211665a96999dfa2.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f33b146b1db3079ffadbc32e3939baa1f36570ee71524251b52696527bce8d73 +size 22104 diff --git a/parse/train/HJem3yHKwH/images/3bee91ae0256038f09eca8070d5b300c413bf2734e0dd232fd66edbbc5337e30.jpg b/parse/train/HJem3yHKwH/images/3bee91ae0256038f09eca8070d5b300c413bf2734e0dd232fd66edbbc5337e30.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5a40d4ce542e267240759fbc8d2e26da1b2b2bc9 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/3bee91ae0256038f09eca8070d5b300c413bf2734e0dd232fd66edbbc5337e30.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32b2a19eca18164feb158a5ec6e226804e60b1451cf8b480735cfebb601d4058 +size 13073 diff --git a/parse/train/HJem3yHKwH/images/40e2eb40c0beae6fcd9a409aca3b3033d324a3b512b8c8c47ebacbdc3425b035.jpg b/parse/train/HJem3yHKwH/images/40e2eb40c0beae6fcd9a409aca3b3033d324a3b512b8c8c47ebacbdc3425b035.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5d1e0f9ff9f5123f473d05710ac2b2a60754b0fa --- /dev/null +++ b/parse/train/HJem3yHKwH/images/40e2eb40c0beae6fcd9a409aca3b3033d324a3b512b8c8c47ebacbdc3425b035.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8606042001c1f475256a8b1e23d646d69a3e323c4a205909f644328dd552c468 +size 1937 diff --git a/parse/train/HJem3yHKwH/images/730a1976c9578f0b444868cda424a296e007d5159c5ec5903befea8a25359a4a.jpg b/parse/train/HJem3yHKwH/images/730a1976c9578f0b444868cda424a296e007d5159c5ec5903befea8a25359a4a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..08ae7182be659d5f95063458ca03c2052b8d7b92 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/730a1976c9578f0b444868cda424a296e007d5159c5ec5903befea8a25359a4a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50100a0ef6143eec8b4b731d8431587c4dabddfcfbbe3a4ee73049aaaf6f2fa1 +size 4484 diff --git a/parse/train/HJem3yHKwH/images/7b6948367ed96b0d7589991223f362e5067fd0a284fcbb0ad1724ecae06d5a27.jpg b/parse/train/HJem3yHKwH/images/7b6948367ed96b0d7589991223f362e5067fd0a284fcbb0ad1724ecae06d5a27.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7143d0f87160af10627c86eab860ea15c860dc83 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/7b6948367ed96b0d7589991223f362e5067fd0a284fcbb0ad1724ecae06d5a27.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a83d983c725c3b85c2d7211ce8631b22d9983bd42130688b6b303afdad9f94f2 +size 101346 diff --git a/parse/train/HJem3yHKwH/images/ac5fcadbe231bc4c806345e114719243cf6eac72e7623666b32e62e36d36b113.jpg b/parse/train/HJem3yHKwH/images/ac5fcadbe231bc4c806345e114719243cf6eac72e7623666b32e62e36d36b113.jpg new file mode 100644 index 0000000000000000000000000000000000000000..186a1e133a0f7793accd48212f5206c3848692d2 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/ac5fcadbe231bc4c806345e114719243cf6eac72e7623666b32e62e36d36b113.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:883a20365443851285b0541864906f7de1a5348e1544fce4e7a21d12e10e8c74 +size 5609 diff --git a/parse/train/HJem3yHKwH/images/ba687c90ee0600a7c31af5b8d203808e632b7dba1fa834f3e670ed9f266b6b3f.jpg b/parse/train/HJem3yHKwH/images/ba687c90ee0600a7c31af5b8d203808e632b7dba1fa834f3e670ed9f266b6b3f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4f497679d703a0fb474334de8b76e536c41fe2dd --- /dev/null +++ b/parse/train/HJem3yHKwH/images/ba687c90ee0600a7c31af5b8d203808e632b7dba1fa834f3e670ed9f266b6b3f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be7093c603a1d71cfda6afee1099bc2ad4269a444e08333d71eed1b097ad90db +size 63263 diff --git a/parse/train/HJem3yHKwH/images/ce03aafd71f0a52811b4fc528282178ec148ee8ac35a27a27eef83d49b6e0d47.jpg b/parse/train/HJem3yHKwH/images/ce03aafd71f0a52811b4fc528282178ec148ee8ac35a27a27eef83d49b6e0d47.jpg new file mode 100644 index 0000000000000000000000000000000000000000..52baa77637e3b9e4fd7cf6fb1da72807bd05db32 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/ce03aafd71f0a52811b4fc528282178ec148ee8ac35a27a27eef83d49b6e0d47.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e1db92af112d6964efadd3dc96108aef0668aee2796e077f7f96b90eda4142e +size 45258 diff --git a/parse/train/HJem3yHKwH/images/cf63ab6219d96b7da2c311c167cef1a4de0dd494c65b7930c2460aa0df1e4291.jpg b/parse/train/HJem3yHKwH/images/cf63ab6219d96b7da2c311c167cef1a4de0dd494c65b7930c2460aa0df1e4291.jpg new file mode 100644 index 0000000000000000000000000000000000000000..dd153ddc0bbe9ac03cfffe7d5f3f5f8ef514ca58 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/cf63ab6219d96b7da2c311c167cef1a4de0dd494c65b7930c2460aa0df1e4291.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20465a3bb2ac95486a76cd25c10537010eb56285a164d17cfc4220437f613251 +size 15603 diff --git a/parse/train/HJem3yHKwH/images/cffe734d2de60751977344e955799e230845a671b869abc6ed862c096520b350.jpg b/parse/train/HJem3yHKwH/images/cffe734d2de60751977344e955799e230845a671b869abc6ed862c096520b350.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d40abab3e77820f9563e5fb6bb9650a5cd10bd75 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/cffe734d2de60751977344e955799e230845a671b869abc6ed862c096520b350.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10dee731820148bac59bb2fc8933f03ac5c4d3ceb5340927ddca212aa333b2bf +size 57526 diff --git a/parse/train/HJem3yHKwH/images/fd49e0d1f6a26ce7a897444acd4836b09cb6a0e38269033a9c2fcc1e6377827d.jpg b/parse/train/HJem3yHKwH/images/fd49e0d1f6a26ce7a897444acd4836b09cb6a0e38269033a9c2fcc1e6377827d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..632851bc0da00367b32541e8dddd3ff4562e60e5 --- /dev/null +++ b/parse/train/HJem3yHKwH/images/fd49e0d1f6a26ce7a897444acd4836b09cb6a0e38269033a9c2fcc1e6377827d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d27037754c1162a868d1f6619eef2b3541c06767c2e7283e1791363b88701a6c +size 52940 diff --git a/parse/train/KCzRX9N8BIH/KCzRX9N8BIH.md b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH.md new file mode 100644 index 0000000000000000000000000000000000000000..230e89ebc9f02baf3c791d6f0775cfb0cb177509 --- /dev/null +++ b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH.md @@ -0,0 +1,292 @@ +# IT IS LIKELY THAT YOUR LOSS SHOULD BE A LIKELIHOOD + +Mark Hamilton MIT, Microsoft markth@mit.edu + +Evan Shelhamer MIT, Adobe Research shelhamer@adobe.com + +William Freeman MIT, Google billf@mit.edu + +# ABSTRACT + +Many common loss functions such as mean-squared-error, cross-entropy, and reconstruction loss are unnecessarily rigid. Under a probabilistic interpretation, these common losses correspond to distributions with fixed shapes and scales. We instead argue for optimizing full likelihoods that include parameters like the normal variance and softmax temperature. Joint optimization of these “likelihood parameters” with model parameters can adaptively tune the scales and shapes of losses in addition to the strength of regularization. We explore and systematically evaluate how to parameterize and apply likelihood parameters for robust modeling, outlier-detection, and re-calibration. Additionally, we propose adaptively tuning $L _ { 2 }$ and $L _ { 1 }$ weights by fitting the scale parameters of normal and Laplace priors and introduce more flexible element-wise regularizers. + +# 1 INTRODUCTION + +Choosing the right loss matters. Many common losses arise from likelihoods, such as the squared error loss from the normal distribution , absolute error from the Laplace distribution, and the cross entropy loss from the softmax distribution. The same is true of regularizers, where $L _ { 2 }$ arises from a normal prior and $L _ { 1 }$ from a Laplace prior. + +Deriving losses from likelihoods recasts the problem as a choice of distribution which allows datadependent adaptation. Standard losses and regularizers implicitly fix key distribution parameters, limiting flexibility. For instance, the squared error corresponds to fixing the normal variance at a constant. The full normal likelihood retains its scale parameter and allows optimization over a parametrized set of distributions. This work examines how to jointly optimize distribution and model parameters to select losses and regularizers that encourage generalization, calibration, and robustness to outliers. We explore three key likelihoods: the normal, softmax, and the robust regression likelihood $\rho$ of Barron (2019). Additionally, we cast adaptive priors in the same light and introduce adaptive regularizers. Our contributions: + +1. We systematically survey and evaluate global, data, and predicted likelihood parameters and introduce a new self-tuning variant of the robust adaptive loss $\rho$ +2. We apply likelihood parameters to create new classes of robust models, outlier detectors, and re-calibrators. +3. We propose adaptive versions of $L 1$ and $L 2$ regularization using parameterized normal and Laplace priors on model parameters. + +# 2 BACKGROUND + +Notation We consider a dataset $\mathcal { D }$ of points $x _ { i }$ and targets $y _ { i }$ indexed by $i \in \{ 1 , \ldots , N \}$ . Targets for regression are real numbers and targets for classification are one-hot vectors. The model $f$ with parameters $\theta$ makes predictions $\hat { y } _ { i } = f _ { \theta } ( x )$ . A loss $L ( \hat { y } , y )$ measures the quality of the prediction given the target. To learn model parameters we solve the following loss optimization: + +$$ +\operatorname* { m i n } _ { \theta } \underset { ( x , y ) \sim \mathcal { D } } { \mathbb { E } } L ( \hat { y } = f _ { \theta } ( x ) , y ) +$$ + +![](images/b1ea4f3c37fb9dbc501c429bffa66313c780d218dcb161499e075885f826da71.jpg) +Figure 1: Optimizing likelihood parameters adapts the loss without manual hyperparameter tuning to balance accuracy and certainty. + +A likelihood $\mathcal { L } ( \hat { y } | y , \phi )$ measures the quality of the prediction as a distribution over $\hat { y }$ given the target $y$ and likelihood parameters $\phi$ . We use the negative log-likelihood $\ell$ (NLL), and the likelihood interchangeably since both have the same optima. We define the full likelihood optimization: + +$$ +\operatorname* { m i n } _ { \theta , \phi } \quad \mathbb { E } _ { \mathbf { \Phi } ( x , y ) \sim \mathcal { D } } \ell ( \hat { y } = f _ { \theta } ( x ) | y , \phi ) +$$ + +to jointly learn model and likelihood parameters. “Full” indicates the inclusion of $\phi$ , which controls the distribution and induced NLL loss. We focus on full likelihood optimization in this work. We note that the target, $y$ , is the only supervision needed to optimize model and likelihood parameters, $\theta$ and $\phi$ respectively. Additionally, though the shape and scale varies with $\phi$ , reducing the error ${ \hat { y } } - y$ always reduces the NLL for our distributions. + +Distributions Under Investigation This work considers the normal likelihood with variance $\sigma$ (Bishop et al., 2006; Hastie et al., 2009), the softmax likelihood with temperature $\tau$ (Hinton et al., 2015), and the robust likelihood $\rho$ (Barron, 2019) with shape $\alpha$ and scale $\sigma$ that control the scale and shape of the likelihood. The first two are among the most common losses in machine learning, and the last loss provides an important illustration of a likelihood parameter that affects “shape” instead of “scale”. We note that changing the scale and shape of the likelihood distribution is not “cheating” as there is a trade-off between uncertainty and credit. Figure 1 shows how this trade-off affects the Normal and softmax distributions and their NLLs. + +The normal likelihood has terms for the residual ${ \hat { y } } - y$ and the variance $\sigma$ as + +$$ +\mathcal { N } ( \hat { y } | y , \sigma ) = ( 2 \pi \sigma ^ { 2 } ) ^ { - \frac 1 2 } \exp \left( - \frac 1 2 \frac { ( \hat { y } - y ) ^ { 2 } } { \sigma ^ { 2 } } \right) , +$$ + +with $\sigma \in ( 0 , \infty )$ scaling the distribution. The normal NLL can be written $\begin{array} { r } { \ell _ { N } = \frac { 1 } { 2 \sigma ^ { 2 } } ( \hat { y } - y ) ^ { 2 } + \log \sigma , } \end{array}$ after simplifying and omitting constants that do not affect minimization. We recover the squared error by substituting $\sigma = 1$ . + +The softmax defines a categorical distribution defined by scores $z$ for each class $c$ as + +$$ +\mathrm { s o f t m a x } ( \hat { y } = y | z , \tau ) = \frac { e ^ { z _ { y } \tau } } { \sum _ { c } e ^ { z _ { c } \tau } } , +$$ + +with the temperature, $\tau \in ( 0 , \infty )$ , adjusting the entropy of the distribution. We recover the classification cross-entropy loss, $- \log p ( \hat { y } = y )$ , by substituting $\tau = 1$ in the respective NLL. We state the gradients of these likelihoods with respect to their $\sigma$ and $\tau$ in Section A of the supplement. + +The robust loss $\rho$ and its likelihood are + +$$ +{ \begin{array} { r l } & { \rho \left( x , \alpha , \sigma \right) = { \frac { \left| \alpha - 2 \right| } { \alpha } } \left( \left( { \frac { \left( x / \sigma \right) ^ { 2 } } { \left| \alpha - 2 \right| } } + 1 \right) ^ { \alpha / 2 } - 1 \right) { \mathrm { a n d } } } \\ & { p \left( { \hat { y } } \mid y , \alpha , \sigma \right) = { \frac { 1 } { \sigma Z \left( \alpha \right) } } \exp \left( - \rho \left( { \hat { y } } - y , \alpha , \sigma \right) \right) , } \end{array} } +$$ + +with shape $\alpha \in [ 0 , \infty )$ , scale $\sigma \in ( 0 , \infty )$ , and normalization function $Z \left( \alpha \right)$ . This robust loss, $\rho _ { ; }$ , has the interesting property that it generalizes several different loss functions commonly used in robust learning such as the L2 loss $( \alpha = 2 )$ , pseudo-huber loss (Charbonnier et al., 1997) $( \alpha = 1 )$ , Cauchy loss (Li et al., 2018) $( \alpha = 0 )$ , Geman-McClure loss (Ganan & McClure, 1985), $( \alpha = - 2 )$ , and Welsch (Dennis Jr & Welsch, 1978) loss $( a l p h a = - \infty )$ . Learning the shape parameter allows models to adapt the shape of their noise distribution. + +# 3 RELATED WORK + +Likelihood optimization follows from maximum likelihood estimation (Hastie et al., 2009; Bishop et al., 2006), yet is uncommon in practice for fitting deep regressors and classifiers for discriminative tasks. However Kendall & Gal (2017); Kendall et al. (2018); Barron (2019); Saxena et al. (2019) optimize likelihood parameters to their advantage yet differ in their tasks, likelihoods, and parameterizations. In this work we aim to systematically experiment, clarify usage, and encourage their wider adoption. + +Early work on regressing means and variances (Nix & Weigend, 1994) had the key insight that optimizing the full likelihood can fit these parameters and adapt the loss. Some recent works use likelihoods for loss adaptation, and interpret their parameters as the uncertainty (Kendall & Gal, 2017; Kendall et al., 2018), robustness (Kendall & Gal, 2017; Barron, 2019; Saxena et al., 2019), and curricula (Saxena et al., 2019) of losses. MacKay & Mac Kay (2003) uses Bayesian evidence to select hyper-parameters and losses based on proper likelihood normalization. Barron (2019) define a generalized robust regression loss, $\rho$ , to jointly optimize the type and degree of robustness with global, data-independent, parameters. Kendall & Gal (2017) predict variances for regression and classification to handle data-dependent uncertainty. Kendall et al. (2018) balance multi-task loss weights by optimizing variances for regression and temperatures for classification. These global parameters depend on the task but not the data, and are interpreted as inherent task uncertainty. Saxena et al. (2019) define a differentiable curriculum for classification by assigning each training point its own temperature. These data parameters depend on the index of the data but not its value. We compare these different likelihood parameterizations across tasks and distributions. + +In the calibration literature, Guo et al. (2017) have found that deep networks are often miscalibrated, but they can be re-calibrated by cross-validating the temperature of the softmax. In this work we explore several generalizations of this concept. Alternatively, Platt scaling (Platt, 1999) fits a sigmoid regressor to model predictions to calibrate probabilities. Kuleshov et al. (2018) re-calibrate regressors by fitting an Isotonic regressor to the empirical cumulative distribution function. + +# 4 LIKELIHOOD PARAMETER TYPES + +We explore the space of likelihood parameter representations for model optimization and inference. Though we note that some losses, like adversarial losses, are difficult to represent as likelihoods, many different losses in the community have a natural probabilistic interpretation. Often, these probabilistic interpretations can be parametrized in a variety of ways. We explore two key axes of generality when building these loss functions: conditioning and dimensionality. + +Conditioning We represent the likelihood parameters by three functional classes: global, data, and predicted. Global parameters, $\phi = c$ , are independent of the data and model and define the same likelihood distribution for all points. Data parameters, $\phi _ { i }$ , are conditioned on the index, $i$ , of the data, $x _ { i }$ , but not its value. Every training point is assigned an independent likelihood parameter, $\phi _ { i }$ that define different likelihoods for each training point. Predicted parameters, $\grave { \phi ( x ) = { g _ { \eta } ( x ) } }$ , are determined by a model, $g$ , with parameters $\eta$ (not to be confused with the task model parameters $\theta$ ). Global and predicted parameters can be used during training and testing, but data parameters are only assigned to each training point and are undefined for testing. We show a simple example of predicted temperature in Figure 4, and an illustration of the parameter types in Figure 2. + +We note that for certain global parameters like a learned Normal scale, changing the scale does not affect the optima, but does change the probabilistic interpretation. This invariance has led many authors to drop the scale from their formulations. However, when models can predict these scale parameters they can naturally remain calibrated in the presence of heteroskedasticity and outliers. + +![](images/c49dec68844ff1791035e87e8e2cdb47745b7e10693a93dd6a1ceca2a971d27a.jpg) +Figure 2: Illustration of an image classifier with three different types of likelihood temperature conditioning: global, predicted, and data. Each represents a different way to parametrize the model’s temperature. + +![](images/ad6ea3516753f68625ef40ead2859868d8abf970b4463448775e0270e5989662.jpg) +Figure 3: An image loss function with three different likelihood parameter dimensionalities. Each represents a possible way to parametrize the additional scale parameter added to the loss. + +![](images/6e6527859ba4e3af8468001a873c2fb1f83aa9e57ce95e50bc2c24f811b78155.jpg) +Figure 4: A synthetic logistic regression experiment. Regressing softmax temperature reduces the influence of outliers (blue, bottom-left), by locally raising temperature. The jointly optimized model (center and right panel) achieves a more accurate classification that a model trained without adaptive temperature (left panel). + +Additionally we note that for the shape parameter of the robust likelihood, $\rho$ , changing global parameters does affect model fitting. Previous works have adapted a global softmax temperature for model distillation (Hinton et al., 2015), and recalibration (Guo et al., 2017). Barron (2019) also experiments with global values of loss function shape and scale parameters. The main work on Data parameters is that of Saxena et al. (2019) who use these to learn a curriculum. Model-based parameters appear in earlier work on regressing variance (Nix & Weigend, 1994), and more recent work by Kendall & Gal (2017). + +Dimensionality The dimensionality, $| \phi |$ , of likelihood parameters can vary with the dimension of the task prediction, $\hat { y }$ . For example, image regressors can use a single likelihood parameter for each image $| \phi | = 1$ , RGB image channel $| \phi | = C$ , or even every pixel $| \phi | = \bar { W \times H \times C }$ as in Figure 3. These choices correspond to different likelihood distribution classes. Dimensionality and Conditioning of likelihood parameters can interact. For example, data parameters with $| \phi | \overset { \cdot } { = }$ $W \times H \times C$ would result in $N \times W \times H \times C$ additional parameters, where $N$ is the size of the dataset. This can complicate implementations and slow down optimization due to disk I/O when their size exceeds memory. Table 5 in the appendix contrasts the computational requirements of different likelihood parameter types. The work of Barron (2019) explores both scalar and pixel-wise dimensionalities for his robust loss. + +# 5 APPLICATIONS + +Table 1: MSE, Time, and Memory increase (compared to standard normal likelihood) for reconstruction by variational auto-encoders with different parameterizations of the robust loss, $\rho$ . Predicted likelihood parameters yield more accurate reconstruction models. + +
Param.DimMSETimeMem
Global1x1x1225.81.04×<1KB
Data1x1x1244.22.70×0.6GB
Pred.1x1x1228.51.04×<1MB
GlobalHxWxC231.11.08×<1MB
DataHxWxC252.69.42×4.4GB
Pred.HxWxC222.31.08×<1MB
+ +# 5.1 ROBUSTNESS AND OUTLIER DETECTION + +Data in the wild is noisy, and machine learning methods should be robust to noise, heteroskedasticity, and corruption. Unfortunately, models trained with the standard mean squared error (MSE) loss are highly susceptible to outliers, and cannot naturally handle heteroskedasticity due to this loss’ fixed variance (Huber, 2004). Allowing models to predict and optimize their likelihood parameters allows models to generalize to these more complex settings. More specifically, likelihood parameters naturally transform standard methods such as regressors, classifiers, and manifold learners into robust variants without expensive outer-loop of model fitting such as RANSAC (Fischler & Bolles, 1981) and Theil-Sen (Theil, 1992). Figure 4 demonstrates this effect with a simple classification dataset, and we point readers to Figures 9 of the Supplement for similar examples for regression and manifold learning. + +In certain datasets, even the assumption of Gaussianity is too restrictive and one must consider more robust and long-tailed distributions. This has led many to investigate broader classes of likelihoods such as Generalized Linear Models (GLMs) (Nelder & Wedderburn, 1972) or the more recent general robust loss, $\rho$ , of (Barron, 2019). To systematically explore how likelihood parameter dimension and conditioning affect model robustness and quality, we reproduce Barron (2019)’s variational auto-encoding (Kingma & Ba, 2015) (VAE) experiments on faces from the CelebA dataset (Liu et al., 2015) in Table 1. We explore learned data (Saxena et al., 2019) and model parameters in addition to Barron’s learned global parameters. We also include two natural parameter dimensionalities: a single set of parameters for the whole image, and a set of parameters for each pixel and channel. We find that predicted parameters achieve the best performance while maintaining fast training time and a small memory footprint. We also find that pixel-wise learned parameters correlate with challenging areas of images and we visualize these parameters in Section D of the Appendix. + +This experiment uses a $1 \times 1$ convolution on the last hidden layer of the decoder as a likelihood parameter model and has the same resolution as the output. The low and high dimensional losses use the same convolutional regressor, but the 1 dimensional case averages over pixels. In the high dimensional case, the output has three channels (for RGB), with six channels total for shape and scale regression. We use the same non-linearities to constrain the shape and scale outputs to reasonable ranges as in (Barron, 2019). More specifically, we use an affine sigmoid to keep the shape $\alpha \in [ 0 , 3 ]$ and the softplus to keep scale $c \in [ 1 0 ^ { - } 8 , \infty )$ . Table 1 gives the results of evaluating each method by MSE on the validation set, while training each method with their respective loss parameters. Data parameter optimization uses Tensorflow’s implementation of sparse RMSProp (Tieleman & Hinton, 2012). We also inherit weight decay $\| \phi \| _ { 2 } ^ { 2 }$ , gradient clipping $\bar { \nabla } _ { \phi } / \| \nabla _ { \phi } \| _ { 2 } ^ { 2 }$ , and learning rate scaling $\alpha _ { \phi } = \alpha \cdot m$ for learning rate $\alpha$ and multiplier $m$ from Barron (2019). + +The robustness we see in our VAE experiments stems from the fact that likelihood parameter prediction gives models a direct channel to express their “uncertainty” for each data-point with respect to the task. This allows models to naturally down-weight and clean outliers from the dataset which can improve model robustness. Consequently, one can harness this effect to create outlier detectors from any underlying model architecture by using learned scales or temperatures as an outlier score function. Furthermore, predicted likelihood parameters allow these methods to detect outliers in unseen data. In Figure 5 we show how auditing temperature or noise parameters can help practitioners spot erroneous labels and poor quality examples. In particular, the model-parameterized temperatures of an image classifier (trained using the setup of 5.3) correlates strongly with blurry, dark, and difficult examples on the Street View House Number (SVHN) dataset. We use this approach to create simple outlier detection algorithms by considering deep $\left( \mathrm { A E } { + } \mathrm { S } \right)$ and linear $( \mathrm { P C A } { + } \mathrm { S } )$ auto-encoders (Kramer, 1991) with data-conditioned scale parameters as outlier scores. We evaluate this approach on tabular datasets using deep and linear auto-encoders with model-parameterized scales. In Table 2 we quantitatively demonstrate the quality of these simple likelihood parameter approaches across 22 datasets from the Outlier Detection Datasets (ODDS), a standard outlier detection benchmark (Rayana, 2016). The ODDS benchmark supplies ground truth outlier labels for each dataset, which allows one to treat outlier detection as an unsupervised classification problem. We compare against a variety of established outlier detection approaches included in the pyOD (Zhao et al., 2019) framework including: One-Class SVMs (OCSVM) (Scholkopf et al., 2000), Local Outlier Fraction (LOF) ¨ (Breunig et al., 2000), Angle Based Outlier Detection (ABOD) (Kriegel et al., 2008), Feature Bagging (FB) (Lazarevic & Kumar, 2005), Auto Encoder Distance (AE) (Aggarwal, 2015), K-Nearest Neighbors (KNN) (Ramaswamy et al., 2000; Angiulli & Pizzuti, 2002), Copula Based Outlier Detection (COPOD) (Li et al., 2020), Variational AutoEncoders (VAE) (Kingma & Welling, 2013), Minimum Covariance Determinants with Mahlanohbis Distance (MCD) (Rousseeuw & Driessen, 1999; Hardin & Rocke, 2004), Histogram-based Outlier Scores (HBOS) (Goldstein & Dengel, 2012), Principal Component Analysis (PCA) (Shyu et al., 2003), Isolation Forests (IF) (Liu et al., 2008; 2012), and the Clustering-Based Local Outlier Factor (CBLOF) (He et al., 2003). + +![](images/13cf97ec3dd147a060f3c2aae3cb2a76c02361a9cc6b8cfbd652566d92745bec.jpg) +Figure 5: The data with the lowest (top) and highest (bottom) predicted temperatures in the SVHN dataset. High temperature entries are blurry, cropped poorly, and generally difficult to classify. + +Table 2: Median outlier detection performance of several methods across 22 benchmark datasets from ODDS. + +
MethodMedian AUC
LOF.669
FB.702
ABOD.727
AE.737
VAE.792
COPOD.799
PCA.808
OCSVM.814
MCD.820
KNN.822
HBOS.822
IF.823
CBLOF.836
AE+S (Ours).846
PCA+S (Ours).868
+ +![](images/e93d5f5fd0e2d1653d0cf1d7ea585fa9b470765348bd2fcdb1405cb7f0328d0e.jpg) +Figure 6: Distribution of Outlier Detection AUC across the ODDS Benchmark. Our approaches, $\mathrm { P C A } { + } \mathrm { S }$ and $_ \mathrm { A E + S }$ , are competitive with other Outlier Detection systems. + +Our predicted scale auto-encoders use PyTorch’s layers API (Paszke et al., 2019) with rectified linear unit (ReLU) activations for deep auto-encoders and Glorot uniform initialization (Dahl et al., 2013; Glorot & Bengio, 2010) for all layers. We use Adam (Kingma & Ba, 2015) with a learning rate of .0005 for 4000 steps with $20 \%$ dropout before the code space. We follow ODDS guidelines and standard scale the data prior to fitting. + +![](images/14ccfaff7d7b98120fd5a934955e0726811cf1fc2af4f91e79fd5ad741b81a41.jpg) +Figure 7: Performance of $L 2$ (left) and $L 1$ (middle) regularized linear regression on a 500 dimensional synthetic dataset where the true parameters, $w ^ { * }$ , are known. Dynamic Ridge (D-Ridge) and D-LASSO regression find the regularization strength that best estimates the true parameters. M-LASSO outperforms any single global regularization strength and does not shrink informative weights. (right) Performance of adaptive $L 1$ regularization methods as a function of true model sparsity. In all cases, Multi-LASSO outperforms other methods by orders of magnitude. + +Our methods $\mathrm { P C A } { + } \mathrm { S }$ and ${ \mathrm { A E } } { + } S _ { \cdot }$ ) use a similar principle as isolation-based approaches that determine outliers based on how difficult they are to model. In existing approaches, outliers influence and skew the isolation model which causes the model to exhibit less confidence on the whole. This hurts a model’s ability to distinguish between inliers and outliers. In contrast, our approach allows the underlying model to down-weight outliers. This yields a more consistent model with a clearer decision boundary between outliers and inliers as shown in Figure 4. As a future direction of investigation we note that our approach is model-architecture agnostic, and can be combined with domain-specific architectures to create outlier detection methods tailored to images, text, and audio. + +# 5.2 ADAPTIVE REGULARIZATION WITH PRIOR PARAMETERS + +In addition to optimizing the shape and scale of the likelihood distribution of the model output, we can use the same approach to optimize the prior distribution of the model parameters. More specifically, we propose adaptive regularizers for a model’s parameters, $\theta$ . This approach optimizes the distribution parameters of the prior, $\phi _ { \mathrm { p r i o r } }$ , to naturally tune the degree of regularization. In particular, the Normal (Ridge, L2) and Laplace (LASSO, L1) priors, with scale parameters $\sigma$ and $b$ , regularize model parameters for small magnitude and sparsity respectively (Hastie et al., 2009). The degree of regularization, $\lambda \in [ 0 , \infty )$ , is conventionally a hyperparameter of the regularized loss function: + +$$ +\operatorname* { m i n } _ { \theta } \sum _ { i } ^ { N } ( \hat { y _ { i } } : = f _ { \theta } ( x _ { i } ) - y _ { i } ) ^ { 2 } + \lambda \sum _ { j } ^ { P } | \theta _ { j } | . +$$ + +We note that we cannot choose $\lambda$ by direct minimization because it admits a trivial minimum at $\lambda = 0$ . In the linear case, one can select this weight efficiently using Least Angle Regression (Efron et al., 2004). However, in general $\lambda$ is usually learned through expensive cross validation methods. Instead, we retain the prior with its scale parameter, and jointly optimize over the full likelihood: + +$$ +\operatorname* { m i n } _ { \theta , \sigma , b } \sum _ { i } ^ { N } \left( \frac { 1 } { 2 \sigma ^ { 2 } } ( \hat { y } _ { i } - y _ { i } ) ^ { 2 } + \log \sigma \right) + \sum _ { j } ^ { P } \left( \frac { | \theta _ { j } | } { b } + \log b \right) +$$ + +This approach, the Dynamic Lasso (D-LASSO), admits no trivial solution for the prior parameter $b$ , and must balance the effective regularization strength, $\frac { 1 } { b }$ , with the normalization factor, $\log b$ . D-LASSO selects the degree of regularization by gradient descent, rather than expensive black-box search. In Figure 7 (left) and (middle) we show that this approach, and its Ridge equivalent, yield ideal settings of the regularization strength on a suite of synthetic regression problems. Figure 7 (right) shows D-LASSO converges to the best LASSO regularization strength for a variety of truemodel sparsities. As a further extension, we replace the global $\sigma$ or $b$ with a $\sigma _ { j }$ or $b _ { j }$ for each model parameter, $\theta _ { j }$ , to locally adapt regularization to each model weight (Multi-Lasso). This consistently outperforms any global setting of the regularization strength and shields important weights from undue shrinkage 7 (middle). For our experiments we use 500 samples of 500 dimensional normal distributions mapped through linear functions with additive gaussian noise. Linear transformations use Uniform $[ 1 , 2 ]$ weights and LASSO experiments use sparse transformations. We use tensorflow’s Adam optimizer with $l r = . 0 0 0 5$ for 100000 steps. + +Our approach of learning regularizer scale parameters can be viewed naturally through the lens of hierarchical priors (Gelman et al., 2013). More specifically this approach is implicitly performing maximum a posteriori (MAP) inference on the prior’s scale with respect to a uniform prior on that parameter. We note that though these methods for hyperparameter selection are common in the Bayesian literature, they are not widely used in practice in the deep learning community. This work aims to bring these parameters back within the scope of deep learning where they can be easily expanded to more flexible forms such as our introduced Multi-Lasso. + +# 5.3 RE-CALIBRATION + +The work of (Guo et al., 2017) shows that modern networks are accurate, yet systematically overconfident, a phenomenon called mis-calibration. We investigate the role of optimizing likelihood parameters to re-calibrate models. More specifically, we can fit likelihood parameter regressors on a validation set to modify an existing model’s confidence to better align with the validation set. This approach is a generalization of Guo et al. (2017)’s Temperature Scaling method, which we refer to as Global Scaling (GS) for notational consistency. Global Scaling re-calibrates classifiers with a learned global parameter, $\tau$ in the loss function: $\sigma ( \vec { x } , \tau )$ . + +Fitting model-conditioned likelihood parameters to a validation set defines a broad class of recalibration strategies. From these we introduce three new re-calibration methods. Linear Scaling (LS) learns a linear mapping, $l$ , to transform logits to a softmax temperature: $\sigma ( \vec { x } , l ( \vec { x } ) )$ . Linear Feature Scaling (LFS) learns a linear mapping, $l$ , to transform the features prior to the logits, $\bar { f }$ , to a softmax temperature: $\sigma ( \vec { x } , l ( \vec { f } ) )$ . Finally, we introduce Deep Scaling (DS) for regressors which learns a nonlinear network, $N$ , to transform features, $\bar { f }$ , into a temperature: $\sigma ( \vec { x } , N ( \vec { f } ) )$ . + +In Table 3 we compare our recalibration approaches to the previous state of the art: Global Scaling. We note that (Guo et al., 2017) have already shown that Global Scaling outperform Bayesian Binning into Quantiles (Naeini et al., 2015), Histogram binning (Zadrozny & Elkan, 2001), and Isotonic Regression. We recalibrate both ResNet50 (He et al., 2016) and DenseNet121 (Huang et al., 2017) on a variety of vision datasets. We measure classifier miscalibration using the Expected Calibration Error (ECE) (Guo et al., 2017) to align with prior art. We additionally evaluate Isotonic recalibration, Platt Scaling (Platt, 1999), and Vector Scaling (VS) (Guo et al., 2017), which learns a vector, $\vec { v }$ , to re-weight logits: $\sigma ( \vec { v } \vec { x } , 1 )$ . LS and LFS tend to outperform other approaches like GS and VS, which demonstrates that richer likelihood parametrizations can improve calibration akin to how richer models can improve prediction. + +Our experiments leverage Tensorflow’s Dataset APIs that include the SVHN, (Netzer et al., 2011), ImageNet (Deng et al., 2009), CIFAR-100, CIFAR-10 (Krizhevsky, 2009) datasets. We use Keras implementations of DenseNet-121 (Huang et al., 2017) and ResNet-50 (He et al., 2016) with default initializations. For optimization we use Adam with $l r = 0 . 0 0 0 1$ , $\beta _ { 1 } = . 9 , \beta _ { 2 } = . 9 9$ (Kingma & Ba, 2015) and train for 300 epoch with a batch size of 512. + +For recalibrating regressors, we compare against the previous state of the art, Kuleshov et al. (2018), who use an Isotonic regressor to correct a regressors’ confidence. We use the same experimental setting as Kuleshov et al. (2018) including the UCI datasets (Dua & Graff, 2017), and regressor calibration metric (CAL). Table 4 shows that our approaches can outperform this baseline as well as the regression equivalent of Global Scaling. Inputs and targets are scaled to unit norm and variance prior to fitting for all regression experiments and missing values are imputed using scikit-learn’s “SimpleImputer” (Pedregosa et al., 2011). Experiments utilize Keras’ layers API with two hidden rectified linear unit (ReLU) layers, Glorot uniform initialization (Dahl et al., 2013; Glorot & Bengio, 2010) and Adam optimization with $l r = 0 . 0 0 1$ for 3000 steps without minibatching. + +Table 3: Comparison of calibration methods by ECE for ResNet-50 (RN50) and DenseNet-121 (DN121) architectures on test data. Our predicted likelihood parameter methods: Linear Scaling (LS) and Linear Feature Scaling (LFS) outperform other approaches. In all cases our methods reduce miscalibration with comparable computation time as GS. + +
ModelDatasetUncalibratedPlattIsotonicGSVSLSLFS
RN50CIFAR-10.250.034.053.046.037.018.018
RN50CIFAR-100.642.061.072.035.044.030.173
RN50SVHN.072.053.010.029.022.009.009
RN50ImageNet.430.018.070.019.023.026.015
DN121CIFAR-10.253.048.042.039.034.028.028
DN121CIFAR-100.537.049.067.024.024.014.031
DN121SVHN.079.018.010.022.017.011.010
DN121ImageNet.229.028.095.021.019.043.019
+ +Table 4: Comparison of regression calibration methods as evaluated by their calibration error as defined in (Kuleshov et al., 2018). Predicted likelihood parameters often outperform other methods. + +
DatasetUncalibrated IsotonicGSLSDS
crime0.36240.34990.0693 0.0125 0.0310
kinematics0.01640.01030.0022 0.0021 0.0032
bank0.01220.00560.0027 0.0024 0.0020
wine0.00910.01080.0152 0.01310.0064
mpg0.21530.22000.1964 0.14830.0233
cpu0.08620.03400.3018 0.20780.1740
soil0.30830.30000.3130 0.3175 0.3137
fried0.00060.00020.00020.00020.0002
+ +# 6 EXPERIMENTAL DETAILS + +We run all experiments on Ubuntu 16.04 Azure Standard NV24 virtual machines (24 CPUs, 224 Gb memory, and $4 \times \mathrm { M } 6 0 \ : \mathrm { G P U s } )$ with Tensorflow 1.15 (Abadi et al., 2015) and PyTorch 1.17 (Paszke et al., 2019). Many likelihood parameters have constrained domains, such as the normal variance $\sigma \in [ 0 , \infty )$ . To evade the complexity of constrained optimization, we define unconstrained parameters $\phi _ { u }$ and choose a transformation $t ( \cdot )$ with inverse $t ^ { - 1 } ( \cdot )$ to map to and from the constrained $\phi$ . For positivity, exp/log parameterization is standard (Kendall & Gal, 2017; Kendall et al., 2018; Saxena et al., 2019). However, this parameterization can lead to instabilities and we use the softplus, $s ^ { + } ( x ) = \log ( 1 + \exp ( x ) )$ , instead. Shifting the softplus, $s _ { c } ^ { + } ( x ) = ( l n ( 1 + e ^ { x } ) + c ) / ( l n ( 2 ) + c ) ,$ , further improves stability and we explore this effect in Figure 12 of the Appendix. We use an affine softplus $s _ { . 0 1 } ^ { \mp }$ and $s _ { . 2 } ^ { + }$ respectively for adaptive scales and temperatures respectively. The one exception is adaptive regularizer scales where we found exp led to faster convergence. For the constrained interval $[ a , b ]$ we use affine transformations of the sigmoid $\begin{array} { r } { s ( x ) = \frac { \tilde { 1 } } { 1 + \exp ( - x ) } } \end{array}$ (Barron, 2019). We initialize likelihood parameter biases to settings that yield MSE and Cross Entropy $\sigma = \tau = 1 \mathrm { { } } $ . + +# 7 CONCLUSION + +Optimizing the full likelihood can improve model quality by adapting losses and regularizers. Full likelihoods are agnostic to the architecture, optimizer, and task, which makes them simple substitutes for standard losses. Global, data, and predicted likelihood parameters offer different degrees of expressivity and efficiency. In particular, predicted parameters adapt the likelihood to each data point during training and testing without significant time and space overhead. By including these parameters in a loss function one can improve a model’s robustness and generalization ability and create new classes of outlier detectors and recalibrators that outperform baselines. More generally, we hope this work encourages joint optimization of model and likelihood parameters, and argue it is likely that your loss should be a likelihood. + +# REFERENCES + +Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´ Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´ berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org. + +Charu C Aggarwal. Outlier analysis. In Data mining, pp. 237–263. Springer, 2015. + +Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European conference on principles of data mining and knowledge discovery, pp. 15–27. Springer, 2002. + +Jonathan T Barron. A general and adaptive robust loss function. In CVPR, 2019. + +Christopher M Bishop. Mixture density networks. 1994. + +Christopher M Bishop et al. Pattern recognition and machine learning, volume 4. springer New York, 2006. + +Markus M Breunig, Hans-Peter Kriegel, Raymond T $\mathrm { N g }$ , and Jorg Sander. Lof: identifying density- ¨ based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pp. 93–104, 2000. + +Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Deterministic edge- ´ preserving regularization in computed imaging. IEEE Transactions on image processing, 6(2): 298–311, 1997. + +George E Dahl, Tara N Sainath, and Geoffrey E Hinton. Improving deep neural networks for lvcsr using rectified linear units and dropout. In ICASSP, pp. 8609–8613. IEEE, 2013. + +Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pp. 248–255. Ieee, 2009. + +John E Dennis Jr and Roy E Welsch. Techniques for nonlinear least squares and robust regression. Communications in Statistics-simulation and Computation, 7(4):345–359, 1978. + +Dheeru Dua and Casey Graff. UCI machine learning repository, 2017. URL http://archive. ics.uci.edu/ml. + +Bradley Efron, Trevor Hastie, Iain Johnstone, Robert Tibshirani, et al. Least angle regression. The Annals of statistics, 32(2):407–499, 2004. + +Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24 (6):381–395, 1981. + +Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission tomography. Amer. Statist. Assoc, pp. 12–18, 1985. + +Andrew Gelman, John B Carlin, Hal S Stern, David B Dunson, Aki Vehtari, and Donald B Rubin. Bayesian data analysis. CRC press, 2013. + +Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, pp. 249–256, 2010. + +Markus Goldstein and Andreas Dengel. Histogram-based outlier score (hbos): A fast unsupervised anomaly detection algorithm. KI-2012: Poster and Demo Track, pp. 59–63, 2012. + +Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. + +Johanna Hardin and David M Rocke. Outlier detection in the multiple cluster setting using the minimum covariance determinant estimator. Computational Statistics & Data Analysis, 44(4): 625–638, 2004. + +Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning: data mining, inference, and prediction. Springer Science & Business Media, 2009. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. + +Zengyou He, Xiaofei Xu, and Shengchun Deng. Discovering cluster-based local outliers. Pattern Recognition Letters, 24(9-10):1641–1650, 2003. + +Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. + +Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In CVPR, 2017. + +Peter J Huber. Robust statistics, volume 523. John Wiley & Sons, 2004. + +Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, pp. 5574–5584, 2017. + +Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pp. 7482–7491, 2018. + +Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. + +Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. + +Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks. AIChE journal, 37(2):233–243, 1991. + +Hans-Peter Kriegel, Matthias Schubert, and Arthur Zimek. Angle-based outlier detection in highdimensional data. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 444–452, 2008. + +Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. + +Volodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated regression. In Jennifer Dy and Andreas Krause (eds.), ICML, volume 80 of Proceedings of Machine Learning Research, pp. 2801–2809, Stockholmsmassan, Stockholm ¨ Sweden, 10–15 Jul 2018. PMLR. + +Aleksandar Lazarevic and Vipin Kumar. Feature bagging for outlier detection. In Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, pp. 157–166, 2005. + +Xuelong Li, Quanmao Lu, Yongsheng Dong, and Dacheng Tao. Robust subspace clustering by cauchy loss function. IEEE transactions on neural networks and learning systems, 30(7):2067– 2078, 2018. + +Zheng Li, Yue Zhao, Nicola Botta, Cezar Ionescu, and Xiyang Hu. Copod: copula-based outlier detection. arXiv preprint arXiv:2009.09463, 2020. + +Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 Eighth IEEE International Conference on Data Mining, pp. 413–422. IEEE, 2008. + +Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation-based anomaly detection. ACM Transactions on Knowledge Discovery from Data (TKDD), 6(1):1–39, 2012. + +Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In ICCV, December 2015. + +David JC MacKay and David JC Mac Kay. Information theory, inference and learning algorithms. Cambridge university press, 2003. + +Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. + +John Ashworth Nelder and Robert WM Wedderburn. Generalized linear models. Journal of the Royal Statistical Society: Series A (General), 135(3):370–384, 1972. + +Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. + +David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In ICNN, volume 1, pp. 55–60. IEEE, 1994. + +Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´ R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 8024–8035. Curran Associates, Inc., 2019. + +Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier ¨ Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011. + +John C Platt. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61–74, 1999. + +Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. Efficient algorithms for mining outliers from large data sets. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pp. 427–438, 2000. + +Shebuti Rayana. Odds library. Stony Brook,-2016, (2017), 2016. + +Peter J Rousseeuw and Katrien Van Driessen. A fast algorithm for the minimum covariance determinant estimator. Technometrics, 41(3):212–223, 1999. + +Shreyas Saxena, Oncel Tuzel, and Dennis DeCoste. Data parameters: A new family of parameters for learning a differentiable curriculum. In NeurIPS, pp. 11093–11103, 2019. + +Bernhard Scholkopf, Robert C Williamson, Alex J Smola, John Shawe-Taylor, and John C Platt.¨ Support vector method for novelty detection. In Advances in neural information processing systems, pp. 582–588, 2000. + +Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and LiWu Chang. A novel anomaly detection scheme based on principal component classifier. Technical report, MIAMI UNIV CORAL GABLES FL DEPT OF ELECTRICAL AND COMPUTER ENGINEERING, 2003. + +Henri Theil. A rank-invariant method of linear and polynomial regression analysis. In Henri Theil’s contributions to economics and econometrics, pp. 345–381. Springer, 1992. + +Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26– 31, 2012. + +Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML, volume 1, pp. 609–616. Citeseer, 2001. + +Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detection. Journal of Machine Learning Research, 20(96):1–7, 2019. URL http://jmlr.org/ papers/v20/19-011.html. \ No newline at end of file diff --git a/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_content_list.json b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..26d8158aec335891f61f76534e0ab0f083c1f2a7 --- /dev/null +++ b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_content_list.json @@ -0,0 +1,1551 @@ +[ + { + "type": "text", + "text": "IT IS LIKELY THAT YOUR LOSS SHOULD BE A LIKELIHOOD ", + "text_level": 1, + "bbox": [ + 174, + 98, + 555, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mark Hamilton MIT, Microsoft markth@mit.edu ", + "bbox": [ + 183, + 170, + 321, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Evan Shelhamer MIT, Adobe Research shelhamer@adobe.com ", + "bbox": [ + 385, + 170, + 575, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "William Freeman MIT, Google billf@mit.edu ", + "bbox": [ + 637, + 170, + 766, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 248, + 544, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many common loss functions such as mean-squared-error, cross-entropy, and reconstruction loss are unnecessarily rigid. Under a probabilistic interpretation, these common losses correspond to distributions with fixed shapes and scales. We instead argue for optimizing full likelihoods that include parameters like the normal variance and softmax temperature. Joint optimization of these “likelihood parameters” with model parameters can adaptively tune the scales and shapes of losses in addition to the strength of regularization. We explore and systematically evaluate how to parameterize and apply likelihood parameters for robust modeling, outlier-detection, and re-calibration. Additionally, we propose adaptively tuning $L _ { 2 }$ and $L _ { 1 }$ weights by fitting the scale parameters of normal and Laplace priors and introduce more flexible element-wise regularizers. ", + "bbox": [ + 233, + 281, + 764, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 463, + 336, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Choosing the right loss matters. Many common losses arise from likelihoods, such as the squared error loss from the normal distribution , absolute error from the Laplace distribution, and the cross entropy loss from the softmax distribution. The same is true of regularizers, where $L _ { 2 }$ arises from a normal prior and $L _ { 1 }$ from a Laplace prior. ", + "bbox": [ + 174, + 496, + 825, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deriving losses from likelihoods recasts the problem as a choice of distribution which allows datadependent adaptation. Standard losses and regularizers implicitly fix key distribution parameters, limiting flexibility. For instance, the squared error corresponds to fixing the normal variance at a constant. The full normal likelihood retains its scale parameter and allows optimization over a parametrized set of distributions. This work examines how to jointly optimize distribution and model parameters to select losses and regularizers that encourage generalization, calibration, and robustness to outliers. We explore three key likelihoods: the normal, softmax, and the robust regression likelihood $\\rho$ of Barron (2019). Additionally, we cast adaptive priors in the same light and introduce adaptive regularizers. Our contributions: ", + "bbox": [ + 174, + 559, + 825, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. We systematically survey and evaluate global, data, and predicted likelihood parameters and introduce a new self-tuning variant of the robust adaptive loss $\\rho$ \n2. We apply likelihood parameters to create new classes of robust models, outlier detectors, and re-calibrators. \n3. We propose adaptive versions of $L 1$ and $L 2$ regularization using parameterized normal and Laplace priors on model parameters. ", + "bbox": [ + 212, + 698, + 825, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 BACKGROUND ", + "text_level": 1, + "bbox": [ + 174, + 804, + 326, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Notation We consider a dataset $\\mathcal { D }$ of points $x _ { i }$ and targets $y _ { i }$ indexed by $i \\in \\{ 1 , \\ldots , N \\}$ . Targets for regression are real numbers and targets for classification are one-hot vectors. The model $f$ with parameters $\\theta$ makes predictions $\\hat { y } _ { i } = f _ { \\theta } ( x )$ . A loss $L ( \\hat { y } , y )$ measures the quality of the prediction given the target. To learn model parameters we solve the following loss optimization: ", + "bbox": [ + 174, + 837, + 823, + 893 + ], + "page_idx": 0 + }, + { + "type": "equation", + "img_path": "images/feb3b89e30bd398c8e0cee30d0790e7945f21c973bbfebdfe555fe9624bfe8f7.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta } \\underset { ( x , y ) \\sim \\mathcal { D } } { \\mathbb { E } } L ( \\hat { y } = f _ { \\theta } ( x ) , y )\n$$", + "text_format": "latex", + "bbox": [ + 401, + 901, + 596, + 928 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1ea4f3c37fb9dbc501c429bffa66313c780d218dcb161499e075885f826da71.jpg", + "image_caption": [ + "Figure 1: Optimizing likelihood parameters adapts the loss without manual hyperparameter tuning to balance accuracy and certainty. " + ], + "image_footnote": [], + "bbox": [ + 179, + 104, + 815, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A likelihood $\\mathcal { L } ( \\hat { y } | y , \\phi )$ measures the quality of the prediction as a distribution over $\\hat { y }$ given the target $y$ and likelihood parameters $\\phi$ . We use the negative log-likelihood $\\ell$ (NLL), and the likelihood interchangeably since both have the same optima. We define the full likelihood optimization: ", + "bbox": [ + 173, + 310, + 825, + 353 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/6b7c1f81e8f579b80c819e0155429a965350c54767ecabb1a4d0fead8f057d8a.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta , \\phi } \\quad \\mathbb { E } _ { \\mathbf { \\Phi } ( x , y ) \\sim \\mathcal { D } } \\ell ( \\hat { y } = f _ { \\theta } ( x ) | y , \\phi )\n$$", + "text_format": "latex", + "bbox": [ + 398, + 359, + 602, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to jointly learn model and likelihood parameters. “Full” indicates the inclusion of $\\phi$ , which controls the distribution and induced NLL loss. We focus on full likelihood optimization in this work. We note that the target, $y$ , is the only supervision needed to optimize model and likelihood parameters, $\\theta$ and $\\phi$ respectively. Additionally, though the shape and scale varies with $\\phi$ , reducing the error ${ \\hat { y } } - y$ always reduces the NLL for our distributions. ", + "bbox": [ + 173, + 392, + 825, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Distributions Under Investigation This work considers the normal likelihood with variance $\\sigma$ (Bishop et al., 2006; Hastie et al., 2009), the softmax likelihood with temperature $\\tau$ (Hinton et al., 2015), and the robust likelihood $\\rho$ (Barron, 2019) with shape $\\alpha$ and scale $\\sigma$ that control the scale and shape of the likelihood. The first two are among the most common losses in machine learning, and the last loss provides an important illustration of a likelihood parameter that affects “shape” instead of “scale”. We note that changing the scale and shape of the likelihood distribution is not “cheating” as there is a trade-off between uncertainty and credit. Figure 1 shows how this trade-off affects the Normal and softmax distributions and their NLLs. ", + "bbox": [ + 173, + 469, + 825, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The normal likelihood has terms for the residual ${ \\hat { y } } - y$ and the variance $\\sigma$ as ", + "bbox": [ + 173, + 590, + 676, + 606 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/3c46149cbf1a7efde4f1f949b21aafb63604e38b2e4c71197321aa8be8dab306.jpg", + "text": "$$\n\\mathcal { N } ( \\hat { y } | y , \\sigma ) = ( 2 \\pi \\sigma ^ { 2 } ) ^ { - \\frac 1 2 } \\exp \\left( - \\frac 1 2 \\frac { ( \\hat { y } - y ) ^ { 2 } } { \\sigma ^ { 2 } } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 343, + 613, + 651, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with $\\sigma \\in ( 0 , \\infty )$ scaling the distribution. The normal NLL can be written $\\begin{array} { r } { \\ell _ { N } = \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } - y ) ^ { 2 } + \\log \\sigma , } \\end{array}$ after simplifying and omitting constants that do not affect minimization. We recover the squared error by substituting $\\sigma = 1$ . ", + "bbox": [ + 174, + 656, + 825, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The softmax defines a categorical distribution defined by scores $z$ for each class $c$ as ", + "bbox": [ + 173, + 705, + 723, + 720 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/972e7ded4f3db2c759825046dfc1de10135f752823d6c1e5f59c9f683b2cf5ad.jpg", + "text": "$$\n\\mathrm { s o f t m a x } ( \\hat { y } = y | z , \\tau ) = \\frac { e ^ { z _ { y } \\tau } } { \\sum _ { c } e ^ { z _ { c } \\tau } } ,\n$$", + "text_format": "latex", + "bbox": [ + 385, + 727, + 611, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with the temperature, $\\tau \\in ( 0 , \\infty )$ , adjusting the entropy of the distribution. We recover the classification cross-entropy loss, $- \\log p ( \\hat { y } = y )$ , by substituting $\\tau = 1$ in the respective NLL. We state the gradients of these likelihoods with respect to their $\\sigma$ and $\\tau$ in Section A of the supplement. ", + "bbox": [ + 174, + 767, + 825, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The robust loss $\\rho$ and its likelihood are ", + "bbox": [ + 176, + 818, + 429, + 832 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/06652b49497df2b81c83be6eceb21ed82ceeec1780ea4c0aaed390fe4ed31d95.jpg", + "text": "$$\n{ \\begin{array} { r l } & { \\rho \\left( x , \\alpha , \\sigma \\right) = { \\frac { \\left| \\alpha - 2 \\right| } { \\alpha } } \\left( \\left( { \\frac { \\left( x / \\sigma \\right) ^ { 2 } } { \\left| \\alpha - 2 \\right| } } + 1 \\right) ^ { \\alpha / 2 } - 1 \\right) { \\mathrm { a n d } } } \\\\ & { p \\left( { \\hat { y } } \\mid y , \\alpha , \\sigma \\right) = { \\frac { 1 } { \\sigma Z \\left( \\alpha \\right) } } \\exp \\left( - \\rho \\left( { \\hat { y } } - y , \\alpha , \\sigma \\right) \\right) , } \\end{array} }\n$$", + "text_format": "latex", + "bbox": [ + 312, + 838, + 687, + 922 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with shape $\\alpha \\in [ 0 , \\infty )$ , scale $\\sigma \\in ( 0 , \\infty )$ , and normalization function $Z \\left( \\alpha \\right)$ . This robust loss, $\\rho _ { ; }$ , has the interesting property that it generalizes several different loss functions commonly used in robust learning such as the L2 loss $( \\alpha = 2 )$ , pseudo-huber loss (Charbonnier et al., 1997) $( \\alpha = 1 )$ , Cauchy loss (Li et al., 2018) $( \\alpha = 0 )$ , Geman-McClure loss (Ganan & McClure, 1985), $( \\alpha = - 2 )$ , and Welsch (Dennis Jr & Welsch, 1978) loss $( a l p h a = - \\infty )$ . Learning the shape parameter allows models to adapt the shape of their noise distribution. ", + "bbox": [ + 174, + 102, + 825, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 214, + 344, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Likelihood optimization follows from maximum likelihood estimation (Hastie et al., 2009; Bishop et al., 2006), yet is uncommon in practice for fitting deep regressors and classifiers for discriminative tasks. However Kendall & Gal (2017); Kendall et al. (2018); Barron (2019); Saxena et al. (2019) optimize likelihood parameters to their advantage yet differ in their tasks, likelihoods, and parameterizations. In this work we aim to systematically experiment, clarify usage, and encourage their wider adoption. ", + "bbox": [ + 174, + 246, + 825, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early work on regressing means and variances (Nix & Weigend, 1994) had the key insight that optimizing the full likelihood can fit these parameters and adapt the loss. Some recent works use likelihoods for loss adaptation, and interpret their parameters as the uncertainty (Kendall & Gal, 2017; Kendall et al., 2018), robustness (Kendall & Gal, 2017; Barron, 2019; Saxena et al., 2019), and curricula (Saxena et al., 2019) of losses. MacKay & Mac Kay (2003) uses Bayesian evidence to select hyper-parameters and losses based on proper likelihood normalization. Barron (2019) define a generalized robust regression loss, $\\rho$ , to jointly optimize the type and degree of robustness with global, data-independent, parameters. Kendall & Gal (2017) predict variances for regression and classification to handle data-dependent uncertainty. Kendall et al. (2018) balance multi-task loss weights by optimizing variances for regression and temperatures for classification. These global parameters depend on the task but not the data, and are interpreted as inherent task uncertainty. Saxena et al. (2019) define a differentiable curriculum for classification by assigning each training point its own temperature. These data parameters depend on the index of the data but not its value. We compare these different likelihood parameterizations across tasks and distributions. ", + "bbox": [ + 174, + 337, + 825, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the calibration literature, Guo et al. (2017) have found that deep networks are often miscalibrated, but they can be re-calibrated by cross-validating the temperature of the softmax. In this work we explore several generalizations of this concept. Alternatively, Platt scaling (Platt, 1999) fits a sigmoid regressor to model predictions to calibrate probabilities. Kuleshov et al. (2018) re-calibrate regressors by fitting an Isotonic regressor to the empirical cumulative distribution function. ", + "bbox": [ + 174, + 537, + 825, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 LIKELIHOOD PARAMETER TYPES ", + "text_level": 1, + "bbox": [ + 176, + 627, + 482, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We explore the space of likelihood parameter representations for model optimization and inference. Though we note that some losses, like adversarial losses, are difficult to represent as likelihoods, many different losses in the community have a natural probabilistic interpretation. Often, these probabilistic interpretations can be parametrized in a variety of ways. We explore two key axes of generality when building these loss functions: conditioning and dimensionality. ", + "bbox": [ + 174, + 659, + 825, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conditioning We represent the likelihood parameters by three functional classes: global, data, and predicted. Global parameters, $\\phi = c$ , are independent of the data and model and define the same likelihood distribution for all points. Data parameters, $\\phi _ { i }$ , are conditioned on the index, $i$ , of the data, $x _ { i }$ , but not its value. Every training point is assigned an independent likelihood parameter, $\\phi _ { i }$ that define different likelihoods for each training point. Predicted parameters, $\\grave { \\phi ( x ) = { g _ { \\eta } ( x ) } }$ , are determined by a model, $g$ , with parameters $\\eta$ (not to be confused with the task model parameters $\\theta$ ). Global and predicted parameters can be used during training and testing, but data parameters are only assigned to each training point and are undefined for testing. We show a simple example of predicted temperature in Figure 4, and an illustration of the parameter types in Figure 2. ", + "bbox": [ + 174, + 734, + 825, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We note that for certain global parameters like a learned Normal scale, changing the scale does not affect the optima, but does change the probabilistic interpretation. This invariance has led many authors to drop the scale from their formulations. However, when models can predict these scale parameters they can naturally remain calibrated in the presence of heteroskedasticity and outliers. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c49dec68844ff1791035e87e8e2cdb47745b7e10693a93dd6a1ceca2a971d27a.jpg", + "image_caption": [ + "Figure 2: Illustration of an image classifier with three different types of likelihood temperature conditioning: global, predicted, and data. Each represents a different way to parametrize the model’s temperature. " + ], + "image_footnote": [], + "bbox": [ + 183, + 104, + 488, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ad6ea3516753f68625ef40ead2859868d8abf970b4463448775e0270e5989662.jpg", + "image_caption": [ + "Figure 3: An image loss function with three different likelihood parameter dimensionalities. Each represents a possible way to parametrize the additional scale parameter added to the loss. " + ], + "image_footnote": [], + "bbox": [ + 504, + 106, + 813, + 204 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6e6527859ba4e3af8468001a873c2fb1f83aa9e57ce95e50bc2c24f811b78155.jpg", + "image_caption": [ + "Figure 4: A synthetic logistic regression experiment. Regressing softmax temperature reduces the influence of outliers (blue, bottom-left), by locally raising temperature. The jointly optimized model (center and right panel) achieves a more accurate classification that a model trained without adaptive temperature (left panel). " + ], + "image_footnote": [], + "bbox": [ + 174, + 296, + 825, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally we note that for the shape parameter of the robust likelihood, $\\rho$ , changing global parameters does affect model fitting. Previous works have adapted a global softmax temperature for model distillation (Hinton et al., 2015), and recalibration (Guo et al., 2017). Barron (2019) also experiments with global values of loss function shape and scale parameters. The main work on Data parameters is that of Saxena et al. (2019) who use these to learn a curriculum. Model-based parameters appear in earlier work on regressing variance (Nix & Weigend, 1994), and more recent work by Kendall & Gal (2017). ", + "bbox": [ + 173, + 554, + 825, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dimensionality The dimensionality, $| \\phi |$ , of likelihood parameters can vary with the dimension of the task prediction, $\\hat { y }$ . For example, image regressors can use a single likelihood parameter for each image $| \\phi | = 1$ , RGB image channel $| \\phi | = C$ , or even every pixel $| \\phi | = \\bar { W \\times H \\times C }$ as in Figure 3. These choices correspond to different likelihood distribution classes. Dimensionality and Conditioning of likelihood parameters can interact. For example, data parameters with $| \\phi | \\overset { \\cdot } { = }$ $W \\times H \\times C$ would result in $N \\times W \\times H \\times C$ additional parameters, where $N$ is the size of the dataset. This can complicate implementations and slow down optimization due to disk I/O when their size exceeds memory. Table 5 in the appendix contrasts the computational requirements of different likelihood parameter types. The work of Barron (2019) explores both scalar and pixel-wise dimensionalities for his robust loss. ", + "bbox": [ + 173, + 666, + 826, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 APPLICATIONS ", + "text_level": 1, + "bbox": [ + 176, + 829, + 330, + 844 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3097495452486ac35715d02502296356e6c54163934240ac695a578863e89140.jpg", + "table_caption": [ + "Table 1: MSE, Time, and Memory increase (compared to standard normal likelihood) for reconstruction by variational auto-encoders with different parameterizations of the robust loss, $\\rho$ . Predicted likelihood parameters yield more accurate reconstruction models. " + ], + "table_footnote": [], + "table_body": "
Param.DimMSETimeMem
Global1x1x1225.81.04×<1KB
Data1x1x1244.22.70×0.6GB
Pred.1x1x1228.51.04×<1MB
GlobalHxWxC231.11.08×<1MB
DataHxWxC252.69.42×4.4GB
Pred.HxWxC222.31.08×<1MB
", + "bbox": [ + 330, + 154, + 665, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 ROBUSTNESS AND OUTLIER DETECTION ", + "text_level": 1, + "bbox": [ + 176, + 281, + 496, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data in the wild is noisy, and machine learning methods should be robust to noise, heteroskedasticity, and corruption. Unfortunately, models trained with the standard mean squared error (MSE) loss are highly susceptible to outliers, and cannot naturally handle heteroskedasticity due to this loss’ fixed variance (Huber, 2004). Allowing models to predict and optimize their likelihood parameters allows models to generalize to these more complex settings. More specifically, likelihood parameters naturally transform standard methods such as regressors, classifiers, and manifold learners into robust variants without expensive outer-loop of model fitting such as RANSAC (Fischler & Bolles, 1981) and Theil-Sen (Theil, 1992). Figure 4 demonstrates this effect with a simple classification dataset, and we point readers to Figures 9 of the Supplement for similar examples for regression and manifold learning. ", + "bbox": [ + 173, + 306, + 825, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In certain datasets, even the assumption of Gaussianity is too restrictive and one must consider more robust and long-tailed distributions. This has led many to investigate broader classes of likelihoods such as Generalized Linear Models (GLMs) (Nelder & Wedderburn, 1972) or the more recent general robust loss, $\\rho$ , of (Barron, 2019). To systematically explore how likelihood parameter dimension and conditioning affect model robustness and quality, we reproduce Barron (2019)’s variational auto-encoding (Kingma & Ba, 2015) (VAE) experiments on faces from the CelebA dataset (Liu et al., 2015) in Table 1. We explore learned data (Saxena et al., 2019) and model parameters in addition to Barron’s learned global parameters. We also include two natural parameter dimensionalities: a single set of parameters for the whole image, and a set of parameters for each pixel and channel. We find that predicted parameters achieve the best performance while maintaining fast training time and a small memory footprint. We also find that pixel-wise learned parameters correlate with challenging areas of images and we visualize these parameters in Section D of the Appendix. ", + "bbox": [ + 173, + 453, + 825, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This experiment uses a $1 \\times 1$ convolution on the last hidden layer of the decoder as a likelihood parameter model and has the same resolution as the output. The low and high dimensional losses use the same convolutional regressor, but the 1 dimensional case averages over pixels. In the high dimensional case, the output has three channels (for RGB), with six channels total for shape and scale regression. We use the same non-linearities to constrain the shape and scale outputs to reasonable ranges as in (Barron, 2019). More specifically, we use an affine sigmoid to keep the shape $\\alpha \\in [ 0 , 3 ]$ and the softplus to keep scale $c \\in [ 1 0 ^ { - } 8 , \\infty )$ . Table 1 gives the results of evaluating each method by MSE on the validation set, while training each method with their respective loss parameters. Data parameter optimization uses Tensorflow’s implementation of sparse RMSProp (Tieleman & Hinton, 2012). We also inherit weight decay $\\| \\phi \\| _ { 2 } ^ { 2 }$ , gradient clipping $\\bar { \\nabla } _ { \\phi } / \\| \\nabla _ { \\phi } \\| _ { 2 } ^ { 2 }$ , and learning rate scaling $\\alpha _ { \\phi } = \\alpha \\cdot m$ for learning rate $\\alpha$ and multiplier $m$ from Barron (2019). ", + "bbox": [ + 174, + 627, + 825, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The robustness we see in our VAE experiments stems from the fact that likelihood parameter prediction gives models a direct channel to express their “uncertainty” for each data-point with respect to the task. This allows models to naturally down-weight and clean outliers from the dataset which can improve model robustness. Consequently, one can harness this effect to create outlier detectors from any underlying model architecture by using learned scales or temperatures as an outlier score function. Furthermore, predicted likelihood parameters allow these methods to detect outliers in unseen data. In Figure 5 we show how auditing temperature or noise parameters can help practitioners spot erroneous labels and poor quality examples. In particular, the model-parameterized temperatures of an image classifier (trained using the setup of 5.3) correlates strongly with blurry, dark, and difficult examples on the Street View House Number (SVHN) dataset. We use this approach to create simple outlier detection algorithms by considering deep $\\left( \\mathrm { A E } { + } \\mathrm { S } \\right)$ and linear $( \\mathrm { P C A } { + } \\mathrm { S } )$ auto-encoders (Kramer, 1991) with data-conditioned scale parameters as outlier scores. We evaluate this approach on tabular datasets using deep and linear auto-encoders with model-parameterized scales. In Table 2 we quantitatively demonstrate the quality of these simple likelihood parameter approaches across 22 datasets from the Outlier Detection Datasets (ODDS), a standard outlier detection benchmark (Rayana, 2016). The ODDS benchmark supplies ground truth outlier labels for each dataset, which allows one to treat outlier detection as an unsupervised classification problem. We compare against a variety of established outlier detection approaches included in the pyOD (Zhao et al., 2019) framework including: One-Class SVMs (OCSVM) (Scholkopf et al., 2000), Local Outlier Fraction (LOF) ¨ (Breunig et al., 2000), Angle Based Outlier Detection (ABOD) (Kriegel et al., 2008), Feature Bagging (FB) (Lazarevic & Kumar, 2005), Auto Encoder Distance (AE) (Aggarwal, 2015), K-Nearest Neighbors (KNN) (Ramaswamy et al., 2000; Angiulli & Pizzuti, 2002), Copula Based Outlier Detection (COPOD) (Li et al., 2020), Variational AutoEncoders (VAE) (Kingma & Welling, 2013), Minimum Covariance Determinants with Mahlanohbis Distance (MCD) (Rousseeuw & Driessen, 1999; Hardin & Rocke, 2004), Histogram-based Outlier Scores (HBOS) (Goldstein & Dengel, 2012), Principal Component Analysis (PCA) (Shyu et al., 2003), Isolation Forests (IF) (Liu et al., 2008; 2012), and the Clustering-Based Local Outlier Factor (CBLOF) (He et al., 2003). ", + "bbox": [ + 173, + 789, + 825, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13cf97ec3dd147a060f3c2aae3cb2a76c02361a9cc6b8cfbd652566d92745bec.jpg", + "image_caption": [ + "Figure 5: The data with the lowest (top) and highest (bottom) predicted temperatures in the SVHN dataset. High temperature entries are blurry, cropped poorly, and generally difficult to classify. " + ], + "image_footnote": [], + "bbox": [ + 187, + 103, + 812, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b5df76bf37dce536a934fc7313890a362bb93b0b5236c93444add7b70b98c027.jpg", + "table_caption": [ + "Table 2: Median outlier detection performance of several methods across 22 benchmark datasets from ODDS. " + ], + "table_footnote": [], + "table_body": "
MethodMedian AUC
LOF.669
FB.702
ABOD.727
AE.737
VAE.792
COPOD.799
PCA.808
OCSVM.814
MCD.820
KNN.822
HBOS.822
IF.823
CBLOF.836
AE+S (Ours).846
PCA+S (Ours).868
", + "bbox": [ + 230, + 338, + 436, + 563 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e93d5f5fd0e2d1653d0cf1d7ea585fa9b470765348bd2fcdb1405cb7f0328d0e.jpg", + "image_caption": [ + "Figure 6: Distribution of Outlier Detection AUC across the ODDS Benchmark. Our approaches, $\\mathrm { P C A } { + } \\mathrm { S }$ and $_ \\mathrm { A E + S }$ , are competitive with other Outlier Detection systems. " + ], + "image_footnote": [], + "bbox": [ + 506, + 291, + 818, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 587, + 825, + 856 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our predicted scale auto-encoders use PyTorch’s layers API (Paszke et al., 2019) with rectified linear unit (ReLU) activations for deep auto-encoders and Glorot uniform initialization (Dahl et al., 2013; Glorot & Bengio, 2010) for all layers. We use Adam (Kingma & Ba, 2015) with a learning rate of .0005 for 4000 steps with $20 \\%$ dropout before the code space. We follow ODDS guidelines and standard scale the data prior to fitting. ", + "bbox": [ + 176, + 863, + 823, + 906 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14ccfaff7d7b98120fd5a934955e0726811cf1fc2af4f91e79fd5ad741b81a41.jpg", + "image_caption": [ + "Figure 7: Performance of $L 2$ (left) and $L 1$ (middle) regularized linear regression on a 500 dimensional synthetic dataset where the true parameters, $w ^ { * }$ , are known. Dynamic Ridge (D-Ridge) and D-LASSO regression find the regularization strength that best estimates the true parameters. M-LASSO outperforms any single global regularization strength and does not shrink informative weights. (right) Performance of adaptive $L 1$ regularization methods as a function of true model sparsity. In all cases, Multi-LASSO outperforms other methods by orders of magnitude. " + ], + "image_footnote": [], + "bbox": [ + 178, + 102, + 803, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 371, + 820, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our methods $\\mathrm { P C A } { + } \\mathrm { S }$ and ${ \\mathrm { A E } } { + } S _ { \\cdot }$ ) use a similar principle as isolation-based approaches that determine outliers based on how difficult they are to model. In existing approaches, outliers influence and skew the isolation model which causes the model to exhibit less confidence on the whole. This hurts a model’s ability to distinguish between inliers and outliers. In contrast, our approach allows the underlying model to down-weight outliers. This yields a more consistent model with a clearer decision boundary between outliers and inliers as shown in Figure 4. As a future direction of investigation we note that our approach is model-architecture agnostic, and can be combined with domain-specific architectures to create outlier detection methods tailored to images, text, and audio. ", + "bbox": [ + 174, + 409, + 825, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 ADAPTIVE REGULARIZATION WITH PRIOR PARAMETERS ", + "text_level": 1, + "bbox": [ + 173, + 540, + 607, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to optimizing the shape and scale of the likelihood distribution of the model output, we can use the same approach to optimize the prior distribution of the model parameters. More specifically, we propose adaptive regularizers for a model’s parameters, $\\theta$ . This approach optimizes the distribution parameters of the prior, $\\phi _ { \\mathrm { p r i o r } }$ , to naturally tune the degree of regularization. In particular, the Normal (Ridge, L2) and Laplace (LASSO, L1) priors, with scale parameters $\\sigma$ and $b$ , regularize model parameters for small magnitude and sparsity respectively (Hastie et al., 2009). The degree of regularization, $\\lambda \\in [ 0 , \\infty )$ , is conventionally a hyperparameter of the regularized loss function: ", + "bbox": [ + 173, + 566, + 826, + 678 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/f627f0efe268e4bc649489c0cbab3f901c21e69ae5fcc84e0630628c62c374cd.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta } \\sum _ { i } ^ { N } ( \\hat { y _ { i } } : = f _ { \\theta } ( x _ { i } ) - y _ { i } ) ^ { 2 } + \\lambda \\sum _ { j } ^ { P } | \\theta _ { j } | .\n$$", + "text_format": "latex", + "bbox": [ + 357, + 675, + 640, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We note that we cannot choose $\\lambda$ by direct minimization because it admits a trivial minimum at $\\lambda = 0$ . In the linear case, one can select this weight efficiently using Least Angle Regression (Efron et al., 2004). However, in general $\\lambda$ is usually learned through expensive cross validation methods. Instead, we retain the prior with its scale parameter, and jointly optimize over the full likelihood: ", + "bbox": [ + 174, + 723, + 825, + 781 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/c16fdc665e8bcdc32122a2a851920c4d8ca32265b9b9927fbec62b60a2ed0519.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta , \\sigma , b } \\sum _ { i } ^ { N } \\left( \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } _ { i } - y _ { i } ) ^ { 2 } + \\log \\sigma \\right) + \\sum _ { j } ^ { P } \\left( \\frac { | \\theta _ { j } | } { b } + \\log b \\right)\n$$", + "text_format": "latex", + "bbox": [ + 308, + 797, + 691, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This approach, the Dynamic Lasso (D-LASSO), admits no trivial solution for the prior parameter $b$ , and must balance the effective regularization strength, $\\frac { 1 } { b }$ , with the normalization factor, $\\log b$ . D-LASSO selects the degree of regularization by gradient descent, rather than expensive black-box search. In Figure 7 (left) and (middle) we show that this approach, and its Ridge equivalent, yield ideal settings of the regularization strength on a suite of synthetic regression problems. Figure 7 (right) shows D-LASSO converges to the best LASSO regularization strength for a variety of truemodel sparsities. As a further extension, we replace the global $\\sigma$ or $b$ with a $\\sigma _ { j }$ or $b _ { j }$ for each model parameter, $\\theta _ { j }$ , to locally adapt regularization to each model weight (Multi-Lasso). This consistently outperforms any global setting of the regularization strength and shields important weights from undue shrinkage 7 (middle). For our experiments we use 500 samples of 500 dimensional normal distributions mapped through linear functions with additive gaussian noise. Linear transformations use Uniform $[ 1 , 2 ]$ weights and LASSO experiments use sparse transformations. We use tensorflow’s Adam optimizer with $l r = . 0 0 0 5$ for 100000 steps. ", + "bbox": [ + 174, + 853, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach of learning regularizer scale parameters can be viewed naturally through the lens of hierarchical priors (Gelman et al., 2013). More specifically this approach is implicitly performing maximum a posteriori (MAP) inference on the prior’s scale with respect to a uniform prior on that parameter. We note that though these methods for hyperparameter selection are common in the Bayesian literature, they are not widely used in practice in the deep learning community. This work aims to bring these parameters back within the scope of deep learning where they can be easily expanded to more flexible forms such as our introduced Multi-Lasso. ", + "bbox": [ + 174, + 224, + 823, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 RE-CALIBRATION ", + "text_level": 1, + "bbox": [ + 176, + 342, + 338, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work of (Guo et al., 2017) shows that modern networks are accurate, yet systematically overconfident, a phenomenon called mis-calibration. We investigate the role of optimizing likelihood parameters to re-calibrate models. More specifically, we can fit likelihood parameter regressors on a validation set to modify an existing model’s confidence to better align with the validation set. This approach is a generalization of Guo et al. (2017)’s Temperature Scaling method, which we refer to as Global Scaling (GS) for notational consistency. Global Scaling re-calibrates classifiers with a learned global parameter, $\\tau$ in the loss function: $\\sigma ( \\vec { x } , \\tau )$ . ", + "bbox": [ + 173, + 367, + 825, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fitting model-conditioned likelihood parameters to a validation set defines a broad class of recalibration strategies. From these we introduce three new re-calibration methods. Linear Scaling (LS) learns a linear mapping, $l$ , to transform logits to a softmax temperature: $\\sigma ( \\vec { x } , l ( \\vec { x } ) )$ . Linear Feature Scaling (LFS) learns a linear mapping, $l$ , to transform the features prior to the logits, $\\bar { f }$ , to a softmax temperature: $\\sigma ( \\vec { x } , l ( \\vec { f } ) )$ . Finally, we introduce Deep Scaling (DS) for regressors which learns a nonlinear network, $N$ , to transform features, $\\bar { f }$ , into a temperature: $\\sigma ( \\vec { x } , N ( \\vec { f } ) )$ . ", + "bbox": [ + 173, + 473, + 825, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 3 we compare our recalibration approaches to the previous state of the art: Global Scaling. We note that (Guo et al., 2017) have already shown that Global Scaling outperform Bayesian Binning into Quantiles (Naeini et al., 2015), Histogram binning (Zadrozny & Elkan, 2001), and Isotonic Regression. We recalibrate both ResNet50 (He et al., 2016) and DenseNet121 (Huang et al., 2017) on a variety of vision datasets. We measure classifier miscalibration using the Expected Calibration Error (ECE) (Guo et al., 2017) to align with prior art. We additionally evaluate Isotonic recalibration, Platt Scaling (Platt, 1999), and Vector Scaling (VS) (Guo et al., 2017), which learns a vector, $\\vec { v }$ , to re-weight logits: $\\sigma ( \\vec { v } \\vec { x } , 1 )$ . LS and LFS tend to outperform other approaches like GS and VS, which demonstrates that richer likelihood parametrizations can improve calibration akin to how richer models can improve prediction. ", + "bbox": [ + 173, + 570, + 825, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments leverage Tensorflow’s Dataset APIs that include the SVHN, (Netzer et al., 2011), ImageNet (Deng et al., 2009), CIFAR-100, CIFAR-10 (Krizhevsky, 2009) datasets. We use Keras implementations of DenseNet-121 (Huang et al., 2017) and ResNet-50 (He et al., 2016) with default initializations. For optimization we use Adam with $l r = 0 . 0 0 0 1$ , $\\beta _ { 1 } = . 9 , \\beta _ { 2 } = . 9 9$ (Kingma & Ba, 2015) and train for 300 epoch with a batch size of 512. ", + "bbox": [ + 174, + 713, + 823, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For recalibrating regressors, we compare against the previous state of the art, Kuleshov et al. (2018), who use an Isotonic regressor to correct a regressors’ confidence. We use the same experimental setting as Kuleshov et al. (2018) including the UCI datasets (Dua & Graff, 2017), and regressor calibration metric (CAL). Table 4 shows that our approaches can outperform this baseline as well as the regression equivalent of Global Scaling. Inputs and targets are scaled to unit norm and variance prior to fitting for all regression experiments and missing values are imputed using scikit-learn’s “SimpleImputer” (Pedregosa et al., 2011). Experiments utilize Keras’ layers API with two hidden rectified linear unit (ReLU) layers, Glorot uniform initialization (Dahl et al., 2013; Glorot & Bengio, 2010) and Adam optimization with $l r = 0 . 0 0 1$ for 3000 steps without minibatching. ", + "bbox": [ + 174, + 791, + 825, + 915 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/369344b0ae6cfc25469ef3c2cd30adf97e0a11d84f88602e26aec2e434c106e2.jpg", + "table_caption": [ + "Table 3: Comparison of calibration methods by ECE for ResNet-50 (RN50) and DenseNet-121 (DN121) architectures on test data. Our predicted likelihood parameter methods: Linear Scaling (LS) and Linear Feature Scaling (LFS) outperform other approaches. In all cases our methods reduce miscalibration with comparable computation time as GS. " + ], + "table_footnote": [], + "table_body": "
ModelDatasetUncalibratedPlattIsotonicGSVSLSLFS
RN50CIFAR-10.250.034.053.046.037.018.018
RN50CIFAR-100.642.061.072.035.044.030.173
RN50SVHN.072.053.010.029.022.009.009
RN50ImageNet.430.018.070.019.023.026.015
DN121CIFAR-10.253.048.042.039.034.028.028
DN121CIFAR-100.537.049.067.024.024.014.031
DN121SVHN.079.018.010.022.017.011.010
DN121ImageNet.229.028.095.021.019.043.019
", + "bbox": [ + 235, + 165, + 761, + 292 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ce104ddd599fdb776ec7eda3b07530a3b5dc8713e4fa687011f6a95cf6631b79.jpg", + "table_caption": [ + "Table 4: Comparison of regression calibration methods as evaluated by their calibration error as defined in (Kuleshov et al., 2018). Predicted likelihood parameters often outperform other methods. " + ], + "table_footnote": [], + "table_body": "
DatasetUncalibrated IsotonicGSLSDS
crime0.36240.34990.0693 0.0125 0.0310
kinematics0.01640.01030.0022 0.0021 0.0032
bank0.01220.00560.0027 0.0024 0.0020
wine0.00910.01080.0152 0.01310.0064
mpg0.21530.22000.1964 0.14830.0233
cpu0.08620.03400.3018 0.20780.1740
soil0.30830.30000.3130 0.3175 0.3137
fried0.00060.00020.00020.00020.0002
", + "bbox": [ + 307, + 352, + 689, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 EXPERIMENTAL DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 522, + 416, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We run all experiments on Ubuntu 16.04 Azure Standard NV24 virtual machines (24 CPUs, 224 Gb memory, and $4 \\times \\mathrm { M } 6 0 \\ : \\mathrm { G P U s } )$ with Tensorflow 1.15 (Abadi et al., 2015) and PyTorch 1.17 (Paszke et al., 2019). Many likelihood parameters have constrained domains, such as the normal variance $\\sigma \\in [ 0 , \\infty )$ . To evade the complexity of constrained optimization, we define unconstrained parameters $\\phi _ { u }$ and choose a transformation $t ( \\cdot )$ with inverse $t ^ { - 1 } ( \\cdot )$ to map to and from the constrained $\\phi$ . For positivity, exp/log parameterization is standard (Kendall & Gal, 2017; Kendall et al., 2018; Saxena et al., 2019). However, this parameterization can lead to instabilities and we use the softplus, $s ^ { + } ( x ) = \\log ( 1 + \\exp ( x ) )$ , instead. Shifting the softplus, $s _ { c } ^ { + } ( x ) = ( l n ( 1 + e ^ { x } ) + c ) / ( l n ( 2 ) + c ) ,$ , further improves stability and we explore this effect in Figure 12 of the Appendix. We use an affine softplus $s _ { . 0 1 } ^ { \\mp }$ and $s _ { . 2 } ^ { + }$ respectively for adaptive scales and temperatures respectively. The one exception is adaptive regularizer scales where we found exp led to faster convergence. For the constrained interval $[ a , b ]$ we use affine transformations of the sigmoid $\\begin{array} { r } { s ( x ) = \\frac { \\tilde { 1 } } { 1 + \\exp ( - x ) } } \\end{array}$ (Barron, 2019). We initialize likelihood parameter biases to settings that yield MSE and Cross Entropy $\\sigma = \\tau = 1 \\mathrm { { } } $ . ", + "bbox": [ + 174, + 554, + 825, + 738 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 765, + 318, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Optimizing the full likelihood can improve model quality by adapting losses and regularizers. Full likelihoods are agnostic to the architecture, optimizer, and task, which makes them simple substitutes for standard losses. Global, data, and predicted likelihood parameters offer different degrees of expressivity and efficiency. In particular, predicted parameters adapt the likelihood to each data point during training and testing without significant time and space overhead. By including these parameters in a loss function one can improve a model’s robustness and generalization ability and create new classes of outlier detectors and recalibrators that outperform baselines. More generally, we hope this work encourages joint optimization of model and likelihood parameters, and argue it is likely that your loss should be a likelihood. ", + "bbox": [ + 173, + 797, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´ Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´ berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org. ", + "bbox": [ + 176, + 126, + 825, + 251 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Charu C Aggarwal. Outlier analysis. In Data mining, pp. 237–263. Springer, 2015. ", + "bbox": [ + 176, + 260, + 718, + 275 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European conference on principles of data mining and knowledge discovery, pp. 15–27. Springer, 2002. ", + "bbox": [ + 174, + 284, + 820, + 311 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jonathan T Barron. A general and adaptive robust loss function. In CVPR, 2019. ", + "bbox": [ + 173, + 320, + 702, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Christopher M Bishop. Mixture density networks. 1994. ", + "bbox": [ + 174, + 344, + 545, + 358 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Christopher M Bishop et al. Pattern recognition and machine learning, volume 4. springer New York, 2006. ", + "bbox": [ + 173, + 368, + 823, + 396 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Markus M Breunig, Hans-Peter Kriegel, Raymond T $\\mathrm { N g }$ , and Jorg Sander. Lof: identifying density- ¨ based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pp. 93–104, 2000. ", + "bbox": [ + 176, + 405, + 823, + 448 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Deterministic edge- ´ preserving regularization in computed imaging. IEEE Transactions on image processing, 6(2): 298–311, 1997. ", + "bbox": [ + 174, + 457, + 825, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "George E Dahl, Tara N Sainath, and Geoffrey E Hinton. Improving deep neural networks for lvcsr using rectified linear units and dropout. In ICASSP, pp. 8609–8613. IEEE, 2013. ", + "bbox": [ + 173, + 507, + 820, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pp. 248–255. Ieee, 2009. ", + "bbox": [ + 174, + 544, + 820, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John E Dennis Jr and Roy E Welsch. Techniques for nonlinear least squares and robust regression. Communications in Statistics-simulation and Computation, 7(4):345–359, 1978. ", + "bbox": [ + 176, + 582, + 820, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dheeru Dua and Casey Graff. UCI machine learning repository, 2017. URL http://archive. ics.uci.edu/ml. ", + "bbox": [ + 174, + 619, + 818, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Bradley Efron, Trevor Hastie, Iain Johnstone, Robert Tibshirani, et al. Least angle regression. The Annals of statistics, 32(2):407–499, 2004. ", + "bbox": [ + 173, + 656, + 820, + 685 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24 (6):381–395, 1981. ", + "bbox": [ + 173, + 694, + 825, + 737 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission tomography. Amer. Statist. Assoc, pp. 12–18, 1985. ", + "bbox": [ + 173, + 746, + 821, + 775 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andrew Gelman, John B Carlin, Hal S Stern, David B Dunson, Aki Vehtari, and Donald B Rubin. Bayesian data analysis. CRC press, 2013. ", + "bbox": [ + 173, + 782, + 820, + 811 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, pp. 249–256, 2010. ", + "bbox": [ + 174, + 820, + 821, + 849 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Markus Goldstein and Andreas Dengel. Histogram-based outlier score (hbos): A fast unsupervised anomaly detection algorithm. KI-2012: Poster and Demo Track, pp. 59–63, 2012. ", + "bbox": [ + 173, + 858, + 820, + 887 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Johanna Hardin and David M Rocke. Outlier detection in the multiple cluster setting using the minimum covariance determinant estimator. Computational Statistics & Data Analysis, 44(4): 625–638, 2004. ", + "bbox": [ + 173, + 103, + 821, + 146 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning: data mining, inference, and prediction. Springer Science & Business Media, 2009. ", + "bbox": [ + 171, + 156, + 823, + 185 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. ", + "bbox": [ + 173, + 194, + 820, + 223 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zengyou He, Xiaofei Xu, and Shengchun Deng. Discovering cluster-based local outliers. Pattern Recognition Letters, 24(9-10):1641–1650, 2003. ", + "bbox": [ + 173, + 233, + 821, + 262 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. ", + "bbox": [ + 173, + 272, + 823, + 301 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In CVPR, 2017. ", + "bbox": [ + 174, + 310, + 823, + 340 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Peter J Huber. Robust statistics, volume 523. John Wiley & Sons, 2004. ", + "bbox": [ + 173, + 349, + 648, + 364 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, pp. 5574–5584, 2017. ", + "bbox": [ + 176, + 375, + 823, + 404 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pp. 7482–7491, 2018. ", + "bbox": [ + 173, + 412, + 823, + 443 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. ", + "bbox": [ + 173, + 452, + 799, + 468 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. ", + "bbox": [ + 171, + 477, + 823, + 506 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks. AIChE journal, 37(2):233–243, 1991. ", + "bbox": [ + 169, + 516, + 823, + 545 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hans-Peter Kriegel, Matthias Schubert, and Arthur Zimek. Angle-based outlier detection in highdimensional data. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 444–452, 2008. ", + "bbox": [ + 174, + 554, + 821, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. ", + "bbox": [ + 173, + 607, + 821, + 636 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Volodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated regression. In Jennifer Dy and Andreas Krause (eds.), ICML, volume 80 of Proceedings of Machine Learning Research, pp. 2801–2809, Stockholmsmassan, Stockholm ¨ Sweden, 10–15 Jul 2018. PMLR. ", + "bbox": [ + 173, + 645, + 825, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aleksandar Lazarevic and Vipin Kumar. Feature bagging for outlier detection. In Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, pp. 157–166, 2005. ", + "bbox": [ + 173, + 712, + 826, + 756 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xuelong Li, Quanmao Lu, Yongsheng Dong, and Dacheng Tao. Robust subspace clustering by cauchy loss function. IEEE transactions on neural networks and learning systems, 30(7):2067– 2078, 2018. ", + "bbox": [ + 173, + 765, + 825, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zheng Li, Yue Zhao, Nicola Botta, Cezar Ionescu, and Xiyang Hu. Copod: copula-based outlier detection. arXiv preprint arXiv:2009.09463, 2020. ", + "bbox": [ + 169, + 816, + 823, + 847 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 Eighth IEEE International Conference on Data Mining, pp. 413–422. IEEE, 2008. ", + "bbox": [ + 173, + 856, + 821, + 886 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation-based anomaly detection. ACM Transactions on Knowledge Discovery from Data (TKDD), 6(1):1–39, 2012. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In ICCV, December 2015. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "David JC MacKay and David JC Mac Kay. Information theory, inference and learning algorithms. Cambridge university press, 2003. ", + "bbox": [ + 174, + 138, + 821, + 167 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. ", + "bbox": [ + 174, + 174, + 823, + 217 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "John Ashworth Nelder and Robert WM Wedderburn. Generalized linear models. Journal of the Royal Statistical Society: Series A (General), 135(3):370–384, 1972. ", + "bbox": [ + 174, + 223, + 825, + 252 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. ", + "bbox": [ + 176, + 258, + 823, + 289 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In ICNN, volume 1, pp. 55–60. IEEE, 1994. ", + "bbox": [ + 173, + 295, + 821, + 324 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´ R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 8024–8035. Curran Associates, Inc., 2019. ", + "bbox": [ + 174, + 330, + 825, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier ¨ Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011. ", + "bbox": [ + 173, + 435, + 825, + 478 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "John C Platt. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61–74, 1999. ", + "bbox": [ + 173, + 484, + 821, + 513 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. Efficient algorithms for mining outliers from large data sets. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pp. 427–438, 2000. ", + "bbox": [ + 173, + 520, + 825, + 563 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Shebuti Rayana. Odds library. Stony Brook,-2016, (2017), 2016. ", + "bbox": [ + 173, + 569, + 598, + 584 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Peter J Rousseeuw and Katrien Van Driessen. A fast algorithm for the minimum covariance determinant estimator. Technometrics, 41(3):212–223, 1999. ", + "bbox": [ + 173, + 590, + 823, + 619 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Shreyas Saxena, Oncel Tuzel, and Dennis DeCoste. Data parameters: A new family of parameters for learning a differentiable curriculum. In NeurIPS, pp. 11093–11103, 2019. ", + "bbox": [ + 171, + 626, + 821, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Bernhard Scholkopf, Robert C Williamson, Alex J Smola, John Shawe-Taylor, and John C Platt.¨ Support vector method for novelty detection. In Advances in neural information processing systems, pp. 582–588, 2000. ", + "bbox": [ + 173, + 661, + 821, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and LiWu Chang. A novel anomaly detection scheme based on principal component classifier. Technical report, MIAMI UNIV CORAL GABLES FL DEPT OF ELECTRICAL AND COMPUTER ENGINEERING, 2003. ", + "bbox": [ + 173, + 712, + 825, + 753 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Henri Theil. A rank-invariant method of linear and polynomial regression analysis. In Henri Theil’s contributions to economics and econometrics, pp. 345–381. Springer, 1992. ", + "bbox": [ + 173, + 761, + 823, + 790 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26– 31, 2012. ", + "bbox": [ + 173, + 796, + 825, + 838 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML, volume 1, pp. 609–616. Citeseer, 2001. ", + "bbox": [ + 173, + 845, + 820, + 876 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detection. Journal of Machine Learning Research, 20(96):1–7, 2019. URL http://jmlr.org/ papers/v20/19-011.html. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_middle.json b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..a5852335ca3baa545ac66ae8ad336a8d05ea0a28 --- /dev/null +++ b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_middle.json @@ -0,0 +1,34873 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 340, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 341, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 341, + 96 + ], + "score": 1.0, + "content": "IT IS LIKELY THAT YOUR LOSS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 309, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 309, + 118 + ], + "score": 1.0, + "content": "SHOULD BE A LIKELIHOOD", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 197, + 168 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 182, + 146 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 182, + 146 + ], + "score": 1.0, + "content": "Mark Hamilton", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 177, + 156 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 177, + 156 + ], + "score": 1.0, + "content": "MIT, Microsoft", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 158, + 198, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 158, + 198, + 168 + ], + "score": 1.0, + "content": "markth@mit.edu", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 236, + 135, + 352, + 168 + ], + "lines": [ + { + "bbox": [ + 236, + 135, + 309, + 146 + ], + "spans": [ + { + "bbox": [ + 236, + 135, + 309, + 146 + ], + "score": 1.0, + "content": "Evan Shelhamer", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 145, + 327, + 157 + ], + "spans": [ + { + "bbox": [ + 235, + 145, + 327, + 157 + ], + "score": 1.0, + "content": "MIT, Adobe Research", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 158, + 353, + 168 + ], + "spans": [ + { + "bbox": [ + 236, + 158, + 353, + 168 + ], + "score": 1.0, + "content": "shelhamer@adobe.com", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 390, + 135, + 469, + 168 + ], + "lines": [ + { + "bbox": [ + 390, + 134, + 467, + 147 + ], + "spans": [ + { + "bbox": [ + 390, + 134, + 467, + 147 + ], + "score": 1.0, + "content": "William Freeman", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 389, + 145, + 446, + 158 + ], + "spans": [ + { + "bbox": [ + 389, + 145, + 446, + 158 + ], + "score": 1.0, + "content": "MIT, Google", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 390, + 158, + 470, + 168 + ], + "spans": [ + { + "bbox": [ + 390, + 158, + 470, + 168 + ], + "score": 1.0, + "content": "billf@mit.edu", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 278, + 197, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 211 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 211 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 143, + 223, + 468, + 344 + ], + "lines": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "score": 1.0, + "content": "Many common loss functions such as mean-squared-error, cross-entropy, and re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "score": 1.0, + "content": "construction loss are unnecessarily rigid. Under a probabilistic interpretation,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "these common losses correspond to distributions with fixed shapes and scales.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "score": 1.0, + "content": "We instead argue for optimizing full likelihoods that include parameters like the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "normal variance and softmax temperature. Joint optimization of these “likelihood", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "parameters” with model parameters can adaptively tune the scales and shapes of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "losses in addition to the strength of regularization. We explore and systematically", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "score": 1.0, + "content": "evaluate how to parameterize and apply likelihood parameters for robust mod-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "eling, outlier-detection, and re-calibration. Additionally, we propose adaptively", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 142, + 322, + 171, + 334 + ], + "score": 1.0, + "content": "tuning", + "type": "text" + }, + { + "bbox": [ + 171, + 322, + 183, + 333 + ], + "score": 0.87, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 322, + 202, + 334 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 202, + 322, + 215, + 333 + ], + "score": 0.88, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "weights by fitting the scale parameters of normal and Laplace", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 332, + 388, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 388, + 346 + ], + "score": 1.0, + "content": "priors and introduce more flexible element-wise regularizers.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 367, + 206, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 208, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 208, + 383 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "Choosing the right loss matters. Many common losses arise from likelihoods, such as the squared", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 405, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 416 + ], + "score": 1.0, + "content": "error loss from the normal distribution , absolute error from the Laplace distribution, and the cross", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 437, + 428 + ], + "score": 1.0, + "content": "entropy loss from the softmax distribution. The same is true of regularizers, where", + "type": "text" + }, + { + "bbox": [ + 437, + 416, + 450, + 426 + ], + "score": 0.88, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "arises from a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 426, + 276, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 176, + 438 + ], + "score": 1.0, + "content": "normal prior and", + "type": "text" + }, + { + "bbox": [ + 176, + 426, + 188, + 437 + ], + "score": 0.89, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 427, + 276, + 438 + ], + "score": 1.0, + "content": "from a Laplace prior.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 504, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 504, + 454 + ], + "score": 1.0, + "content": "Deriving losses from likelihoods recasts the problem as a choice of distribution which allows data-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "score": 1.0, + "content": "dependent adaptation. Standard losses and regularizers implicitly fix key distribution parameters,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "limiting flexibility. For instance, the squared error corresponds to fixing the normal variance at", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "score": 1.0, + "content": "a constant. The full normal likelihood retains its scale parameter and allows optimization over", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "a parametrized set of distributions. This work examines how to jointly optimize distribution and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "model parameters to select losses and regularizers that encourage generalization, calibration, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "robustness to outliers. We explore three key likelihoods: the normal, softmax, and the robust re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 185, + 533 + ], + "score": 1.0, + "content": "gression likelihood", + "type": "text" + }, + { + "bbox": [ + 185, + 522, + 192, + 532 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "of Barron (2019). Additionally, we cast adaptive priors in the same light and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 531, + 311, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 311, + 543 + ], + "score": 1.0, + "content": "introduce adaptive regularizers. Our contributions:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 130, + 553, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 129, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 129, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "1. We systematically survey and evaluate global, data, and predicted likelihood parameters", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 564, + 414, + 578 + ], + "spans": [ + { + "bbox": [ + 141, + 564, + 407, + 578 + ], + "score": 1.0, + "content": "and introduce a new self-tuning variant of the robust adaptive loss", + "type": "text" + }, + { + "bbox": [ + 407, + 567, + 414, + 576 + ], + "score": 0.71, + "content": "\\rho", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 574, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 127, + 574, + 506, + 590 + ], + "score": 1.0, + "content": "2. We apply likelihood parameters to create new classes of robust models, outlier detectors,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 142, + 588, + 217, + 598 + ], + "spans": [ + { + "bbox": [ + 142, + 588, + 217, + 598 + ], + "score": 1.0, + "content": "and re-calibrators.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 128, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 128, + 597, + 273, + 612 + ], + "score": 1.0, + "content": "3. We propose adaptive versions of", + "type": "text" + }, + { + "bbox": [ + 273, + 599, + 286, + 609 + ], + "score": 0.8, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 597, + 303, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 304, + 599, + 317, + 609 + ], + "score": 0.81, + "content": "L 2", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "regularization using parameterized normal and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 141, + 610, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 141, + 610, + 291, + 622 + ], + "score": 1.0, + "content": "Laplace priors on model parameters.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 107, + 637, + 200, + 651 + ], + "lines": [ + { + "bbox": [ + 104, + 636, + 201, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 201, + 653 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 663, + 504, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 239, + 677 + ], + "score": 1.0, + "content": "Notation We consider a dataset", + "type": "text" + }, + { + "bbox": [ + 239, + 664, + 249, + 674 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 663, + 288, + 677 + ], + "score": 1.0, + "content": "of points", + "type": "text" + }, + { + "bbox": [ + 289, + 666, + 298, + 675 + ], + "score": 0.85, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 663, + 346, + 677 + ], + "score": 1.0, + "content": "and targets", + "type": "text" + }, + { + "bbox": [ + 347, + 666, + 356, + 675 + ], + "score": 0.85, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 663, + 405, + 677 + ], + "score": 1.0, + "content": "indexed by", + "type": "text" + }, + { + "bbox": [ + 405, + 664, + 468, + 676 + ], + "score": 0.92, + "content": "i \\in \\{ 1 , \\ldots , N \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 663, + 506, + 677 + ], + "score": 1.0, + "content": ". Targets", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 476, + 687 + ], + "score": 1.0, + "content": "for regression are real numbers and targets for classification are one-hot vectors. The model", + "type": "text" + }, + { + "bbox": [ + 476, + 676, + 484, + 687 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 686, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 153, + 699 + ], + "score": 1.0, + "content": "parameters", + "type": "text" + }, + { + "bbox": [ + 153, + 686, + 159, + 696 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 686, + 237, + 699 + ], + "score": 1.0, + "content": "makes predictions", + "type": "text" + }, + { + "bbox": [ + 237, + 686, + 283, + 698 + ], + "score": 0.94, + "content": "\\hat { y } _ { i } = f _ { \\theta } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 686, + 318, + 699 + ], + "score": 1.0, + "content": ". A loss", + "type": "text" + }, + { + "bbox": [ + 318, + 686, + 348, + 698 + ], + "score": 0.93, + "content": "L ( \\hat { y } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 686, + 506, + 699 + ], + "score": 1.0, + "content": "measures the quality of the prediction", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 697, + 449, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 449, + 710 + ], + "score": 1.0, + "content": "given the target. To learn model parameters we solve the following loss optimization:", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 714, + 365, + 735 + ], + "lines": [ + { + "bbox": [ + 246, + 714, + 365, + 735 + ], + "spans": [ + { + "bbox": [ + 246, + 714, + 365, + 735 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta } \\underset { ( x , y ) \\sim \\mathcal { D } } { \\mathbb { E } } L ( \\hat { y } = f _ { \\theta } ( x ) , y )", + "type": "interline_equation", + "image_path": "feb3b89e30bd398c8e0cee30d0790e7945f21c973bbfebdfe555fe9624bfe8f7.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 246, + 714, + 365, + 735 + ], + "spans": [], + "index": 48 + } + ] + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 340, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 341, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 341, + 96 + ], + "score": 1.0, + "content": "IT IS LIKELY THAT YOUR LOSS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 309, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 309, + 118 + ], + "score": 1.0, + "content": "SHOULD BE A LIKELIHOOD", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 197, + 168 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 182, + 146 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 182, + 146 + ], + "score": 1.0, + "content": "Mark Hamilton", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 177, + 156 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 177, + 156 + ], + "score": 1.0, + "content": "MIT, Microsoft", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 158, + 198, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 158, + 198, + 168 + ], + "score": 1.0, + "content": "markth@mit.edu", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 111, + 134, + 198, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 236, + 135, + 352, + 168 + ], + "lines": [ + { + "bbox": [ + 236, + 135, + 309, + 146 + ], + "spans": [ + { + "bbox": [ + 236, + 135, + 309, + 146 + ], + "score": 1.0, + "content": "Evan Shelhamer", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 145, + 327, + 157 + ], + "spans": [ + { + "bbox": [ + 235, + 145, + 327, + 157 + ], + "score": 1.0, + "content": "MIT, Adobe Research", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 158, + 353, + 168 + ], + "spans": [ + { + "bbox": [ + 236, + 158, + 353, + 168 + ], + "score": 1.0, + "content": "shelhamer@adobe.com", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6, + "bbox_fs": [ + 235, + 135, + 353, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 390, + 135, + 469, + 168 + ], + "lines": [ + { + "bbox": [ + 390, + 134, + 467, + 147 + ], + "spans": [ + { + "bbox": [ + 390, + 134, + 467, + 147 + ], + "score": 1.0, + "content": "William Freeman", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 389, + 145, + 446, + 158 + ], + "spans": [ + { + "bbox": [ + 389, + 145, + 446, + 158 + ], + "score": 1.0, + "content": "MIT, Google", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 390, + 158, + 470, + 168 + ], + "spans": [ + { + "bbox": [ + 390, + 158, + 470, + 168 + ], + "score": 1.0, + "content": "billf@mit.edu", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7, + "bbox_fs": [ + 389, + 134, + 470, + 168 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 197, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 211 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 211 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 143, + 223, + 468, + 344 + ], + "lines": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "score": 1.0, + "content": "Many common loss functions such as mean-squared-error, cross-entropy, and re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "score": 1.0, + "content": "construction loss are unnecessarily rigid. Under a probabilistic interpretation,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "these common losses correspond to distributions with fixed shapes and scales.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "score": 1.0, + "content": "We instead argue for optimizing full likelihoods that include parameters like the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "normal variance and softmax temperature. Joint optimization of these “likelihood", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "parameters” with model parameters can adaptively tune the scales and shapes of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "losses in addition to the strength of regularization. We explore and systematically", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "score": 1.0, + "content": "evaluate how to parameterize and apply likelihood parameters for robust mod-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "eling, outlier-detection, and re-calibration. Additionally, we propose adaptively", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 142, + 322, + 171, + 334 + ], + "score": 1.0, + "content": "tuning", + "type": "text" + }, + { + "bbox": [ + 171, + 322, + 183, + 333 + ], + "score": 0.87, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 322, + 202, + 334 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 202, + 322, + 215, + 333 + ], + "score": 0.88, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "weights by fitting the scale parameters of normal and Laplace", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 332, + 388, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 388, + 346 + ], + "score": 1.0, + "content": "priors and introduce more flexible element-wise regularizers.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17, + "bbox_fs": [ + 141, + 223, + 470, + 346 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 367, + 206, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 208, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 208, + 383 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "Choosing the right loss matters. Many common losses arise from likelihoods, such as the squared", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 405, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 416 + ], + "score": 1.0, + "content": "error loss from the normal distribution , absolute error from the Laplace distribution, and the cross", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 437, + 428 + ], + "score": 1.0, + "content": "entropy loss from the softmax distribution. The same is true of regularizers, where", + "type": "text" + }, + { + "bbox": [ + 437, + 416, + 450, + 426 + ], + "score": 0.88, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "arises from a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 426, + 276, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 176, + 438 + ], + "score": 1.0, + "content": "normal prior and", + "type": "text" + }, + { + "bbox": [ + 176, + 426, + 188, + 437 + ], + "score": 0.89, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 427, + 276, + 438 + ], + "score": 1.0, + "content": "from a Laplace prior.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 393, + 505, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 504, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 504, + 454 + ], + "score": 1.0, + "content": "Deriving losses from likelihoods recasts the problem as a choice of distribution which allows data-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 468 + ], + "score": 1.0, + "content": "dependent adaptation. Standard losses and regularizers implicitly fix key distribution parameters,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "limiting flexibility. For instance, the squared error corresponds to fixing the normal variance at", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "score": 1.0, + "content": "a constant. The full normal likelihood retains its scale parameter and allows optimization over", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "a parametrized set of distributions. This work examines how to jointly optimize distribution and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "model parameters to select losses and regularizers that encourage generalization, calibration, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "robustness to outliers. We explore three key likelihoods: the normal, softmax, and the robust re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 185, + 533 + ], + "score": 1.0, + "content": "gression likelihood", + "type": "text" + }, + { + "bbox": [ + 185, + 522, + 192, + 532 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "of Barron (2019). Additionally, we cast adaptive priors in the same light and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 531, + 311, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 311, + 543 + ], + "score": 1.0, + "content": "introduce adaptive regularizers. Our contributions:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 443, + 506, + 543 + ] + }, + { + "type": "list", + "bbox": [ + 130, + 553, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 129, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 129, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "1. We systematically survey and evaluate global, data, and predicted likelihood parameters", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 564, + 414, + 578 + ], + "spans": [ + { + "bbox": [ + 141, + 564, + 407, + 578 + ], + "score": 1.0, + "content": "and introduce a new self-tuning variant of the robust adaptive loss", + "type": "text" + }, + { + "bbox": [ + 407, + 567, + 414, + 576 + ], + "score": 0.71, + "content": "\\rho", + "type": "inline_equation" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 127, + 574, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 127, + 574, + 506, + 590 + ], + "score": 1.0, + "content": "2. We apply likelihood parameters to create new classes of robust models, outlier detectors,", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 588, + 217, + 598 + ], + "spans": [ + { + "bbox": [ + 142, + 588, + 217, + 598 + ], + "score": 1.0, + "content": "and re-calibrators.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 128, + 597, + 273, + 612 + ], + "score": 1.0, + "content": "3. We propose adaptive versions of", + "type": "text" + }, + { + "bbox": [ + 273, + 599, + 286, + 609 + ], + "score": 0.8, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 597, + 303, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 304, + 599, + 317, + 609 + ], + "score": 0.81, + "content": "L 2", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "regularization using parameterized normal and", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 610, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 141, + 610, + 291, + 622 + ], + "score": 1.0, + "content": "Laplace priors on model parameters.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + } + ], + "index": 39.5, + "bbox_fs": [ + 127, + 553, + 506, + 622 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 637, + 200, + 651 + ], + "lines": [ + { + "bbox": [ + 104, + 636, + 201, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 201, + 653 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 663, + 504, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 239, + 677 + ], + "score": 1.0, + "content": "Notation We consider a dataset", + "type": "text" + }, + { + "bbox": [ + 239, + 664, + 249, + 674 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 663, + 288, + 677 + ], + "score": 1.0, + "content": "of points", + "type": "text" + }, + { + "bbox": [ + 289, + 666, + 298, + 675 + ], + "score": 0.85, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 663, + 346, + 677 + ], + "score": 1.0, + "content": "and targets", + "type": "text" + }, + { + "bbox": [ + 347, + 666, + 356, + 675 + ], + "score": 0.85, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 663, + 405, + 677 + ], + "score": 1.0, + "content": "indexed by", + "type": "text" + }, + { + "bbox": [ + 405, + 664, + 468, + 676 + ], + "score": 0.92, + "content": "i \\in \\{ 1 , \\ldots , N \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 663, + 506, + 677 + ], + "score": 1.0, + "content": ". Targets", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 476, + 687 + ], + "score": 1.0, + "content": "for regression are real numbers and targets for classification are one-hot vectors. The model", + "type": "text" + }, + { + "bbox": [ + 476, + 676, + 484, + 687 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 686, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 153, + 699 + ], + "score": 1.0, + "content": "parameters", + "type": "text" + }, + { + "bbox": [ + 153, + 686, + 159, + 696 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 686, + 237, + 699 + ], + "score": 1.0, + "content": "makes predictions", + "type": "text" + }, + { + "bbox": [ + 237, + 686, + 283, + 698 + ], + "score": 0.94, + "content": "\\hat { y } _ { i } = f _ { \\theta } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 686, + 318, + 699 + ], + "score": 1.0, + "content": ". A loss", + "type": "text" + }, + { + "bbox": [ + 318, + 686, + 348, + 698 + ], + "score": 0.93, + "content": "L ( \\hat { y } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 686, + 506, + 699 + ], + "score": 1.0, + "content": "measures the quality of the prediction", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 697, + 449, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 449, + 710 + ], + "score": 1.0, + "content": "given the target. To learn model parameters we solve the following loss optimization:", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 663, + 506, + 710 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 714, + 365, + 735 + ], + "lines": [ + { + "bbox": [ + 246, + 714, + 365, + 735 + ], + "spans": [ + { + "bbox": [ + 246, + 714, + 365, + 735 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta } \\underset { ( x , y ) \\sim \\mathcal { D } } { \\mathbb { E } } L ( \\hat { y } = f _ { \\theta } ( x ) , y )", + "type": "interline_equation", + "image_path": "feb3b89e30bd398c8e0cee30d0790e7945f21c973bbfebdfe555fe9624bfe8f7.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 246, + 714, + 365, + 735 + ], + "spans": [], + "index": 48 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 83, + 499, + 190 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 83, + 499, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 83, + 499, + 190 + ], + "spans": [ + { + "bbox": [ + 110, + 83, + 499, + 190 + ], + "score": 0.965, + "type": "image", + "image_path": "b1ea4f3c37fb9dbc501c429bffa66313c780d218dcb161499e075885f826da71.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 83, + 499, + 118.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 118.66666666666666, + 499, + 154.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 154.33333333333331, + 499, + 189.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 202, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 200, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 216 + ], + "score": 1.0, + "content": "Figure 1: Optimizing likelihood parameters adapts the loss without manual hyperparameter tuning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 212, + 244, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 244, + 226 + ], + "score": 1.0, + "content": "to balance accuracy and certainty.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 246, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 161, + 260 + ], + "score": 1.0, + "content": "A likelihood", + "type": "text" + }, + { + "bbox": [ + 162, + 246, + 201, + 259 + ], + "score": 0.93, + "content": "\\mathcal { L } ( \\hat { y } | y , \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 246, + 455, + 260 + ], + "score": 1.0, + "content": "measures the quality of the prediction as a distribution over", + "type": "text" + }, + { + "bbox": [ + 456, + 247, + 462, + 258 + ], + "score": 0.82, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 246, + 505, + 260 + ], + "score": 1.0, + "content": "given the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 131, + 270 + ], + "score": 1.0, + "content": "target", + "type": "text" + }, + { + "bbox": [ + 131, + 259, + 138, + 269 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 257, + 244, + 270 + ], + "score": 1.0, + "content": "and likelihood parameters", + "type": "text" + }, + { + "bbox": [ + 244, + 258, + 251, + 269 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 257, + 394, + 270 + ], + "score": 1.0, + "content": ". We use the negative log-likelihood", + "type": "text" + }, + { + "bbox": [ + 394, + 258, + 400, + 268 + ], + "score": 0.68, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 257, + 506, + 270 + ], + "score": 1.0, + "content": "(NLL), and the likelihood", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 269, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 479, + 281 + ], + "score": 1.0, + "content": "interchangeably since both have the same optima. We define the full likelihood optimization:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 285, + 369, + 306 + ], + "lines": [ + { + "bbox": [ + 244, + 285, + 369, + 306 + ], + "spans": [ + { + "bbox": [ + 244, + 285, + 369, + 306 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta , \\phi } \\quad \\mathbb { E } _ { \\mathbf { \\Phi } ( x , y ) \\sim \\mathcal { D } } \\ell ( \\hat { y } = f _ { \\theta } ( x ) | y , \\phi )", + "type": "interline_equation", + "image_path": "6b7c1f81e8f579b80c819e0155429a965350c54767ecabb1a4d0fead8f057d8a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 244, + 285, + 369, + 306 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 433, + 324 + ], + "score": 1.0, + "content": "to jointly learn model and likelihood parameters. “Full” indicates the inclusion of", + "type": "text" + }, + { + "bbox": [ + 433, + 312, + 440, + 324 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 312, + 505, + 324 + ], + "score": 1.0, + "content": ", which controls", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "score": 1.0, + "content": "the distribution and induced NLL loss. We focus on full likelihood optimization in this work. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 332, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 104, + 332, + 186, + 348 + ], + "score": 1.0, + "content": "note that the target,", + "type": "text" + }, + { + "bbox": [ + 187, + 336, + 193, + 345 + ], + "score": 0.72, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 332, + 506, + 348 + ], + "score": 1.0, + "content": ", is the only supervision needed to optimize model and likelihood parameters,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 343, + 504, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 113, + 355 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 343, + 130, + 360 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 130, + 345, + 138, + 356 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 343, + 398, + 360 + ], + "score": 1.0, + "content": "respectively. Additionally, though the shape and scale varies with", + "type": "text" + }, + { + "bbox": [ + 399, + 345, + 406, + 357 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 343, + 482, + 360 + ], + "score": 1.0, + "content": ", reducing the error", + "type": "text" + }, + { + "bbox": [ + 482, + 345, + 504, + 357 + ], + "score": 0.9, + "content": "{ \\hat { y } } - y", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 356, + 290, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 290, + 368 + ], + "score": 1.0, + "content": "always reduces the NLL for our distributions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 372, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 104, + 371, + 504, + 386 + ], + "spans": [ + { + "bbox": [ + 104, + 371, + 496, + 386 + ], + "score": 1.0, + "content": "Distributions Under Investigation This work considers the normal likelihood with variance", + "type": "text" + }, + { + "bbox": [ + 496, + 375, + 504, + 383 + ], + "score": 0.7, + "content": "\\sigma", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 438, + 396 + ], + "score": 1.0, + "content": "(Bishop et al., 2006; Hastie et al., 2009), the softmax likelihood with temperature", + "type": "text" + }, + { + "bbox": [ + 438, + 386, + 445, + 394 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "(Hinton et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 393, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 235, + 407 + ], + "score": 1.0, + "content": "2015), and the robust likelihood", + "type": "text" + }, + { + "bbox": [ + 235, + 396, + 241, + 406 + ], + "score": 0.79, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 393, + 348, + 407 + ], + "score": 1.0, + "content": "(Barron, 2019) with shape", + "type": "text" + }, + { + "bbox": [ + 348, + 397, + 356, + 405 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 393, + 395, + 407 + ], + "score": 1.0, + "content": "and scale", + "type": "text" + }, + { + "bbox": [ + 396, + 397, + 403, + 405 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 393, + 506, + 407 + ], + "score": 1.0, + "content": "that control the scale and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "shape of the likelihood. The first two are among the most common losses in machine learning, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "the last loss provides an important illustration of a likelihood parameter that affects “shape” instead", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 428, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 444 + ], + "score": 1.0, + "content": "of “scale”. We note that changing the scale and shape of the likelihood distribution is not “cheating”", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "as there is a trade-off between uncertainty and credit. Figure 1 shows how this trade-off affects the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 451, + 309, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 309, + 464 + ], + "score": 1.0, + "content": "Normal and softmax distributions and their NLLs.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 414, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 467, + 413, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 301, + 482 + ], + "score": 1.0, + "content": "The normal likelihood has terms for the residual", + "type": "text" + }, + { + "bbox": [ + 301, + 469, + 325, + 480 + ], + "score": 0.91, + "content": "{ \\hat { y } } - y", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 467, + 394, + 482 + ], + "score": 1.0, + "content": "and the variance", + "type": "text" + }, + { + "bbox": [ + 394, + 471, + 401, + 479 + ], + "score": 0.79, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 467, + 413, + 482 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 486, + 399, + 514 + ], + "lines": [ + { + "bbox": [ + 210, + 486, + 399, + 514 + ], + "spans": [ + { + "bbox": [ + 210, + 486, + 399, + 514 + ], + "score": 0.94, + "content": "\\mathcal { N } ( \\hat { y } | y , \\sigma ) = ( 2 \\pi \\sigma ^ { 2 } ) ^ { - \\frac 1 2 } \\exp \\left( - \\frac 1 2 \\frac { ( \\hat { y } - y ) ^ { 2 } } { \\sigma ^ { 2 } } \\right) ,", + "type": "interline_equation", + "image_path": "3c46149cbf1a7efde4f1f949b21aafb63604e38b2e4c71197321aa8be8dab306.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 210, + 486, + 399, + 514 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 520, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 103, + 516, + 502, + 537 + ], + "spans": [ + { + "bbox": [ + 103, + 516, + 126, + 537 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 520, + 173, + 532 + ], + "score": 0.93, + "content": "\\sigma \\in ( 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 516, + 397, + 537 + ], + "score": 1.0, + "content": "scaling the distribution. The normal NLL can be written", + "type": "text" + }, + { + "bbox": [ + 397, + 519, + 502, + 533 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\ell _ { N } = \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } - y ) ^ { 2 } + \\log \\sigma , } \\end{array}", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "after simplifying and omitting constants that do not affect minimization. We recover the squared", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 542, + 220, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 190, + 555 + ], + "score": 1.0, + "content": "error by substituting", + "type": "text" + }, + { + "bbox": [ + 190, + 543, + 215, + 552 + ], + "score": 0.89, + "content": "\\sigma = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 542, + 220, + 555 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 443, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 445, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 362, + 572 + ], + "score": 1.0, + "content": "The softmax defines a categorical distribution defined by scores", + "type": "text" + }, + { + "bbox": [ + 363, + 562, + 369, + 569 + ], + "score": 0.79, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 558, + 427, + 572 + ], + "score": 1.0, + "content": "for each class", + "type": "text" + }, + { + "bbox": [ + 427, + 562, + 433, + 569 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 558, + 445, + 572 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 576, + 374, + 603 + ], + "lines": [ + { + "bbox": [ + 236, + 576, + 374, + 603 + ], + "spans": [ + { + "bbox": [ + 236, + 576, + 374, + 603 + ], + "score": 0.93, + "content": "\\mathrm { s o f t m a x } ( \\hat { y } = y | z , \\tau ) = \\frac { e ^ { z _ { y } \\tau } } { \\sum _ { c } e ^ { z _ { c } \\tau } } ,", + "type": "interline_equation", + "image_path": "972e7ded4f3db2c759825046dfc1de10135f752823d6c1e5f59c9f683b2cf5ad.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 236, + 576, + 374, + 603 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 608, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 194, + 622 + ], + "score": 1.0, + "content": "with the temperature,", + "type": "text" + }, + { + "bbox": [ + 195, + 609, + 240, + 621 + ], + "score": 0.92, + "content": "\\tau \\in ( 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 608, + 505, + 622 + ], + "score": 1.0, + "content": ", adjusting the entropy of the distribution. We recover the classifi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 210, + 632 + ], + "score": 1.0, + "content": "cation cross-entropy loss,", + "type": "text" + }, + { + "bbox": [ + 210, + 620, + 272, + 632 + ], + "score": 0.9, + "content": "- \\log p ( \\hat { y } = y )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 620, + 336, + 632 + ], + "score": 1.0, + "content": ", by substituting", + "type": "text" + }, + { + "bbox": [ + 337, + 621, + 362, + 630 + ], + "score": 0.9, + "content": "\\tau = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "in the respective NLL. We state the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 630, + 470, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 308, + 644 + ], + "score": 1.0, + "content": "gradients of these likelihoods with respect to their", + "type": "text" + }, + { + "bbox": [ + 308, + 633, + 315, + 641 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 630, + 333, + 644 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 333, + 633, + 340, + 641 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 630, + 470, + 644 + ], + "score": 1.0, + "content": "in Section A of the supplement.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 108, + 648, + 263, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 263, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 169, + 660 + ], + "score": 1.0, + "content": "The robust loss", + "type": "text" + }, + { + "bbox": [ + 170, + 650, + 176, + 659 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 647, + 263, + 660 + ], + "score": 1.0, + "content": "and its likelihood are", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 664, + 421, + 731 + ], + "lines": [ + { + "bbox": [ + 191, + 664, + 421, + 731 + ], + "spans": [ + { + "bbox": [ + 191, + 664, + 421, + 731 + ], + "score": 0.95, + "content": "{ \\begin{array} { r l } & { \\rho \\left( x , \\alpha , \\sigma \\right) = { \\frac { \\left| \\alpha - 2 \\right| } { \\alpha } } \\left( \\left( { \\frac { \\left( x / \\sigma \\right) ^ { 2 } } { \\left| \\alpha - 2 \\right| } } + 1 \\right) ^ { \\alpha / 2 } - 1 \\right) { \\mathrm { a n d } } } \\\\ & { p \\left( { \\hat { y } } \\mid y , \\alpha , \\sigma \\right) = { \\frac { 1 } { \\sigma Z \\left( \\alpha \\right) } } \\exp \\left( - \\rho \\left( { \\hat { y } } - y , \\alpha , \\sigma \\right) \\right) , } \\end{array} }", + "type": "interline_equation", + "image_path": "06652b49497df2b81c83be6eceb21ed82ceeec1780ea4c0aaed390fe4ed31d95.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 664, + 421, + 680.75 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 191, + 680.75, + 421, + 697.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 191, + 697.5, + 421, + 714.25 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 191, + 714.25, + 421, + 731.0 + ], + "spans": [], + "index": 36 + } + ] + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 83, + 499, + 190 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 83, + 499, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 83, + 499, + 190 + ], + "spans": [ + { + "bbox": [ + 110, + 83, + 499, + 190 + ], + "score": 0.965, + "type": "image", + "image_path": "b1ea4f3c37fb9dbc501c429bffa66313c780d218dcb161499e075885f826da71.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 83, + 499, + 118.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 118.66666666666666, + 499, + 154.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 154.33333333333331, + 499, + 189.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 202, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 200, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 216 + ], + "score": 1.0, + "content": "Figure 1: Optimizing likelihood parameters adapts the loss without manual hyperparameter tuning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 212, + 244, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 244, + 226 + ], + "score": 1.0, + "content": "to balance accuracy and certainty.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 246, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 161, + 260 + ], + "score": 1.0, + "content": "A likelihood", + "type": "text" + }, + { + "bbox": [ + 162, + 246, + 201, + 259 + ], + "score": 0.93, + "content": "\\mathcal { L } ( \\hat { y } | y , \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 246, + 455, + 260 + ], + "score": 1.0, + "content": "measures the quality of the prediction as a distribution over", + "type": "text" + }, + { + "bbox": [ + 456, + 247, + 462, + 258 + ], + "score": 0.82, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 246, + 505, + 260 + ], + "score": 1.0, + "content": "given the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 131, + 270 + ], + "score": 1.0, + "content": "target", + "type": "text" + }, + { + "bbox": [ + 131, + 259, + 138, + 269 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 257, + 244, + 270 + ], + "score": 1.0, + "content": "and likelihood parameters", + "type": "text" + }, + { + "bbox": [ + 244, + 258, + 251, + 269 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 257, + 394, + 270 + ], + "score": 1.0, + "content": ". We use the negative log-likelihood", + "type": "text" + }, + { + "bbox": [ + 394, + 258, + 400, + 268 + ], + "score": 0.68, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 257, + 506, + 270 + ], + "score": 1.0, + "content": "(NLL), and the likelihood", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 269, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 479, + 281 + ], + "score": 1.0, + "content": "interchangeably since both have the same optima. We define the full likelihood optimization:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 246, + 506, + 281 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 285, + 369, + 306 + ], + "lines": [ + { + "bbox": [ + 244, + 285, + 369, + 306 + ], + "spans": [ + { + "bbox": [ + 244, + 285, + 369, + 306 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta , \\phi } \\quad \\mathbb { E } _ { \\mathbf { \\Phi } ( x , y ) \\sim \\mathcal { D } } \\ell ( \\hat { y } = f _ { \\theta } ( x ) | y , \\phi )", + "type": "interline_equation", + "image_path": "6b7c1f81e8f579b80c819e0155429a965350c54767ecabb1a4d0fead8f057d8a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 244, + 285, + 369, + 306 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 433, + 324 + ], + "score": 1.0, + "content": "to jointly learn model and likelihood parameters. “Full” indicates the inclusion of", + "type": "text" + }, + { + "bbox": [ + 433, + 312, + 440, + 324 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 312, + 505, + 324 + ], + "score": 1.0, + "content": ", which controls", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 335 + ], + "score": 1.0, + "content": "the distribution and induced NLL loss. We focus on full likelihood optimization in this work. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 332, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 104, + 332, + 186, + 348 + ], + "score": 1.0, + "content": "note that the target,", + "type": "text" + }, + { + "bbox": [ + 187, + 336, + 193, + 345 + ], + "score": 0.72, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 332, + 506, + 348 + ], + "score": 1.0, + "content": ", is the only supervision needed to optimize model and likelihood parameters,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 343, + 504, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 113, + 355 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 343, + 130, + 360 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 130, + 345, + 138, + 356 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 343, + 398, + 360 + ], + "score": 1.0, + "content": "respectively. Additionally, though the shape and scale varies with", + "type": "text" + }, + { + "bbox": [ + 399, + 345, + 406, + 357 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 343, + 482, + 360 + ], + "score": 1.0, + "content": ", reducing the error", + "type": "text" + }, + { + "bbox": [ + 482, + 345, + 504, + 357 + ], + "score": 0.9, + "content": "{ \\hat { y } } - y", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 356, + 290, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 290, + 368 + ], + "score": 1.0, + "content": "always reduces the NLL for our distributions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 104, + 312, + 506, + 368 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 372, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 104, + 371, + 504, + 386 + ], + "spans": [ + { + "bbox": [ + 104, + 371, + 496, + 386 + ], + "score": 1.0, + "content": "Distributions Under Investigation This work considers the normal likelihood with variance", + "type": "text" + }, + { + "bbox": [ + 496, + 375, + 504, + 383 + ], + "score": 0.7, + "content": "\\sigma", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 438, + 396 + ], + "score": 1.0, + "content": "(Bishop et al., 2006; Hastie et al., 2009), the softmax likelihood with temperature", + "type": "text" + }, + { + "bbox": [ + 438, + 386, + 445, + 394 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "(Hinton et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 393, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 235, + 407 + ], + "score": 1.0, + "content": "2015), and the robust likelihood", + "type": "text" + }, + { + "bbox": [ + 235, + 396, + 241, + 406 + ], + "score": 0.79, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 393, + 348, + 407 + ], + "score": 1.0, + "content": "(Barron, 2019) with shape", + "type": "text" + }, + { + "bbox": [ + 348, + 397, + 356, + 405 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 393, + 395, + 407 + ], + "score": 1.0, + "content": "and scale", + "type": "text" + }, + { + "bbox": [ + 396, + 397, + 403, + 405 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 393, + 506, + 407 + ], + "score": 1.0, + "content": "that control the scale and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "shape of the likelihood. The first two are among the most common losses in machine learning, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "the last loss provides an important illustration of a likelihood parameter that affects “shape” instead", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 428, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 444 + ], + "score": 1.0, + "content": "of “scale”. We note that changing the scale and shape of the likelihood distribution is not “cheating”", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "as there is a trade-off between uncertainty and credit. Figure 1 shows how this trade-off affects the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 451, + 309, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 309, + 464 + ], + "score": 1.0, + "content": "Normal and softmax distributions and their NLLs.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 371, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 414, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 467, + 413, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 301, + 482 + ], + "score": 1.0, + "content": "The normal likelihood has terms for the residual", + "type": "text" + }, + { + "bbox": [ + 301, + 469, + 325, + 480 + ], + "score": 0.91, + "content": "{ \\hat { y } } - y", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 467, + 394, + 482 + ], + "score": 1.0, + "content": "and the variance", + "type": "text" + }, + { + "bbox": [ + 394, + 471, + 401, + 479 + ], + "score": 0.79, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 467, + 413, + 482 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 467, + 413, + 482 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 486, + 399, + 514 + ], + "lines": [ + { + "bbox": [ + 210, + 486, + 399, + 514 + ], + "spans": [ + { + "bbox": [ + 210, + 486, + 399, + 514 + ], + "score": 0.94, + "content": "\\mathcal { N } ( \\hat { y } | y , \\sigma ) = ( 2 \\pi \\sigma ^ { 2 } ) ^ { - \\frac 1 2 } \\exp \\left( - \\frac 1 2 \\frac { ( \\hat { y } - y ) ^ { 2 } } { \\sigma ^ { 2 } } \\right) ,", + "type": "interline_equation", + "image_path": "3c46149cbf1a7efde4f1f949b21aafb63604e38b2e4c71197321aa8be8dab306.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 210, + 486, + 399, + 514 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 520, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 103, + 516, + 502, + 537 + ], + "spans": [ + { + "bbox": [ + 103, + 516, + 126, + 537 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 520, + 173, + 532 + ], + "score": 0.93, + "content": "\\sigma \\in ( 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 516, + 397, + 537 + ], + "score": 1.0, + "content": "scaling the distribution. The normal NLL can be written", + "type": "text" + }, + { + "bbox": [ + 397, + 519, + 502, + 533 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\ell _ { N } = \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } - y ) ^ { 2 } + \\log \\sigma , } \\end{array}", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "after simplifying and omitting constants that do not affect minimization. We recover the squared", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 542, + 220, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 190, + 555 + ], + "score": 1.0, + "content": "error by substituting", + "type": "text" + }, + { + "bbox": [ + 190, + 543, + 215, + 552 + ], + "score": 0.89, + "content": "\\sigma = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 542, + 220, + 555 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 103, + 516, + 505, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 443, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 445, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 362, + 572 + ], + "score": 1.0, + "content": "The softmax defines a categorical distribution defined by scores", + "type": "text" + }, + { + "bbox": [ + 363, + 562, + 369, + 569 + ], + "score": 0.79, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 558, + 427, + 572 + ], + "score": 1.0, + "content": "for each class", + "type": "text" + }, + { + "bbox": [ + 427, + 562, + 433, + 569 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 558, + 445, + 572 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 558, + 445, + 572 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 576, + 374, + 603 + ], + "lines": [ + { + "bbox": [ + 236, + 576, + 374, + 603 + ], + "spans": [ + { + "bbox": [ + 236, + 576, + 374, + 603 + ], + "score": 0.93, + "content": "\\mathrm { s o f t m a x } ( \\hat { y } = y | z , \\tau ) = \\frac { e ^ { z _ { y } \\tau } } { \\sum _ { c } e ^ { z _ { c } \\tau } } ,", + "type": "interline_equation", + "image_path": "972e7ded4f3db2c759825046dfc1de10135f752823d6c1e5f59c9f683b2cf5ad.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 236, + 576, + 374, + 603 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 608, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 194, + 622 + ], + "score": 1.0, + "content": "with the temperature,", + "type": "text" + }, + { + "bbox": [ + 195, + 609, + 240, + 621 + ], + "score": 0.92, + "content": "\\tau \\in ( 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 608, + 505, + 622 + ], + "score": 1.0, + "content": ", adjusting the entropy of the distribution. We recover the classifi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 210, + 632 + ], + "score": 1.0, + "content": "cation cross-entropy loss,", + "type": "text" + }, + { + "bbox": [ + 210, + 620, + 272, + 632 + ], + "score": 0.9, + "content": "- \\log p ( \\hat { y } = y )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 620, + 336, + 632 + ], + "score": 1.0, + "content": ", by substituting", + "type": "text" + }, + { + "bbox": [ + 337, + 621, + 362, + 630 + ], + "score": 0.9, + "content": "\\tau = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "in the respective NLL. We state the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 630, + 470, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 308, + 644 + ], + "score": 1.0, + "content": "gradients of these likelihoods with respect to their", + "type": "text" + }, + { + "bbox": [ + 308, + 633, + 315, + 641 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 630, + 333, + 644 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 333, + 633, + 340, + 641 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 630, + 470, + 644 + ], + "score": 1.0, + "content": "in Section A of the supplement.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 608, + 505, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 648, + 263, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 263, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 169, + 660 + ], + "score": 1.0, + "content": "The robust loss", + "type": "text" + }, + { + "bbox": [ + 170, + 650, + 176, + 659 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 647, + 263, + 660 + ], + "score": 1.0, + "content": "and its likelihood are", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 647, + 263, + 660 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 664, + 421, + 731 + ], + "lines": [ + { + "bbox": [ + 191, + 664, + 421, + 731 + ], + "spans": [ + { + "bbox": [ + 191, + 664, + 421, + 731 + ], + "score": 0.95, + "content": "{ \\begin{array} { r l } & { \\rho \\left( x , \\alpha , \\sigma \\right) = { \\frac { \\left| \\alpha - 2 \\right| } { \\alpha } } \\left( \\left( { \\frac { \\left( x / \\sigma \\right) ^ { 2 } } { \\left| \\alpha - 2 \\right| } } + 1 \\right) ^ { \\alpha / 2 } - 1 \\right) { \\mathrm { a n d } } } \\\\ & { p \\left( { \\hat { y } } \\mid y , \\alpha , \\sigma \\right) = { \\frac { 1 } { \\sigma Z \\left( \\alpha \\right) } } \\exp \\left( - \\rho \\left( { \\hat { y } } - y , \\alpha , \\sigma \\right) \\right) , } \\end{array} }", + "type": "interline_equation", + "image_path": "06652b49497df2b81c83be6eceb21ed82ceeec1780ea4c0aaed390fe4ed31d95.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 664, + 421, + 680.75 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 191, + 680.75, + 421, + 697.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 191, + 697.5, + 421, + 714.25 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 191, + 714.25, + 421, + 731.0 + ], + "spans": [], + "index": 36 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 153, + 96 + ], + "score": 1.0, + "content": "with shape", + "type": "text" + }, + { + "bbox": [ + 153, + 82, + 201, + 95 + ], + "score": 0.92, + "content": "\\alpha \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 81, + 228, + 96 + ], + "score": 1.0, + "content": ", scale", + "type": "text" + }, + { + "bbox": [ + 229, + 82, + 277, + 95 + ], + "score": 0.93, + "content": "\\sigma \\in ( 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 81, + 394, + 96 + ], + "score": 1.0, + "content": ", and normalization function", + "type": "text" + }, + { + "bbox": [ + 394, + 82, + 419, + 95 + ], + "score": 0.91, + "content": "Z \\left( \\alpha \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 81, + 494, + 96 + ], + "score": 1.0, + "content": ". This robust loss,", + "type": "text" + }, + { + "bbox": [ + 495, + 85, + 502, + 94 + ], + "score": 0.47, + "content": "\\rho _ { ; }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "has the interesting property that it generalizes several different loss functions commonly used in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 253, + 120 + ], + "score": 1.0, + "content": "robust learning such as the L2 loss", + "type": "text" + }, + { + "bbox": [ + 253, + 107, + 281, + 118 + ], + "score": 0.84, + "content": "( \\alpha = 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 106, + 471, + 120 + ], + "score": 1.0, + "content": ", pseudo-huber loss (Charbonnier et al., 1997)", + "type": "text" + }, + { + "bbox": [ + 471, + 107, + 501, + 118 + ], + "score": 0.52, + "content": "( \\alpha = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 106, + 505, + 120 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 118, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 225, + 133 + ], + "score": 1.0, + "content": "Cauchy loss (Li et al., 2018)", + "type": "text" + }, + { + "bbox": [ + 225, + 119, + 257, + 131 + ], + "score": 0.83, + "content": "( \\alpha = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 118, + 461, + 133 + ], + "score": 1.0, + "content": ", Geman-McClure loss (Ganan & McClure, 1985),", + "type": "text" + }, + { + "bbox": [ + 461, + 119, + 501, + 131 + ], + "score": 0.85, + "content": "( \\alpha = - 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 118, + 505, + 133 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 288, + 144 + ], + "score": 1.0, + "content": "and Welsch (Dennis Jr & Welsch, 1978) loss", + "type": "text" + }, + { + "bbox": [ + 289, + 131, + 351, + 143 + ], + "score": 0.87, + "content": "( a l p h a = - \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 131, + 505, + 144 + ], + "score": 1.0, + "content": ". Learning the shape parameter allows", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 319, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 319, + 157 + ], + "score": 1.0, + "content": "models to adapt the shape of their noise distribution.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 170, + 211, + 183 + ], + "lines": [ + { + "bbox": [ + 104, + 169, + 213, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 213, + 186 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 195, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "score": 1.0, + "content": "Likelihood optimization follows from maximum likelihood estimation (Hastie et al., 2009; Bishop", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 218 + ], + "score": 1.0, + "content": "et al., 2006), yet is uncommon in practice for fitting deep regressors and classifiers for discriminative", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 218, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 229 + ], + "score": 1.0, + "content": "tasks. However Kendall & Gal (2017); Kendall et al. (2018); Barron (2019); Saxena et al. (2019)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "score": 1.0, + "content": "optimize likelihood parameters to their advantage yet differ in their tasks, likelihoods, and param-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "eterizations. In this work we aim to systematically experiment, clarify usage, and encourage their", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 251, + 171, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 171, + 263 + ], + "score": 1.0, + "content": "wider adoption.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "score": 1.0, + "content": "Early work on regressing means and variances (Nix & Weigend, 1994) had the key insight that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "optimizing the full likelihood can fit these parameters and adapt the loss. Some recent works use", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "likelihoods for loss adaptation, and interpret their parameters as the uncertainty (Kendall & Gal,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "2017; Kendall et al., 2018), robustness (Kendall & Gal, 2017; Barron, 2019; Saxena et al., 2019),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 324 + ], + "score": 1.0, + "content": "and curricula (Saxena et al., 2019) of losses. MacKay & Mac Kay (2003) uses Bayesian evidence to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "select hyper-parameters and losses based on proper likelihood normalization. Barron (2019) define", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 256, + 346 + ], + "score": 1.0, + "content": "a generalized robust regression loss,", + "type": "text" + }, + { + "bbox": [ + 256, + 335, + 263, + 345 + ], + "score": 0.73, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 333, + 506, + 346 + ], + "score": 1.0, + "content": ", to jointly optimize the type and degree of robustness with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "global, data-independent, parameters. Kendall & Gal (2017) predict variances for regression and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "classification to handle data-dependent uncertainty. Kendall et al. (2018) balance multi-task loss", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "weights by optimizing variances for regression and temperatures for classification. These global", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "parameters depend on the task but not the data, and are interpreted as inherent task uncertainty.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "Saxena et al. (2019) define a differentiable curriculum for classification by assigning each training", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "point its own temperature. These data parameters depend on the index of the data but not its value.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 410, + 455, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 455, + 421 + ], + "score": 1.0, + "content": "We compare these different likelihood parameterizations across tasks and distributions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "In the calibration literature, Guo et al. (2017) have found that deep networks are often miscalibrated,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 436, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 450 + ], + "score": 1.0, + "content": "but they can be re-calibrated by cross-validating the temperature of the softmax. In this work we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 447, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 463 + ], + "score": 1.0, + "content": "explore several generalizations of this concept. Alternatively, Platt scaling (Platt, 1999) fits a sig-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "moid regressor to model predictions to calibrate probabilities. Kuleshov et al. (2018) re-calibrate", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 471, + 471, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 471, + 483 + ], + "score": 1.0, + "content": "regressors by fitting an Isotonic regressor to the empirical cumulative distribution function.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 295, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 296, + 511 + ], + "score": 1.0, + "content": "4 LIKELIHOOD PARAMETER TYPES", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "We explore the space of likelihood parameter representations for model optimization and inference.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "Though we note that some losses, like adversarial losses, are difficult to represent as likelihoods,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "many different losses in the community have a natural probabilistic interpretation. Often, these", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "probabilistic interpretations can be parametrized in a variety of ways. We explore two key axes of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 425, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 425, + 579 + ], + "score": 1.0, + "content": "generality when building these loss functions: conditioning and dimensionality.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "Conditioning We represent the likelihood parameters by three functional classes: global, data, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 231, + 607 + ], + "score": 1.0, + "content": "predicted. Global parameters,", + "type": "text" + }, + { + "bbox": [ + 231, + 594, + 258, + 606 + ], + "score": 0.92, + "content": "\\phi = c", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 594, + 506, + 607 + ], + "score": 1.0, + "content": ", are independent of the data and model and define the same", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 331, + 618 + ], + "score": 1.0, + "content": "likelihood distribution for all points. Data parameters,", + "type": "text" + }, + { + "bbox": [ + 332, + 605, + 342, + 617 + ], + "score": 0.88, + "content": "\\phi _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 605, + 469, + 618 + ], + "score": 1.0, + "content": ", are conditioned on the index,", + "type": "text" + }, + { + "bbox": [ + 470, + 606, + 474, + 615 + ], + "score": 0.62, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 605, + 506, + 618 + ], + "score": 1.0, + "content": ", of the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 615, + 504, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 128, + 629 + ], + "score": 1.0, + "content": "data,", + "type": "text" + }, + { + "bbox": [ + 128, + 617, + 138, + 627 + ], + "score": 0.84, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 615, + 493, + 629 + ], + "score": 1.0, + "content": ", but not its value. Every training point is assigned an independent likelihood parameter,", + "type": "text" + }, + { + "bbox": [ + 493, + 616, + 504, + 627 + ], + "score": 0.86, + "content": "\\phi _ { i }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 426, + 640 + ], + "score": 1.0, + "content": "that define different likelihoods for each training point. Predicted parameters,", + "type": "text" + }, + { + "bbox": [ + 426, + 627, + 486, + 639 + ], + "score": 0.94, + "content": "\\grave { \\phi ( x ) = { g _ { \\eta } ( x ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 626, + 506, + 640 + ], + "score": 1.0, + "content": ", are", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 637, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 205, + 652 + ], + "score": 1.0, + "content": "determined by a model,", + "type": "text" + }, + { + "bbox": [ + 206, + 640, + 212, + 649 + ], + "score": 0.73, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 637, + 284, + 652 + ], + "score": 1.0, + "content": ", with parameters", + "type": "text" + }, + { + "bbox": [ + 285, + 640, + 291, + 649 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 637, + 506, + 652 + ], + "score": 1.0, + "content": "(not to be confused with the task model parameters", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 113, + 659 + ], + "score": 0.43, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "). Global and predicted parameters can be used during training and testing, but data parameters are", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "only assigned to each training point and are undefined for testing. We show a simple example of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 670, + 459, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 459, + 684 + ], + "score": 1.0, + "content": "predicted temperature in Figure 4, and an illustration of the parameter types in Figure 2.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "We note that for certain global parameters like a learned Normal scale, changing the scale does not", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "affect the optima, but does change the probabilistic interpretation. This invariance has led many", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "authors to drop the scale from their formulations. However, when models can predict these scale", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "parameters they can naturally remain calibrated in the presence of heteroskedasticity and outliers.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 153, + 96 + ], + "score": 1.0, + "content": "with shape", + "type": "text" + }, + { + "bbox": [ + 153, + 82, + 201, + 95 + ], + "score": 0.92, + "content": "\\alpha \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 81, + 228, + 96 + ], + "score": 1.0, + "content": ", scale", + "type": "text" + }, + { + "bbox": [ + 229, + 82, + 277, + 95 + ], + "score": 0.93, + "content": "\\sigma \\in ( 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 81, + 394, + 96 + ], + "score": 1.0, + "content": ", and normalization function", + "type": "text" + }, + { + "bbox": [ + 394, + 82, + 419, + 95 + ], + "score": 0.91, + "content": "Z \\left( \\alpha \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 81, + 494, + 96 + ], + "score": 1.0, + "content": ". This robust loss,", + "type": "text" + }, + { + "bbox": [ + 495, + 85, + 502, + 94 + ], + "score": 0.47, + "content": "\\rho _ { ; }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "has the interesting property that it generalizes several different loss functions commonly used in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 253, + 120 + ], + "score": 1.0, + "content": "robust learning such as the L2 loss", + "type": "text" + }, + { + "bbox": [ + 253, + 107, + 281, + 118 + ], + "score": 0.84, + "content": "( \\alpha = 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 106, + 471, + 120 + ], + "score": 1.0, + "content": ", pseudo-huber loss (Charbonnier et al., 1997)", + "type": "text" + }, + { + "bbox": [ + 471, + 107, + 501, + 118 + ], + "score": 0.52, + "content": "( \\alpha = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 106, + 505, + 120 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 118, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 225, + 133 + ], + "score": 1.0, + "content": "Cauchy loss (Li et al., 2018)", + "type": "text" + }, + { + "bbox": [ + 225, + 119, + 257, + 131 + ], + "score": 0.83, + "content": "( \\alpha = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 118, + 461, + 133 + ], + "score": 1.0, + "content": ", Geman-McClure loss (Ganan & McClure, 1985),", + "type": "text" + }, + { + "bbox": [ + 461, + 119, + 501, + 131 + ], + "score": 0.85, + "content": "( \\alpha = - 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 118, + 505, + 133 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 288, + 144 + ], + "score": 1.0, + "content": "and Welsch (Dennis Jr & Welsch, 1978) loss", + "type": "text" + }, + { + "bbox": [ + 289, + 131, + 351, + 143 + ], + "score": 0.87, + "content": "( a l p h a = - \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 131, + 505, + 144 + ], + "score": 1.0, + "content": ". Learning the shape parameter allows", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 319, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 319, + 157 + ], + "score": 1.0, + "content": "models to adapt the shape of their noise distribution.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 81, + 506, + 157 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 170, + 211, + 183 + ], + "lines": [ + { + "bbox": [ + 104, + 169, + 213, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 213, + 186 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 195, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 208 + ], + "score": 1.0, + "content": "Likelihood optimization follows from maximum likelihood estimation (Hastie et al., 2009; Bishop", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 218 + ], + "score": 1.0, + "content": "et al., 2006), yet is uncommon in practice for fitting deep regressors and classifiers for discriminative", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 218, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 229 + ], + "score": 1.0, + "content": "tasks. However Kendall & Gal (2017); Kendall et al. (2018); Barron (2019); Saxena et al. (2019)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "score": 1.0, + "content": "optimize likelihood parameters to their advantage yet differ in their tasks, likelihoods, and param-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "eterizations. In this work we aim to systematically experiment, clarify usage, and encourage their", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 251, + 171, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 171, + 263 + ], + "score": 1.0, + "content": "wider adoption.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 194, + 506, + 263 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "score": 1.0, + "content": "Early work on regressing means and variances (Nix & Weigend, 1994) had the key insight that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "optimizing the full likelihood can fit these parameters and adapt the loss. Some recent works use", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "likelihoods for loss adaptation, and interpret their parameters as the uncertainty (Kendall & Gal,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "2017; Kendall et al., 2018), robustness (Kendall & Gal, 2017; Barron, 2019; Saxena et al., 2019),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 324 + ], + "score": 1.0, + "content": "and curricula (Saxena et al., 2019) of losses. MacKay & Mac Kay (2003) uses Bayesian evidence to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "select hyper-parameters and losses based on proper likelihood normalization. Barron (2019) define", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 256, + 346 + ], + "score": 1.0, + "content": "a generalized robust regression loss,", + "type": "text" + }, + { + "bbox": [ + 256, + 335, + 263, + 345 + ], + "score": 0.73, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 333, + 506, + 346 + ], + "score": 1.0, + "content": ", to jointly optimize the type and degree of robustness with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "global, data-independent, parameters. Kendall & Gal (2017) predict variances for regression and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "classification to handle data-dependent uncertainty. Kendall et al. (2018) balance multi-task loss", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "weights by optimizing variances for regression and temperatures for classification. These global", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "parameters depend on the task but not the data, and are interpreted as inherent task uncertainty.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "Saxena et al. (2019) define a differentiable curriculum for classification by assigning each training", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "point its own temperature. These data parameters depend on the index of the data but not its value.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 410, + 455, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 455, + 421 + ], + "score": 1.0, + "content": "We compare these different likelihood parameterizations across tasks and distributions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 268, + 506, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "In the calibration literature, Guo et al. (2017) have found that deep networks are often miscalibrated,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 436, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 450 + ], + "score": 1.0, + "content": "but they can be re-calibrated by cross-validating the temperature of the softmax. In this work we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 447, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 463 + ], + "score": 1.0, + "content": "explore several generalizations of this concept. Alternatively, Platt scaling (Platt, 1999) fits a sig-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "moid regressor to model predictions to calibrate probabilities. Kuleshov et al. (2018) re-calibrate", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 471, + 471, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 471, + 483 + ], + "score": 1.0, + "content": "regressors by fitting an Isotonic regressor to the empirical cumulative distribution function.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 427, + 505, + 483 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 295, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 296, + 511 + ], + "score": 1.0, + "content": "4 LIKELIHOOD PARAMETER TYPES", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "We explore the space of likelihood parameter representations for model optimization and inference.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "Though we note that some losses, like adversarial losses, are difficult to represent as likelihoods,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "many different losses in the community have a natural probabilistic interpretation. Often, these", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "probabilistic interpretations can be parametrized in a variety of ways. We explore two key axes of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 425, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 425, + 579 + ], + "score": 1.0, + "content": "generality when building these loss functions: conditioning and dimensionality.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 522, + 506, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "Conditioning We represent the likelihood parameters by three functional classes: global, data, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 231, + 607 + ], + "score": 1.0, + "content": "predicted. Global parameters,", + "type": "text" + }, + { + "bbox": [ + 231, + 594, + 258, + 606 + ], + "score": 0.92, + "content": "\\phi = c", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 594, + 506, + 607 + ], + "score": 1.0, + "content": ", are independent of the data and model and define the same", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 331, + 618 + ], + "score": 1.0, + "content": "likelihood distribution for all points. Data parameters,", + "type": "text" + }, + { + "bbox": [ + 332, + 605, + 342, + 617 + ], + "score": 0.88, + "content": "\\phi _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 605, + 469, + 618 + ], + "score": 1.0, + "content": ", are conditioned on the index,", + "type": "text" + }, + { + "bbox": [ + 470, + 606, + 474, + 615 + ], + "score": 0.62, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 605, + 506, + 618 + ], + "score": 1.0, + "content": ", of the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 615, + 504, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 128, + 629 + ], + "score": 1.0, + "content": "data,", + "type": "text" + }, + { + "bbox": [ + 128, + 617, + 138, + 627 + ], + "score": 0.84, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 615, + 493, + 629 + ], + "score": 1.0, + "content": ", but not its value. Every training point is assigned an independent likelihood parameter,", + "type": "text" + }, + { + "bbox": [ + 493, + 616, + 504, + 627 + ], + "score": 0.86, + "content": "\\phi _ { i }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 426, + 640 + ], + "score": 1.0, + "content": "that define different likelihoods for each training point. Predicted parameters,", + "type": "text" + }, + { + "bbox": [ + 426, + 627, + 486, + 639 + ], + "score": 0.94, + "content": "\\grave { \\phi ( x ) = { g _ { \\eta } ( x ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 626, + 506, + 640 + ], + "score": 1.0, + "content": ", are", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 637, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 205, + 652 + ], + "score": 1.0, + "content": "determined by a model,", + "type": "text" + }, + { + "bbox": [ + 206, + 640, + 212, + 649 + ], + "score": 0.73, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 637, + 284, + 652 + ], + "score": 1.0, + "content": ", with parameters", + "type": "text" + }, + { + "bbox": [ + 285, + 640, + 291, + 649 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 637, + 506, + 652 + ], + "score": 1.0, + "content": "(not to be confused with the task model parameters", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 113, + 659 + ], + "score": 0.43, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "). Global and predicted parameters can be used during training and testing, but data parameters are", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "only assigned to each training point and are undefined for testing. We show a simple example of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 670, + 459, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 459, + 684 + ], + "score": 1.0, + "content": "predicted temperature in Figure 4, and an illustration of the parameter types in Figure 2.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 582, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "We note that for certain global parameters like a learned Normal scale, changing the scale does not", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "affect the optima, but does change the probabilistic interpretation. This invariance has led many", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "authors to drop the scale from their formulations. However, when models can predict these scale", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "parameters they can naturally remain calibrated in the presence of heteroskedasticity and outliers.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 83, + 299, + 153 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 83, + 299, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 83, + 299, + 153 + ], + "spans": [ + { + "bbox": [ + 112, + 83, + 299, + 153 + ], + "score": 0.956, + "type": "image", + "image_path": "c49dec68844ff1791035e87e8e2cdb47745b7e10693a93dd6a1ceca2a971d27a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 112, + 83, + 299, + 97.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 97.0, + 299, + 111.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 111.0, + 299, + 125.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 125.0, + 299, + 139.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 112, + 139.0, + 299, + 153.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 163, + 302, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 163, + 302, + 175 + ], + "spans": [ + { + "bbox": [ + 109, + 163, + 302, + 175 + ], + "score": 1.0, + "content": "Figure 2: Illustration of an image classifier with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 174, + 303, + 187 + ], + "spans": [ + { + "bbox": [ + 109, + 174, + 303, + 187 + ], + "score": 1.0, + "content": "three different types of likelihood temperature", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 184, + 303, + 198 + ], + "spans": [ + { + "bbox": [ + 109, + 184, + 303, + 198 + ], + "score": 1.0, + "content": "conditioning: global, predicted, and data. Each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 196, + 303, + 210 + ], + "spans": [ + { + "bbox": [ + 108, + 196, + 303, + 210 + ], + "score": 1.0, + "content": "represents a different way to parametrize the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 209, + 196, + 222 + ], + "spans": [ + { + "bbox": [ + 109, + 209, + 196, + 222 + ], + "score": 1.0, + "content": "model’s temperature.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "image", + "bbox": [ + 309, + 84, + 498, + 162 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 309, + 84, + 498, + 162 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 84, + 498, + 162 + ], + "spans": [ + { + "bbox": [ + 309, + 84, + 498, + 162 + ], + "score": 0.947, + "type": "image", + "image_path": "ad6ea3516753f68625ef40ead2859868d8abf970b4463448775e0270e5989662.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 309, + 84, + 498, + 99.6 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 309, + 99.6, + 498, + 115.19999999999999 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 309, + 115.19999999999999, + 498, + 130.79999999999998 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 309, + 130.79999999999998, + 498, + 146.39999999999998 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 309, + 146.39999999999998, + 498, + 161.99999999999997 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 308, + 172, + 501, + 218 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 171, + 502, + 184 + ], + "spans": [ + { + "bbox": [ + 309, + 171, + 502, + 184 + ], + "score": 1.0, + "content": "Figure 3: An image loss function with three", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 181, + 502, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 181, + 502, + 195 + ], + "score": 1.0, + "content": "different likelihood parameter dimensionalities.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 192, + 503, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 192, + 503, + 207 + ], + "score": 1.0, + "content": "Each represents a possible way to parametrize", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 205, + 502, + 219 + ], + "spans": [ + { + "bbox": [ + 309, + 205, + 502, + 219 + ], + "score": 1.0, + "content": "the additional scale parameter added to the loss.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + } + ], + "index": 14.25 + }, + { + "type": "image", + "bbox": [ + 107, + 235, + 505, + 369 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 235, + 505, + 369 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 235, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 107, + 235, + 505, + 369 + ], + "score": 0.965, + "type": "image", + "image_path": "6e6527859ba4e3af8468001a873c2fb1f83aa9e57ce95e50bc2c24f811b78155.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 107, + 235, + 505, + 279.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 107, + 279.6666666666667, + 505, + 324.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 107, + 324.33333333333337, + 505, + 369.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 372, + 505, + 419 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 370, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 385 + ], + "score": 1.0, + "content": "Figure 4: A synthetic logistic regression experiment. Regressing softmax temperature reduces the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "influence of outliers (blue, bottom-left), by locally raising temperature. The jointly optimized model", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "score": 1.0, + "content": "(center and right panel) achieves a more accurate classification that a model trained without adaptive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 406, + 206, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 206, + 420 + ], + "score": 1.0, + "content": "temperature (left panel).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + } + ], + "index": 21.75 + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 411, + 452 + ], + "score": 1.0, + "content": "Additionally we note that for the shape parameter of the robust likelihood,", + "type": "text" + }, + { + "bbox": [ + 411, + 442, + 418, + 451 + ], + "score": 0.74, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 439, + 506, + 452 + ], + "score": 1.0, + "content": ", changing global pa-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "rameters does affect model fitting. Previous works have adapted a global softmax temperature for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "model distillation (Hinton et al., 2015), and recalibration (Guo et al., 2017). Barron (2019) also", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "experiments with global values of loss function shape and scale parameters. The main work on Data", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "score": 1.0, + "content": "parameters is that of Saxena et al. (2019) who use these to learn a curriculum. Model-based param-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "eters appear in earlier work on regressing variance (Nix & Weigend, 1994), and more recent work", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 510, + 212, + 527 + ], + "spans": [ + { + "bbox": [ + 104, + 510, + 212, + 527 + ], + "score": 1.0, + "content": "by Kendall & Gal (2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 506, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 259, + 541 + ], + "score": 1.0, + "content": "Dimensionality The dimensionality,", + "type": "text" + }, + { + "bbox": [ + 259, + 529, + 271, + 541 + ], + "score": 0.89, + "content": "| \\phi |", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 528, + 506, + 541 + ], + "score": 1.0, + "content": ", of likelihood parameters can vary with the dimension of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 189, + 552 + ], + "score": 1.0, + "content": "the task prediction,", + "type": "text" + }, + { + "bbox": [ + 189, + 541, + 195, + 552 + ], + "score": 0.81, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 539, + 505, + 552 + ], + "score": 1.0, + "content": ". For example, image regressors can use a single likelihood parameter for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 156, + 564 + ], + "score": 1.0, + "content": "each image", + "type": "text" + }, + { + "bbox": [ + 156, + 551, + 191, + 563 + ], + "score": 0.91, + "content": "| \\phi | = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 550, + 282, + 564 + ], + "score": 1.0, + "content": ", RGB image channel", + "type": "text" + }, + { + "bbox": [ + 282, + 551, + 320, + 563 + ], + "score": 0.92, + "content": "| \\phi | = C", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 550, + 407, + 564 + ], + "score": 1.0, + "content": ", or even every pixel", + "type": "text" + }, + { + "bbox": [ + 407, + 551, + 492, + 563 + ], + "score": 0.91, + "content": "| \\phi | = \\bar { W \\times H \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "score": 1.0, + "content": "in Figure 3. These choices correspond to different likelihood distribution classes. Dimensionality", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 479, + 585 + ], + "score": 1.0, + "content": "and Conditioning of likelihood parameters can interact. For example, data parameters with", + "type": "text" + }, + { + "bbox": [ + 479, + 573, + 505, + 585 + ], + "score": 0.9, + "content": "| \\phi | \\overset { \\cdot } { = }", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 107, + 584, + 161, + 594 + ], + "score": 0.92, + "content": "W \\times H \\times C", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 583, + 226, + 595 + ], + "score": 1.0, + "content": "would result in", + "type": "text" + }, + { + "bbox": [ + 226, + 584, + 303, + 594 + ], + "score": 0.92, + "content": "N \\times W \\times H \\times C", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 583, + 424, + 595 + ], + "score": 1.0, + "content": "additional parameters, where", + "type": "text" + }, + { + "bbox": [ + 424, + 584, + 434, + 594 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "is the size of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "dataset. This can complicate implementations and slow down optimization due to disk I/O when", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "their size exceeds memory. Table 5 in the appendix contrasts the computational requirements of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "different likelihood parameter types. The work of Barron (2019) explores both scalar and pixel-wise", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 628, + 250, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 250, + 641 + ], + "score": 1.0, + "content": "dimensionalities for his robust loss.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 108, + 657, + 202, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 204, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 204, + 672 + ], + "score": 1.0, + "content": "5 APPLICATIONS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 306, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "4", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 83, + 299, + 153 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 83, + 299, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 83, + 299, + 153 + ], + "spans": [ + { + "bbox": [ + 112, + 83, + 299, + 153 + ], + "score": 0.956, + "type": "image", + "image_path": "c49dec68844ff1791035e87e8e2cdb47745b7e10693a93dd6a1ceca2a971d27a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 112, + 83, + 299, + 97.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 97.0, + 299, + 111.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 111.0, + 299, + 125.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 125.0, + 299, + 139.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 112, + 139.0, + 299, + 153.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 163, + 302, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 163, + 302, + 175 + ], + "spans": [ + { + "bbox": [ + 109, + 163, + 302, + 175 + ], + "score": 1.0, + "content": "Figure 2: Illustration of an image classifier with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 174, + 303, + 187 + ], + "spans": [ + { + "bbox": [ + 109, + 174, + 303, + 187 + ], + "score": 1.0, + "content": "three different types of likelihood temperature", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 184, + 303, + 198 + ], + "spans": [ + { + "bbox": [ + 109, + 184, + 303, + 198 + ], + "score": 1.0, + "content": "conditioning: global, predicted, and data. Each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 196, + 303, + 210 + ], + "spans": [ + { + "bbox": [ + 108, + 196, + 303, + 210 + ], + "score": 1.0, + "content": "represents a different way to parametrize the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 209, + 196, + 222 + ], + "spans": [ + { + "bbox": [ + 109, + 209, + 196, + 222 + ], + "score": 1.0, + "content": "model’s temperature.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "image", + "bbox": [ + 309, + 84, + 498, + 162 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 309, + 84, + 498, + 162 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 84, + 498, + 162 + ], + "spans": [ + { + "bbox": [ + 309, + 84, + 498, + 162 + ], + "score": 0.947, + "type": "image", + "image_path": "ad6ea3516753f68625ef40ead2859868d8abf970b4463448775e0270e5989662.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 309, + 84, + 498, + 99.6 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 309, + 99.6, + 498, + 115.19999999999999 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 309, + 115.19999999999999, + 498, + 130.79999999999998 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 309, + 130.79999999999998, + 498, + 146.39999999999998 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 309, + 146.39999999999998, + 498, + 161.99999999999997 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 308, + 172, + 501, + 218 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 309, + 171, + 502, + 184 + ], + "spans": [ + { + "bbox": [ + 309, + 171, + 502, + 184 + ], + "score": 1.0, + "content": "Figure 3: An image loss function with three", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 181, + 502, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 181, + 502, + 195 + ], + "score": 1.0, + "content": "different likelihood parameter dimensionalities.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 192, + 503, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 192, + 503, + 207 + ], + "score": 1.0, + "content": "Each represents a possible way to parametrize", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 205, + 502, + 219 + ], + "spans": [ + { + "bbox": [ + 309, + 205, + 502, + 219 + ], + "score": 1.0, + "content": "the additional scale parameter added to the loss.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + } + ], + "index": 14.25 + }, + { + "type": "image", + "bbox": [ + 107, + 235, + 505, + 369 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 235, + 505, + 369 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 235, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 107, + 235, + 505, + 369 + ], + "score": 0.965, + "type": "image", + "image_path": "6e6527859ba4e3af8468001a873c2fb1f83aa9e57ce95e50bc2c24f811b78155.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 107, + 235, + 505, + 279.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 107, + 279.6666666666667, + 505, + 324.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 107, + 324.33333333333337, + 505, + 369.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 372, + 505, + 419 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 370, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 385 + ], + "score": 1.0, + "content": "Figure 4: A synthetic logistic regression experiment. Regressing softmax temperature reduces the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "influence of outliers (blue, bottom-left), by locally raising temperature. The jointly optimized model", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "score": 1.0, + "content": "(center and right panel) achieves a more accurate classification that a model trained without adaptive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 406, + 206, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 206, + 420 + ], + "score": 1.0, + "content": "temperature (left panel).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + } + ], + "index": 21.75 + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 411, + 452 + ], + "score": 1.0, + "content": "Additionally we note that for the shape parameter of the robust likelihood,", + "type": "text" + }, + { + "bbox": [ + 411, + 442, + 418, + 451 + ], + "score": 0.74, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 439, + 506, + 452 + ], + "score": 1.0, + "content": ", changing global pa-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "rameters does affect model fitting. Previous works have adapted a global softmax temperature for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "model distillation (Hinton et al., 2015), and recalibration (Guo et al., 2017). Barron (2019) also", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "experiments with global values of loss function shape and scale parameters. The main work on Data", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "score": 1.0, + "content": "parameters is that of Saxena et al. (2019) who use these to learn a curriculum. Model-based param-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "eters appear in earlier work on regressing variance (Nix & Weigend, 1994), and more recent work", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 510, + 212, + 527 + ], + "spans": [ + { + "bbox": [ + 104, + 510, + 212, + 527 + ], + "score": 1.0, + "content": "by Kendall & Gal (2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 104, + 439, + 506, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 506, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 259, + 541 + ], + "score": 1.0, + "content": "Dimensionality The dimensionality,", + "type": "text" + }, + { + "bbox": [ + 259, + 529, + 271, + 541 + ], + "score": 0.89, + "content": "| \\phi |", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 528, + 506, + 541 + ], + "score": 1.0, + "content": ", of likelihood parameters can vary with the dimension of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 189, + 552 + ], + "score": 1.0, + "content": "the task prediction,", + "type": "text" + }, + { + "bbox": [ + 189, + 541, + 195, + 552 + ], + "score": 0.81, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 539, + 505, + 552 + ], + "score": 1.0, + "content": ". For example, image regressors can use a single likelihood parameter for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 156, + 564 + ], + "score": 1.0, + "content": "each image", + "type": "text" + }, + { + "bbox": [ + 156, + 551, + 191, + 563 + ], + "score": 0.91, + "content": "| \\phi | = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 550, + 282, + 564 + ], + "score": 1.0, + "content": ", RGB image channel", + "type": "text" + }, + { + "bbox": [ + 282, + 551, + 320, + 563 + ], + "score": 0.92, + "content": "| \\phi | = C", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 550, + 407, + 564 + ], + "score": 1.0, + "content": ", or even every pixel", + "type": "text" + }, + { + "bbox": [ + 407, + 551, + 492, + 563 + ], + "score": 0.91, + "content": "| \\phi | = \\bar { W \\times H \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "score": 1.0, + "content": "in Figure 3. These choices correspond to different likelihood distribution classes. Dimensionality", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 479, + 585 + ], + "score": 1.0, + "content": "and Conditioning of likelihood parameters can interact. For example, data parameters with", + "type": "text" + }, + { + "bbox": [ + 479, + 573, + 505, + 585 + ], + "score": 0.9, + "content": "| \\phi | \\overset { \\cdot } { = }", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 107, + 584, + 161, + 594 + ], + "score": 0.92, + "content": "W \\times H \\times C", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 583, + 226, + 595 + ], + "score": 1.0, + "content": "would result in", + "type": "text" + }, + { + "bbox": [ + 226, + 584, + 303, + 594 + ], + "score": 0.92, + "content": "N \\times W \\times H \\times C", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 583, + 424, + 595 + ], + "score": 1.0, + "content": "additional parameters, where", + "type": "text" + }, + { + "bbox": [ + 424, + 584, + 434, + 594 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "is the size of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "dataset. This can complicate implementations and slow down optimization due to disk I/O when", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "their size exceeds memory. Table 5 in the appendix contrasts the computational requirements of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "different likelihood parameter types. The work of Barron (2019) explores both scalar and pixel-wise", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 628, + 250, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 250, + 641 + ], + "score": 1.0, + "content": "dimensionalities for his robust loss.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 528, + 506, + 641 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 657, + 202, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 204, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 204, + 672 + ], + "score": 1.0, + "content": "5 APPLICATIONS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 122, + 407, + 204 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 506, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 505, + 93 + ], + "score": 1.0, + "content": "Table 1: MSE, Time, and Memory increase (compared to standard normal likelihood) for reconstruc-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 451, + 104 + ], + "score": 1.0, + "content": "tion by variational auto-encoders with different parameterizations of the robust loss,", + "type": "text" + }, + { + "bbox": [ + 452, + 93, + 458, + 103 + ], + "score": 0.72, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 91, + 505, + 104 + ], + "score": 1.0, + "content": ". Predicted", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 371, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 371, + 115 + ], + "score": 1.0, + "content": "likelihood parameters yield more accurate reconstruction models.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 202, + 122, + 407, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 122, + 407, + 204 + ], + "spans": [ + { + "bbox": [ + 202, + 122, + 407, + 204 + ], + "score": 0.974, + "html": "
Param.DimMSETimeMem
Global1x1x1225.81.04×<1KB
Data1x1x1244.22.70×0.6GB
Pred.1x1x1228.51.04×<1MB
GlobalHxWxC231.11.08×<1MB
DataHxWxC252.69.42×4.4GB
Pred.HxWxC222.31.08×<1MB
", + "type": "table", + "image_path": "3097495452486ac35715d02502296356e6c54163934240ac695a578863e89140.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 122, + 407, + 135.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 202, + 135.66666666666666, + 407, + 149.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 202, + 149.33333333333331, + 407, + 162.99999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 202, + 162.99999999999997, + 407, + 176.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 202, + 176.66666666666663, + 407, + 190.3333333333333 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 202, + 190.3333333333333, + 407, + 203.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 108, + 223, + 304, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 306, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 306, + 236 + ], + "score": 1.0, + "content": "5.1 ROBUSTNESS AND OUTLIER DETECTION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 242, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 504, + 255 + ], + "score": 1.0, + "content": "Data in the wild is noisy, and machine learning methods should be robust to noise, heteroskedastic-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "ity, and corruption. Unfortunately, models trained with the standard mean squared error (MSE) loss", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "score": 1.0, + "content": "are highly susceptible to outliers, and cannot naturally handle heteroskedasticity due to this loss’", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "fixed variance (Huber, 2004). Allowing models to predict and optimize their likelihood parameters", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 286, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 301 + ], + "score": 1.0, + "content": "allows models to generalize to these more complex settings. More specifically, likelihood parame-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "ters naturally transform standard methods such as regressors, classifiers, and manifold learners into", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "robust variants without expensive outer-loop of model fitting such as RANSAC (Fischler & Bolles,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 319, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 333 + ], + "score": 1.0, + "content": "1981) and Theil-Sen (Theil, 1992). Figure 4 demonstrates this effect with a simple classification", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "dataset, and we point readers to Figures 9 of the Supplement for similar examples for regression and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 341, + 182, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 182, + 356 + ], + "score": 1.0, + "content": "manifold learning.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 359, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 370 + ], + "score": 1.0, + "content": "In certain datasets, even the assumption of Gaussianity is too restrictive and one must consider more", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "robust and long-tailed distributions. This has led many to investigate broader classes of likelihoods", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "score": 1.0, + "content": "such as Generalized Linear Models (GLMs) (Nelder & Wedderburn, 1972) or the more recent gen-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 173, + 405 + ], + "score": 1.0, + "content": "eral robust loss,", + "type": "text" + }, + { + "bbox": [ + 173, + 395, + 180, + 404 + ], + "score": 0.71, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 392, + 504, + 405 + ], + "score": 1.0, + "content": ", of (Barron, 2019). To systematically explore how likelihood parameter dimen-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "sion and conditioning affect model robustness and quality, we reproduce Barron (2019)’s variational", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "auto-encoding (Kingma & Ba, 2015) (VAE) experiments on faces from the CelebA dataset (Liu", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "et al., 2015) in Table 1. We explore learned data (Saxena et al., 2019) and model parameters in addi-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "tion to Barron’s learned global parameters. We also include two natural parameter dimensionalities:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "a single set of parameters for the whole image, and a set of parameters for each pixel and chan-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "nel. We find that predicted parameters achieve the best performance while maintaining fast training", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 469, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 480 + ], + "score": 1.0, + "content": "time and a small memory footprint. We also find that pixel-wise learned parameters correlate with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 480, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 480, + 492 + ], + "score": 1.0, + "content": "challenging areas of images and we visualize these parameters in Section D of the Appendix.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 107, + 498, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 107, + 498, + 204, + 508 + ], + "score": 1.0, + "content": "This experiment uses a", + "type": "text" + }, + { + "bbox": [ + 204, + 498, + 229, + 507 + ], + "score": 0.85, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 498, + 505, + 508 + ], + "score": 1.0, + "content": "convolution on the last hidden layer of the decoder as a likelihood", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "parameter model and has the same resolution as the output. The low and high dimensional losses", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 518, + 504, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 504, + 531 + ], + "score": 1.0, + "content": "use the same convolutional regressor, but the 1 dimensional case averages over pixels. In the high di-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 530, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 107, + 530, + 505, + 541 + ], + "score": 1.0, + "content": "mensional case, the output has three channels (for RGB), with six channels total for shape and scale", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "regression. We use the same non-linearities to constrain the shape and scale outputs to reasonable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 551, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 464, + 564 + ], + "score": 1.0, + "content": "ranges as in (Barron, 2019). More specifically, we use an affine sigmoid to keep the shape", + "type": "text" + }, + { + "bbox": [ + 464, + 551, + 504, + 564 + ], + "score": 0.92, + "content": "\\alpha \\in [ 0 , 3 ]", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 563, + 504, + 575 + ], + "spans": [ + { + "bbox": [ + 107, + 563, + 224, + 574 + ], + "score": 1.0, + "content": "and the softplus to keep scale", + "type": "text" + }, + { + "bbox": [ + 225, + 563, + 285, + 575 + ], + "score": 0.93, + "content": "c \\in [ 1 0 ^ { - } 8 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 563, + 504, + 574 + ], + "score": 1.0, + "content": ". Table 1 gives the results of evaluating each method by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "MSE on the validation set, while training each method with their respective loss parameters. Data", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 586, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 107, + 586, + 504, + 597 + ], + "score": 1.0, + "content": "parameter optimization uses Tensorflow’s implementation of sparse RMSProp (Tieleman & Hinton,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 255, + 609 + ], + "score": 1.0, + "content": "2012). We also inherit weight decay", + "type": "text" + }, + { + "bbox": [ + 256, + 596, + 277, + 609 + ], + "score": 0.9, + "content": "\\| \\phi \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 596, + 352, + 609 + ], + "score": 1.0, + "content": ", gradient clipping", + "type": "text" + }, + { + "bbox": [ + 352, + 596, + 400, + 609 + ], + "score": 0.95, + "content": "\\bar { \\nabla } _ { \\phi } / \\| \\nabla _ { \\phi } \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 596, + 505, + 609 + ], + "score": 1.0, + "content": ", and learning rate scaling", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 608, + 384, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 155, + 620 + ], + "score": 0.85, + "content": "\\alpha _ { \\phi } = \\alpha \\cdot m", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 608, + 223, + 620 + ], + "score": 1.0, + "content": "for learning rate", + "type": "text" + }, + { + "bbox": [ + 223, + 610, + 231, + 617 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 608, + 290, + 620 + ], + "score": 1.0, + "content": "and multiplier", + "type": "text" + }, + { + "bbox": [ + 290, + 610, + 300, + 617 + ], + "score": 0.69, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 608, + 384, + 620 + ], + "score": 1.0, + "content": "from Barron (2019).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 625, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 107, + 626, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 107, + 626, + 505, + 637 + ], + "score": 1.0, + "content": "The robustness we see in our VAE experiments stems from the fact that likelihood parameter predic-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "tion gives models a direct channel to express their “uncertainty” for each data-point with respect to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "score": 1.0, + "content": "the task. This allows models to naturally down-weight and clean outliers from the dataset which can", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 659, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 669 + ], + "score": 1.0, + "content": "improve model robustness. Consequently, one can harness this effect to create outlier detectors from", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "any underlying model architecture by using learned scales or temperatures as an outlier score func-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "tion. Furthermore, predicted likelihood parameters allow these methods to detect outliers in unseen", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "data. In Figure 5 we show how auditing temperature or noise parameters can help practitioners spot", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 702, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 505, + 714 + ], + "score": 1.0, + "content": "erroneous labels and poor quality examples. In particular, the model-parameterized temperatures of", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 122, + 407, + 204 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 506, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 505, + 93 + ], + "score": 1.0, + "content": "Table 1: MSE, Time, and Memory increase (compared to standard normal likelihood) for reconstruc-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 451, + 104 + ], + "score": 1.0, + "content": "tion by variational auto-encoders with different parameterizations of the robust loss,", + "type": "text" + }, + { + "bbox": [ + 452, + 93, + 458, + 103 + ], + "score": 0.72, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 91, + 505, + 104 + ], + "score": 1.0, + "content": ". Predicted", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 371, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 371, + 115 + ], + "score": 1.0, + "content": "likelihood parameters yield more accurate reconstruction models.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 202, + 122, + 407, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 202, + 122, + 407, + 204 + ], + "spans": [ + { + "bbox": [ + 202, + 122, + 407, + 204 + ], + "score": 0.974, + "html": "
Param.DimMSETimeMem
Global1x1x1225.81.04×<1KB
Data1x1x1244.22.70×0.6GB
Pred.1x1x1228.51.04×<1MB
GlobalHxWxC231.11.08×<1MB
DataHxWxC252.69.42×4.4GB
Pred.HxWxC222.31.08×<1MB
", + "type": "table", + "image_path": "3097495452486ac35715d02502296356e6c54163934240ac695a578863e89140.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 122, + 407, + 135.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 202, + 135.66666666666666, + 407, + 149.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 202, + 149.33333333333331, + 407, + 162.99999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 202, + 162.99999999999997, + 407, + 176.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 202, + 176.66666666666663, + 407, + 190.3333333333333 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 202, + 190.3333333333333, + 407, + 203.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 108, + 223, + 304, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 306, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 306, + 236 + ], + "score": 1.0, + "content": "5.1 ROBUSTNESS AND OUTLIER DETECTION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 242, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 504, + 255 + ], + "score": 1.0, + "content": "Data in the wild is noisy, and machine learning methods should be robust to noise, heteroskedastic-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "ity, and corruption. Unfortunately, models trained with the standard mean squared error (MSE) loss", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "score": 1.0, + "content": "are highly susceptible to outliers, and cannot naturally handle heteroskedasticity due to this loss’", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "fixed variance (Huber, 2004). Allowing models to predict and optimize their likelihood parameters", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 286, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 301 + ], + "score": 1.0, + "content": "allows models to generalize to these more complex settings. More specifically, likelihood parame-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "ters naturally transform standard methods such as regressors, classifiers, and manifold learners into", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "robust variants without expensive outer-loop of model fitting such as RANSAC (Fischler & Bolles,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 319, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 333 + ], + "score": 1.0, + "content": "1981) and Theil-Sen (Theil, 1992). Figure 4 demonstrates this effect with a simple classification", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "dataset, and we point readers to Figures 9 of the Supplement for similar examples for regression and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 341, + 182, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 182, + 356 + ], + "score": 1.0, + "content": "manifold learning.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 242, + 506, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 359, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 370 + ], + "score": 1.0, + "content": "In certain datasets, even the assumption of Gaussianity is too restrictive and one must consider more", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "robust and long-tailed distributions. This has led many to investigate broader classes of likelihoods", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "score": 1.0, + "content": "such as Generalized Linear Models (GLMs) (Nelder & Wedderburn, 1972) or the more recent gen-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 173, + 405 + ], + "score": 1.0, + "content": "eral robust loss,", + "type": "text" + }, + { + "bbox": [ + 173, + 395, + 180, + 404 + ], + "score": 0.71, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 392, + 504, + 405 + ], + "score": 1.0, + "content": ", of (Barron, 2019). To systematically explore how likelihood parameter dimen-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "sion and conditioning affect model robustness and quality, we reproduce Barron (2019)’s variational", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "auto-encoding (Kingma & Ba, 2015) (VAE) experiments on faces from the CelebA dataset (Liu", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "et al., 2015) in Table 1. We explore learned data (Saxena et al., 2019) and model parameters in addi-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "tion to Barron’s learned global parameters. We also include two natural parameter dimensionalities:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "a single set of parameters for the whole image, and a set of parameters for each pixel and chan-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "nel. We find that predicted parameters achieve the best performance while maintaining fast training", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 469, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 480 + ], + "score": 1.0, + "content": "time and a small memory footprint. We also find that pixel-wise learned parameters correlate with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 480, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 480, + 492 + ], + "score": 1.0, + "content": "challenging areas of images and we visualize these parameters in Section D of the Appendix.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 360, + 506, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 107, + 498, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 107, + 498, + 204, + 508 + ], + "score": 1.0, + "content": "This experiment uses a", + "type": "text" + }, + { + "bbox": [ + 204, + 498, + 229, + 507 + ], + "score": 0.85, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 498, + 505, + 508 + ], + "score": 1.0, + "content": "convolution on the last hidden layer of the decoder as a likelihood", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "parameter model and has the same resolution as the output. The low and high dimensional losses", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 518, + 504, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 504, + 531 + ], + "score": 1.0, + "content": "use the same convolutional regressor, but the 1 dimensional case averages over pixels. In the high di-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 530, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 107, + 530, + 505, + 541 + ], + "score": 1.0, + "content": "mensional case, the output has three channels (for RGB), with six channels total for shape and scale", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "regression. We use the same non-linearities to constrain the shape and scale outputs to reasonable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 551, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 464, + 564 + ], + "score": 1.0, + "content": "ranges as in (Barron, 2019). More specifically, we use an affine sigmoid to keep the shape", + "type": "text" + }, + { + "bbox": [ + 464, + 551, + 504, + 564 + ], + "score": 0.92, + "content": "\\alpha \\in [ 0 , 3 ]", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 563, + 504, + 575 + ], + "spans": [ + { + "bbox": [ + 107, + 563, + 224, + 574 + ], + "score": 1.0, + "content": "and the softplus to keep scale", + "type": "text" + }, + { + "bbox": [ + 225, + 563, + 285, + 575 + ], + "score": 0.93, + "content": "c \\in [ 1 0 ^ { - } 8 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 563, + 504, + 574 + ], + "score": 1.0, + "content": ". Table 1 gives the results of evaluating each method by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "MSE on the validation set, while training each method with their respective loss parameters. Data", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 586, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 107, + 586, + 504, + 597 + ], + "score": 1.0, + "content": "parameter optimization uses Tensorflow’s implementation of sparse RMSProp (Tieleman & Hinton,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 255, + 609 + ], + "score": 1.0, + "content": "2012). We also inherit weight decay", + "type": "text" + }, + { + "bbox": [ + 256, + 596, + 277, + 609 + ], + "score": 0.9, + "content": "\\| \\phi \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 596, + 352, + 609 + ], + "score": 1.0, + "content": ", gradient clipping", + "type": "text" + }, + { + "bbox": [ + 352, + 596, + 400, + 609 + ], + "score": 0.95, + "content": "\\bar { \\nabla } _ { \\phi } / \\| \\nabla _ { \\phi } \\| _ { 2 } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 596, + 505, + 609 + ], + "score": 1.0, + "content": ", and learning rate scaling", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 608, + 384, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 155, + 620 + ], + "score": 0.85, + "content": "\\alpha _ { \\phi } = \\alpha \\cdot m", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 608, + 223, + 620 + ], + "score": 1.0, + "content": "for learning rate", + "type": "text" + }, + { + "bbox": [ + 223, + 610, + 231, + 617 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 608, + 290, + 620 + ], + "score": 1.0, + "content": "and multiplier", + "type": "text" + }, + { + "bbox": [ + 290, + 610, + 300, + 617 + ], + "score": 0.69, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 608, + 384, + 620 + ], + "score": 1.0, + "content": "from Barron (2019).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 498, + 505, + 620 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 625, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 107, + 626, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 107, + 626, + 505, + 637 + ], + "score": 1.0, + "content": "The robustness we see in our VAE experiments stems from the fact that likelihood parameter predic-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "tion gives models a direct channel to express their “uncertainty” for each data-point with respect to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 505, + 659 + ], + "score": 1.0, + "content": "the task. This allows models to naturally down-weight and clean outliers from the dataset which can", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 659, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 669 + ], + "score": 1.0, + "content": "improve model robustness. Consequently, one can harness this effect to create outlier detectors from", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "any underlying model architecture by using learned scales or temperatures as an outlier score func-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "tion. Furthermore, predicted likelihood parameters allow these methods to detect outliers in unseen", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "data. In Figure 5 we show how auditing temperature or noise parameters can help practitioners spot", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 702, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 505, + 714 + ], + "score": 1.0, + "content": "erroneous labels and poor quality examples. In particular, the model-parameterized temperatures of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "an image classifier (trained using the setup of 5.3) correlates strongly with blurry, dark, and diffi-", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "cult examples on the Street View House Number (SVHN) dataset. We use this approach to create", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 332, + 502 + ], + "score": 1.0, + "content": "simple outlier detection algorithms by considering deep", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 332, + 490, + 363, + 501 + ], + "score": 0.66, + "content": "\\left( \\mathrm { A E } { + } \\mathrm { S } \\right)", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 363, + 489, + 408, + 502 + ], + "score": 1.0, + "content": "and linear", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 408, + 491, + 444, + 501 + ], + "score": 0.78, + "content": "( \\mathrm { P C A } { + } \\mathrm { S } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 444, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "auto-encoders", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 107, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "(Kramer, 1991) with data-conditioned scale parameters as outlier scores. We evaluate this approach", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "on tabular datasets using deep and linear auto-encoders with model-parameterized scales. In Table", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "2 we quantitatively demonstrate the quality of these simple likelihood parameter approaches across", + "type": "text", + "cross_page": true + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "22 datasets from the Outlier Detection Datasets (ODDS), a standard outlier detection benchmark", + "type": "text", + "cross_page": true + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "(Rayana, 2016). The ODDS benchmark supplies ground truth outlier labels for each dataset, which", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 556, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 567 + ], + "score": 1.0, + "content": "allows one to treat outlier detection as an unsupervised classification problem. We compare against", + "type": "text", + "cross_page": true + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "a variety of established outlier detection approaches included in the pyOD (Zhao et al., 2019) frame-", + "type": "text", + "cross_page": true + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "work including: One-Class SVMs (OCSVM) (Scholkopf et al., 2000), Local Outlier Fraction (LOF) ¨", + "type": "text", + "cross_page": true + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 588, + 504, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 504, + 604 + ], + "score": 1.0, + "content": "(Breunig et al., 2000), Angle Based Outlier Detection (ABOD) (Kriegel et al., 2008), Feature Bag-", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "ging (FB) (Lazarevic & Kumar, 2005), Auto Encoder Distance (AE) (Aggarwal, 2015), K-Nearest", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 504, + 624 + ], + "score": 1.0, + "content": "Neighbors (KNN) (Ramaswamy et al., 2000; Angiulli & Pizzuti, 2002), Copula Based Outlier Detec-", + "type": "text", + "cross_page": true + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 622, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 504, + 635 + ], + "score": 1.0, + "content": "tion (COPOD) (Li et al., 2020), Variational AutoEncoders (VAE) (Kingma & Welling, 2013), Mini-", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 504, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 504, + 646 + ], + "score": 1.0, + "content": "mum Covariance Determinants with Mahlanohbis Distance (MCD) (Rousseeuw & Driessen, 1999;", + "type": "text", + "cross_page": true + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 645, + 504, + 656 + ], + "spans": [ + { + "bbox": [ + 107, + 645, + 504, + 656 + ], + "score": 1.0, + "content": "Hardin & Rocke, 2004), Histogram-based Outlier Scores (HBOS) (Goldstein & Dengel, 2012), Prin-", + "type": "text", + "cross_page": true + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 655, + 504, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 504, + 668 + ], + "score": 1.0, + "content": "cipal Component Analysis (PCA) (Shyu et al., 2003), Isolation Forests (IF) (Liu et al., 2008; 2012),", + "type": "text", + "cross_page": true + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 667, + 404, + 678 + ], + "spans": [ + { + "bbox": [ + 107, + 667, + 404, + 678 + ], + "score": 1.0, + "content": "and the Clustering-Based Local Outlier Factor (CBLOF) (He et al., 2003).", + "type": "text", + "cross_page": true + } + ], + "index": 42 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 626, + 505, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 82, + 497, + 182 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 82, + 497, + 182 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 82, + 497, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 82, + 497, + 182 + ], + "score": 0.943, + "type": "image", + "image_path": "13cf97ec3dd147a060f3c2aae3cb2a76c02361a9cc6b8cfbd652566d92745bec.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 82, + 497, + 115.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 115.33333333333334, + 497, + 148.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 148.66666666666669, + 497, + 182.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 194, + 505, + 218 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "Figure 5: The data with the lowest (top) and highest (bottom) predicted temperatures in the SVHN", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 205, + 484, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 484, + 218 + ], + "score": 1.0, + "content": "dataset. High temperature entries are blurry, cropped poorly, and generally difficult to classify.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 141, + 268, + 267, + 446 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 228, + 303, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 227, + 303, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 303, + 241 + ], + "score": 1.0, + "content": "Table 2: Median outlier detection performance", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 239, + 303, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 303, + 251 + ], + "score": 1.0, + "content": "of several methods across 22 benchmark datasets", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 250, + 160, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 160, + 262 + ], + "score": 1.0, + "content": "from ODDS.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 141, + 268, + 267, + 446 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 268, + 267, + 446 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 267, + 446 + ], + "score": 0.937, + "html": "
MethodMedian AUC
LOF.669
FB.702
ABOD.727
AE.737
VAE.792
COPOD.799
PCA.808
OCSVM.814
MCD.820
KNN.822
HBOS.822
IF.823
CBLOF.836
AE+S (Ours).846
PCA+S (Ours).868
", + "type": "table", + "image_path": "b5df76bf37dce536a934fc7313890a362bb93b0b5236c93444add7b70b98c027.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 141, + 268, + 267, + 446 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "image", + "bbox": [ + 310, + 231, + 501, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 310, + 231, + 501, + 385 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 310, + 231, + 501, + 385 + ], + "spans": [ + { + "bbox": [ + 310, + 231, + 501, + 385 + ], + "score": 0.969, + "type": "image", + "image_path": "e93d5f5fd0e2d1653d0cf1d7ea585fa9b470765348bd2fcdb1405cb7f0328d0e.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 310, + 231, + 501, + 245.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 310, + 245.0, + 501, + 259.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 310, + 259.0, + 501, + 273.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 310, + 273.0, + 501, + 287.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 310, + 287.0, + 501, + 301.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 310, + 301.0, + 501, + 315.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 310, + 315.0, + 501, + 329.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 310, + 329.0, + 501, + 343.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 310, + 343.0, + 501, + 357.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 310, + 357.0, + 501, + 371.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 310, + 371.0, + 501, + 385.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 307, + 397, + 505, + 442 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 307, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "Figure 6: Distribution of Outlier Detection AUC", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 406, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 505, + 421 + ], + "score": 1.0, + "content": "across the ODDS Benchmark. Our approaches,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 419, + 341, + 430 + ], + "score": 0.81, + "content": "\\mathrm { P C A } { + } \\mathrm { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 417, + 361, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 362, + 419, + 388, + 429 + ], + "score": 0.84, + "content": "_ \\mathrm { A E + S }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 417, + 505, + 431 + ], + "score": 1.0, + "content": ", are competitive with other", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 429, + 417, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 429, + 417, + 443 + ], + "score": 1.0, + "content": "Outlier Detection systems.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + } + ], + "index": 17.75 + }, + { + "type": "text", + "bbox": [ + 106, + 465, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "an image classifier (trained using the setup of 5.3) correlates strongly with blurry, dark, and diffi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "cult examples on the Street View House Number (SVHN) dataset. We use this approach to create", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 332, + 502 + ], + "score": 1.0, + "content": "simple outlier detection algorithms by considering deep", + "type": "text" + }, + { + "bbox": [ + 332, + 490, + 363, + 501 + ], + "score": 0.66, + "content": "\\left( \\mathrm { A E } { + } \\mathrm { S } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 489, + 408, + 502 + ], + "score": 1.0, + "content": "and linear", + "type": "text" + }, + { + "bbox": [ + 408, + 491, + 444, + 501 + ], + "score": 0.78, + "content": "( \\mathrm { P C A } { + } \\mathrm { S } )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "auto-encoders", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 107, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "(Kramer, 1991) with data-conditioned scale parameters as outlier scores. We evaluate this approach", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "on tabular datasets using deep and linear auto-encoders with model-parameterized scales. In Table", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "2 we quantitatively demonstrate the quality of these simple likelihood parameter approaches across", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "22 datasets from the Outlier Detection Datasets (ODDS), a standard outlier detection benchmark", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "(Rayana, 2016). The ODDS benchmark supplies ground truth outlier labels for each dataset, which", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 556, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 567 + ], + "score": 1.0, + "content": "allows one to treat outlier detection as an unsupervised classification problem. We compare against", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "a variety of established outlier detection approaches included in the pyOD (Zhao et al., 2019) frame-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "work including: One-Class SVMs (OCSVM) (Scholkopf et al., 2000), Local Outlier Fraction (LOF) ¨", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 588, + 504, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 504, + 604 + ], + "score": 1.0, + "content": "(Breunig et al., 2000), Angle Based Outlier Detection (ABOD) (Kriegel et al., 2008), Feature Bag-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "ging (FB) (Lazarevic & Kumar, 2005), Auto Encoder Distance (AE) (Aggarwal, 2015), K-Nearest", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 504, + 624 + ], + "score": 1.0, + "content": "Neighbors (KNN) (Ramaswamy et al., 2000; Angiulli & Pizzuti, 2002), Copula Based Outlier Detec-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 622, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 504, + 635 + ], + "score": 1.0, + "content": "tion (COPOD) (Li et al., 2020), Variational AutoEncoders (VAE) (Kingma & Welling, 2013), Mini-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 504, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 504, + 646 + ], + "score": 1.0, + "content": "mum Covariance Determinants with Mahlanohbis Distance (MCD) (Rousseeuw & Driessen, 1999;", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 645, + 504, + 656 + ], + "spans": [ + { + "bbox": [ + 107, + 645, + 504, + 656 + ], + "score": 1.0, + "content": "Hardin & Rocke, 2004), Histogram-based Outlier Scores (HBOS) (Goldstein & Dengel, 2012), Prin-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 655, + 504, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 504, + 668 + ], + "score": 1.0, + "content": "cipal Component Analysis (PCA) (Shyu et al., 2003), Isolation Forests (IF) (Liu et al., 2008; 2012),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 667, + 404, + 678 + ], + "spans": [ + { + "bbox": [ + 107, + 667, + 404, + 678 + ], + "score": 1.0, + "content": "and the Clustering-Based Local Outlier Factor (CBLOF) (He et al., 2003).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 684, + 504, + 718 + ], + "lines": [ + { + "bbox": [ + 107, + 684, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 107, + 684, + 505, + 695 + ], + "score": 1.0, + "content": "Our predicted scale auto-encoders use PyTorch’s layers API (Paszke et al., 2019) with rectified linear", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 693, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 505, + 707 + ], + "score": 1.0, + "content": "unit (ReLU) activations for deep auto-encoders and Glorot uniform initialization (Dahl et al., 2013;", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 704, + 506, + 719 + ], + "spans": [ + { + "bbox": [ + 106, + 704, + 506, + 719 + ], + "score": 1.0, + "content": "Glorot & Bengio, 2010) for all layers. We use Adam (Kingma & Ba, 2015) with a learning rate of", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 82, + 497, + 182 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 82, + 497, + 182 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 82, + 497, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 82, + 497, + 182 + ], + "score": 0.943, + "type": "image", + "image_path": "13cf97ec3dd147a060f3c2aae3cb2a76c02361a9cc6b8cfbd652566d92745bec.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 82, + 497, + 115.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 115.33333333333334, + 497, + 148.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 148.66666666666669, + 497, + 182.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 194, + 505, + 218 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "Figure 5: The data with the lowest (top) and highest (bottom) predicted temperatures in the SVHN", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 205, + 484, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 484, + 218 + ], + "score": 1.0, + "content": "dataset. High temperature entries are blurry, cropped poorly, and generally difficult to classify.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 141, + 268, + 267, + 446 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 228, + 303, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 227, + 303, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 303, + 241 + ], + "score": 1.0, + "content": "Table 2: Median outlier detection performance", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 239, + 303, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 303, + 251 + ], + "score": 1.0, + "content": "of several methods across 22 benchmark datasets", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 250, + 160, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 160, + 262 + ], + "score": 1.0, + "content": "from ODDS.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 141, + 268, + 267, + 446 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 268, + 267, + 446 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 267, + 446 + ], + "score": 0.937, + "html": "
MethodMedian AUC
LOF.669
FB.702
ABOD.727
AE.737
VAE.792
COPOD.799
PCA.808
OCSVM.814
MCD.820
KNN.822
HBOS.822
IF.823
CBLOF.836
AE+S (Ours).846
PCA+S (Ours).868
", + "type": "table", + "image_path": "b5df76bf37dce536a934fc7313890a362bb93b0b5236c93444add7b70b98c027.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 141, + 268, + 267, + 446 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "image", + "bbox": [ + 310, + 231, + 501, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 310, + 231, + 501, + 385 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 310, + 231, + 501, + 385 + ], + "spans": [ + { + "bbox": [ + 310, + 231, + 501, + 385 + ], + "score": 0.969, + "type": "image", + "image_path": "e93d5f5fd0e2d1653d0cf1d7ea585fa9b470765348bd2fcdb1405cb7f0328d0e.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 310, + 231, + 501, + 245.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 310, + 245.0, + 501, + 259.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 310, + 259.0, + 501, + 273.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 310, + 273.0, + 501, + 287.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 310, + 287.0, + 501, + 301.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 310, + 301.0, + 501, + 315.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 310, + 315.0, + 501, + 329.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 310, + 329.0, + 501, + 343.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 310, + 343.0, + 501, + 357.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 310, + 357.0, + 501, + 371.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 310, + 371.0, + 501, + 385.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 307, + 397, + 505, + 442 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 307, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "Figure 6: Distribution of Outlier Detection AUC", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 406, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 505, + 421 + ], + "score": 1.0, + "content": "across the ODDS Benchmark. Our approaches,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 419, + 341, + 430 + ], + "score": 0.81, + "content": "\\mathrm { P C A } { + } \\mathrm { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 417, + 361, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 362, + 419, + 388, + 429 + ], + "score": 0.84, + "content": "_ \\mathrm { A E + S }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 417, + 505, + 431 + ], + "score": 1.0, + "content": ", are competitive with other", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 429, + 417, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 429, + 417, + 443 + ], + "score": 1.0, + "content": "Outlier Detection systems.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + } + ], + "index": 17.75 + }, + { + "type": "text", + "bbox": [ + 106, + 465, + 505, + 678 + ], + "lines": [], + "index": 33, + "bbox_fs": [ + 105, + 466, + 506, + 678 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 684, + 504, + 718 + ], + "lines": [ + { + "bbox": [ + 107, + 684, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 107, + 684, + 505, + 695 + ], + "score": 1.0, + "content": "Our predicted scale auto-encoders use PyTorch’s layers API (Paszke et al., 2019) with rectified linear", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 693, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 505, + 707 + ], + "score": 1.0, + "content": "unit (ReLU) activations for deep auto-encoders and Glorot uniform initialization (Dahl et al., 2013;", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 704, + 506, + 719 + ], + "spans": [ + { + "bbox": [ + 106, + 704, + 506, + 719 + ], + "score": 1.0, + "content": "Glorot & Bengio, 2010) for all layers. We use Adam (Kingma & Ba, 2015) with a learning rate of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 292, + 504, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 215, + 309 + ], + "score": 1.0, + "content": ".0005 for 4000 steps with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 215, + 296, + 235, + 306 + ], + "score": 0.86, + "content": "20 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 235, + 292, + 504, + 309 + ], + "score": 1.0, + "content": "dropout before the code space. We follow ODDS guidelines and", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 303, + 261, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 261, + 320 + ], + "score": 1.0, + "content": "standard scale the data prior to fitting.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 684, + 506, + 719 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 81, + 492, + 197 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 81, + 492, + 197 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 81, + 492, + 197 + ], + "spans": [ + { + "bbox": [ + 109, + 81, + 492, + 197 + ], + "score": 0.97, + "type": "image", + "image_path": "14ccfaff7d7b98120fd5a934955e0726811cf1fc2af4f91e79fd5ad741b81a41.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 81, + 492, + 119.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 119.66666666666666, + 492, + 158.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 158.33333333333331, + 492, + 196.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 206, + 505, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 207, + 504, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 212, + 218 + ], + "score": 1.0, + "content": "Figure 7: Performance of", + "type": "text" + }, + { + "bbox": [ + 212, + 207, + 226, + 217 + ], + "score": 0.72, + "content": "L 2", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 207, + 267, + 218 + ], + "score": 1.0, + "content": "(left) and", + "type": "text" + }, + { + "bbox": [ + 267, + 207, + 281, + 217 + ], + "score": 0.64, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 207, + 504, + 218 + ], + "score": 1.0, + "content": "(middle) regularized linear regression on a 500 dimen-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 320, + 231 + ], + "score": 1.0, + "content": "sional synthetic dataset where the true parameters,", + "type": "text" + }, + { + "bbox": [ + 321, + 218, + 334, + 228 + ], + "score": 0.86, + "content": "w ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 218, + 505, + 231 + ], + "score": 1.0, + "content": ", are known. Dynamic Ridge (D-Ridge)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 228, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 505, + 241 + ], + "score": 1.0, + "content": "and D-LASSO regression find the regularization strength that best estimates the true parameters.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "M-LASSO outperforms any single global regularization strength and does not shrink informative", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 278, + 263 + ], + "score": 1.0, + "content": "weights. (right) Performance of adaptive", + "type": "text" + }, + { + "bbox": [ + 279, + 251, + 292, + 261 + ], + "score": 0.63, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "regularization methods as a function of true model", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 262, + 459, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 459, + 274 + ], + "score": 1.0, + "content": "sparsity. In all cases, Multi-LASSO outperforms other methods by orders of magnitude.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 108, + 294, + 502, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 292, + 504, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 215, + 309 + ], + "score": 1.0, + "content": ".0005 for 4000 steps with", + "type": "text" + }, + { + "bbox": [ + 215, + 296, + 235, + 306 + ], + "score": 0.86, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 292, + 504, + 309 + ], + "score": 1.0, + "content": "dropout before the code space. We follow ODDS guidelines and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 303, + 261, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 261, + 320 + ], + "score": 1.0, + "content": "standard scale the data prior to fitting.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 164, + 336 + ], + "score": 1.0, + "content": "Our methods", + "type": "text" + }, + { + "bbox": [ + 165, + 324, + 197, + 335 + ], + "score": 0.8, + "content": "\\mathrm { P C A } { + } \\mathrm { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 323, + 216, + 336 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 216, + 324, + 242, + 335 + ], + "score": 0.86, + "content": "{ \\mathrm { A E } } { + } S _ { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 323, + 505, + 336 + ], + "score": 1.0, + "content": ") use a similar principle as isolation-based approaches that deter-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "mine outliers based on how difficult they are to model. In existing approaches, outliers influence and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "skew the isolation model which causes the model to exhibit less confidence on the whole. This hurts", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "a model’s ability to distinguish between inliers and outliers. In contrast, our approach allows the un-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "derlying model to down-weight outliers. This yields a more consistent model with a clearer decision", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "score": 1.0, + "content": "boundary between outliers and inliers as shown in Figure 4. As a future direction of investigation we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "note that our approach is model-architecture agnostic, and can be combined with domain-specific", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 401, + 439, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 439, + 412 + ], + "score": 1.0, + "content": "architectures to create outlier detection methods tailored to images, text, and audio.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 106, + 428, + 372, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 373, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 373, + 441 + ], + "score": 1.0, + "content": "5.2 ADAPTIVE REGULARIZATION WITH PRIOR PARAMETERS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 506, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "score": 1.0, + "content": "In addition to optimizing the shape and scale of the likelihood distribution of the model output,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "score": 1.0, + "content": "we can use the same approach to optimize the prior distribution of the model parameters. More", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 393, + 484 + ], + "score": 1.0, + "content": "specifically, we propose adaptive regularizers for a model’s parameters,", + "type": "text" + }, + { + "bbox": [ + 394, + 472, + 400, + 481 + ], + "score": 0.7, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 472, + 505, + 484 + ], + "score": 1.0, + "content": ". This approach optimizes", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 273, + 496 + ], + "score": 1.0, + "content": "the distribution parameters of the prior,", + "type": "text" + }, + { + "bbox": [ + 273, + 483, + 294, + 495 + ], + "score": 0.91, + "content": "\\phi _ { \\mathrm { p r i o r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 482, + 506, + 496 + ], + "score": 1.0, + "content": ", to naturally tune the degree of regularization. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 479, + 506 + ], + "score": 1.0, + "content": "particular, the Normal (Ridge, L2) and Laplace (LASSO, L1) priors, with scale parameters", + "type": "text" + }, + { + "bbox": [ + 479, + 495, + 486, + 503 + ], + "score": 0.75, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 107, + 505, + 112, + 514 + ], + "score": 0.65, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 505, + 505, + 516 + ], + "score": 1.0, + "content": ", regularize model parameters for small magnitude and sparsity respectively (Hastie et al., 2009).", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 514, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 224, + 529 + ], + "score": 1.0, + "content": "The degree of regularization,", + "type": "text" + }, + { + "bbox": [ + 224, + 515, + 269, + 527 + ], + "score": 0.92, + "content": "\\lambda \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 514, + 506, + 529 + ], + "score": 1.0, + "content": ", is conventionally a hyperparameter of the regularized loss", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 525, + 145, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 145, + 539 + ], + "score": 1.0, + "content": "function:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 535, + 392, + 570 + ], + "lines": [ + { + "bbox": [ + 219, + 535, + 392, + 570 + ], + "spans": [ + { + "bbox": [ + 219, + 535, + 392, + 570 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta } \\sum _ { i } ^ { N } ( \\hat { y _ { i } } : = f _ { \\theta } ( x _ { i } ) - y _ { i } ) ^ { 2 } + \\lambda \\sum _ { j } ^ { P } | \\theta _ { j } | .", + "type": "interline_equation", + "image_path": "f627f0efe268e4bc649489c0cbab3f901c21e69ae5fcc84e0630628c62c374cd.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 535, + 392, + 552.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 219, + 552.5, + 392, + 570.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 238, + 586 + ], + "score": 1.0, + "content": "We note that we cannot choose", + "type": "text" + }, + { + "bbox": [ + 239, + 574, + 246, + 584 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "by direct minimization because it admits a trivial minimum at", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 107, + 585, + 132, + 595 + ], + "score": 0.88, + "content": "\\lambda = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 585, + 505, + 597 + ], + "score": 1.0, + "content": ". In the linear case, one can select this weight efficiently using Least Angle Regression (Efron", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 244, + 608 + ], + "score": 1.0, + "content": "et al., 2004). However, in general", + "type": "text" + }, + { + "bbox": [ + 244, + 596, + 251, + 605 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "is usually learned through expensive cross validation methods.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 607, + 494, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 494, + 619 + ], + "score": 1.0, + "content": "Instead, we retain the prior with its scale parameter, and jointly optimize over the full likelihood:", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 632, + 423, + 668 + ], + "lines": [ + { + "bbox": [ + 189, + 632, + 423, + 668 + ], + "spans": [ + { + "bbox": [ + 189, + 632, + 423, + 668 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta , \\sigma , b } \\sum _ { i } ^ { N } \\left( \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } _ { i } - y _ { i } ) ^ { 2 } + \\log \\sigma \\right) + \\sum _ { j } ^ { P } \\left( \\frac { | \\theta _ { j } | } { b } + \\log b \\right)", + "type": "interline_equation", + "image_path": "c16fdc665e8bcdc32122a2a851920c4d8ca32265b9b9927fbec62b60a2ed0519.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 632, + 423, + 650.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 189, + 650.0, + 423, + 668.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "This approach, the Dynamic Lasso (D-LASSO), admits no trivial solution for the prior parameter", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 686, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 107, + 688, + 112, + 698 + ], + "score": 0.67, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 686, + 343, + 703 + ], + "score": 1.0, + "content": ", and must balance the effective regularization strength,", + "type": "text" + }, + { + "bbox": [ + 344, + 688, + 351, + 701 + ], + "score": 0.85, + "content": "\\frac { 1 } { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 686, + 480, + 703 + ], + "score": 1.0, + "content": ", with the normalization factor,", + "type": "text" + }, + { + "bbox": [ + 481, + 688, + 501, + 699 + ], + "score": 0.87, + "content": "\\log b", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 686, + 506, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "D-LASSO selects the degree of regularization by gradient descent, rather than expensive black-box", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "search. In Figure 7 (left) and (middle) we show that this approach, and its Ridge equivalent, yield", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "ideal settings of the regularization strength on a suite of synthetic regression problems. Figure 7", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 81, + 492, + 197 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 81, + 492, + 197 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 81, + 492, + 197 + ], + "spans": [ + { + "bbox": [ + 109, + 81, + 492, + 197 + ], + "score": 0.97, + "type": "image", + "image_path": "14ccfaff7d7b98120fd5a934955e0726811cf1fc2af4f91e79fd5ad741b81a41.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 81, + 492, + 119.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 119.66666666666666, + 492, + 158.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 158.33333333333331, + 492, + 196.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 206, + 505, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 207, + 504, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 212, + 218 + ], + "score": 1.0, + "content": "Figure 7: Performance of", + "type": "text" + }, + { + "bbox": [ + 212, + 207, + 226, + 217 + ], + "score": 0.72, + "content": "L 2", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 207, + 267, + 218 + ], + "score": 1.0, + "content": "(left) and", + "type": "text" + }, + { + "bbox": [ + 267, + 207, + 281, + 217 + ], + "score": 0.64, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 207, + 504, + 218 + ], + "score": 1.0, + "content": "(middle) regularized linear regression on a 500 dimen-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 320, + 231 + ], + "score": 1.0, + "content": "sional synthetic dataset where the true parameters,", + "type": "text" + }, + { + "bbox": [ + 321, + 218, + 334, + 228 + ], + "score": 0.86, + "content": "w ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 218, + 505, + 231 + ], + "score": 1.0, + "content": ", are known. Dynamic Ridge (D-Ridge)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 228, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 505, + 241 + ], + "score": 1.0, + "content": "and D-LASSO regression find the regularization strength that best estimates the true parameters.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "M-LASSO outperforms any single global regularization strength and does not shrink informative", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 278, + 263 + ], + "score": 1.0, + "content": "weights. (right) Performance of adaptive", + "type": "text" + }, + { + "bbox": [ + 279, + 251, + 292, + 261 + ], + "score": 0.63, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "regularization methods as a function of true model", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 262, + 459, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 459, + 274 + ], + "score": 1.0, + "content": "sparsity. In all cases, Multi-LASSO outperforms other methods by orders of magnitude.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 108, + 294, + 502, + 318 + ], + "lines": [], + "index": 9.5, + "bbox_fs": [ + 105, + 292, + 504, + 320 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 164, + 336 + ], + "score": 1.0, + "content": "Our methods", + "type": "text" + }, + { + "bbox": [ + 165, + 324, + 197, + 335 + ], + "score": 0.8, + "content": "\\mathrm { P C A } { + } \\mathrm { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 323, + 216, + 336 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 216, + 324, + 242, + 335 + ], + "score": 0.86, + "content": "{ \\mathrm { A E } } { + } S _ { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 323, + 505, + 336 + ], + "score": 1.0, + "content": ") use a similar principle as isolation-based approaches that deter-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "mine outliers based on how difficult they are to model. In existing approaches, outliers influence and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "skew the isolation model which causes the model to exhibit less confidence on the whole. This hurts", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "a model’s ability to distinguish between inliers and outliers. In contrast, our approach allows the un-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "derlying model to down-weight outliers. This yields a more consistent model with a clearer decision", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "score": 1.0, + "content": "boundary between outliers and inliers as shown in Figure 4. As a future direction of investigation we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "note that our approach is model-architecture agnostic, and can be combined with domain-specific", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 401, + 439, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 439, + 412 + ], + "score": 1.0, + "content": "architectures to create outlier detection methods tailored to images, text, and audio.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 323, + 505, + 412 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 428, + 372, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 373, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 373, + 441 + ], + "score": 1.0, + "content": "5.2 ADAPTIVE REGULARIZATION WITH PRIOR PARAMETERS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 506, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 463 + ], + "score": 1.0, + "content": "In addition to optimizing the shape and scale of the likelihood distribution of the model output,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "score": 1.0, + "content": "we can use the same approach to optimize the prior distribution of the model parameters. More", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 393, + 484 + ], + "score": 1.0, + "content": "specifically, we propose adaptive regularizers for a model’s parameters,", + "type": "text" + }, + { + "bbox": [ + 394, + 472, + 400, + 481 + ], + "score": 0.7, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 472, + 505, + 484 + ], + "score": 1.0, + "content": ". This approach optimizes", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 273, + 496 + ], + "score": 1.0, + "content": "the distribution parameters of the prior,", + "type": "text" + }, + { + "bbox": [ + 273, + 483, + 294, + 495 + ], + "score": 0.91, + "content": "\\phi _ { \\mathrm { p r i o r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 482, + 506, + 496 + ], + "score": 1.0, + "content": ", to naturally tune the degree of regularization. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 479, + 506 + ], + "score": 1.0, + "content": "particular, the Normal (Ridge, L2) and Laplace (LASSO, L1) priors, with scale parameters", + "type": "text" + }, + { + "bbox": [ + 479, + 495, + 486, + 503 + ], + "score": 0.75, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 107, + 505, + 112, + 514 + ], + "score": 0.65, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 505, + 505, + 516 + ], + "score": 1.0, + "content": ", regularize model parameters for small magnitude and sparsity respectively (Hastie et al., 2009).", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 514, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 224, + 529 + ], + "score": 1.0, + "content": "The degree of regularization,", + "type": "text" + }, + { + "bbox": [ + 224, + 515, + 269, + 527 + ], + "score": 0.92, + "content": "\\lambda \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 514, + 506, + 529 + ], + "score": 1.0, + "content": ", is conventionally a hyperparameter of the regularized loss", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 525, + 145, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 145, + 539 + ], + "score": 1.0, + "content": "function:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 449, + 506, + 539 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 535, + 392, + 570 + ], + "lines": [ + { + "bbox": [ + 219, + 535, + 392, + 570 + ], + "spans": [ + { + "bbox": [ + 219, + 535, + 392, + 570 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta } \\sum _ { i } ^ { N } ( \\hat { y _ { i } } : = f _ { \\theta } ( x _ { i } ) - y _ { i } ) ^ { 2 } + \\lambda \\sum _ { j } ^ { P } | \\theta _ { j } | .", + "type": "interline_equation", + "image_path": "f627f0efe268e4bc649489c0cbab3f901c21e69ae5fcc84e0630628c62c374cd.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 535, + 392, + 552.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 219, + 552.5, + 392, + 570.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 238, + 586 + ], + "score": 1.0, + "content": "We note that we cannot choose", + "type": "text" + }, + { + "bbox": [ + 239, + 574, + 246, + 584 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "by direct minimization because it admits a trivial minimum at", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 107, + 585, + 132, + 595 + ], + "score": 0.88, + "content": "\\lambda = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 585, + 505, + 597 + ], + "score": 1.0, + "content": ". In the linear case, one can select this weight efficiently using Least Angle Regression (Efron", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 244, + 608 + ], + "score": 1.0, + "content": "et al., 2004). However, in general", + "type": "text" + }, + { + "bbox": [ + 244, + 596, + 251, + 605 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "is usually learned through expensive cross validation methods.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 607, + 494, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 494, + 619 + ], + "score": 1.0, + "content": "Instead, we retain the prior with its scale parameter, and jointly optimize over the full likelihood:", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 573, + 506, + 619 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 632, + 423, + 668 + ], + "lines": [ + { + "bbox": [ + 189, + 632, + 423, + 668 + ], + "spans": [ + { + "bbox": [ + 189, + 632, + 423, + 668 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\theta , \\sigma , b } \\sum _ { i } ^ { N } \\left( \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } _ { i } - y _ { i } ) ^ { 2 } + \\log \\sigma \\right) + \\sum _ { j } ^ { P } \\left( \\frac { | \\theta _ { j } | } { b } + \\log b \\right)", + "type": "interline_equation", + "image_path": "c16fdc665e8bcdc32122a2a851920c4d8ca32265b9b9927fbec62b60a2ed0519.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 632, + 423, + 650.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 189, + 650.0, + 423, + 668.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "This approach, the Dynamic Lasso (D-LASSO), admits no trivial solution for the prior parameter", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 686, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 107, + 688, + 112, + 698 + ], + "score": 0.67, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 686, + 343, + 703 + ], + "score": 1.0, + "content": ", and must balance the effective regularization strength,", + "type": "text" + }, + { + "bbox": [ + 344, + 688, + 351, + 701 + ], + "score": 0.85, + "content": "\\frac { 1 } { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 686, + 480, + 703 + ], + "score": 1.0, + "content": ", with the normalization factor,", + "type": "text" + }, + { + "bbox": [ + 481, + 688, + 501, + 699 + ], + "score": 0.87, + "content": "\\log b", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 686, + 506, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "D-LASSO selects the degree of regularization by gradient descent, rather than expensive black-box", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "search. In Figure 7 (left) and (middle) we show that this approach, and its Ridge equivalent, yield", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "ideal settings of the regularization strength on a suite of synthetic regression problems. Figure 7", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "(right) shows D-LASSO converges to the best LASSO regularization strength for a variety of true-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 356, + 106 + ], + "score": 1.0, + "content": "model sparsities. As a further extension, we replace the global", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 356, + 95, + 364, + 104 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 364, + 93, + 375, + 106 + ], + "score": 1.0, + "content": "or", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 375, + 94, + 381, + 104 + ], + "score": 0.64, + "content": "b", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 381, + 93, + 408, + 106 + ], + "score": 1.0, + "content": "with a", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 409, + 95, + 420, + 106 + ], + "score": 0.85, + "content": "\\sigma _ { j }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 420, + 93, + 432, + 106 + ], + "score": 1.0, + "content": "or", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 432, + 94, + 442, + 106 + ], + "score": 0.87, + "content": "b _ { j }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 442, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "for each model", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 150, + 117 + ], + "score": 1.0, + "content": "parameter,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 151, + 105, + 161, + 117 + ], + "score": 0.88, + "content": "\\theta _ { j }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 161, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ", to locally adapt regularization to each model weight (Multi-Lasso). This consistently", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "outperforms any global setting of the regularization strength and shields important weights from", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "undue shrinkage 7 (middle). For our experiments we use 500 samples of 500 dimensional normal", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "distributions mapped through linear functions with additive gaussian noise. Linear transformations", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 157, + 163 + ], + "score": 1.0, + "content": "use Uniform", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 157, + 151, + 177, + 163 + ], + "score": 0.6, + "content": "[ 1 , 2 ]", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 178, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "weights and LASSO experiments use sparse transformations. We use tensorflow’s", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 162, + 312, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 194, + 176 + ], + "score": 1.0, + "content": "Adam optimizer with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 194, + 163, + 240, + 173 + ], + "score": 0.87, + "content": "l r = . 0 0 0 5", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 240, + 162, + 312, + 176 + ], + "score": 1.0, + "content": "for 100000 steps.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "(right) shows D-LASSO converges to the best LASSO regularization strength for a variety of true-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 356, + 106 + ], + "score": 1.0, + "content": "model sparsities. As a further extension, we replace the global", + "type": "text" + }, + { + "bbox": [ + 356, + 95, + 364, + 104 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 93, + 375, + 106 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 375, + 94, + 381, + 104 + ], + "score": 0.64, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 93, + 408, + 106 + ], + "score": 1.0, + "content": "with a", + "type": "text" + }, + { + "bbox": [ + 409, + 95, + 420, + 106 + ], + "score": 0.85, + "content": "\\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 93, + 432, + 106 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 432, + 94, + 442, + 106 + ], + "score": 0.87, + "content": "b _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "for each model", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 150, + 117 + ], + "score": 1.0, + "content": "parameter,", + "type": "text" + }, + { + "bbox": [ + 151, + 105, + 161, + 117 + ], + "score": 0.88, + "content": "\\theta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ", to locally adapt regularization to each model weight (Multi-Lasso). This consistently", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "outperforms any global setting of the regularization strength and shields important weights from", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "undue shrinkage 7 (middle). For our experiments we use 500 samples of 500 dimensional normal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "distributions mapped through linear functions with additive gaussian noise. Linear transformations", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 157, + 163 + ], + "score": 1.0, + "content": "use Uniform", + "type": "text" + }, + { + "bbox": [ + 157, + 151, + 177, + 163 + ], + "score": 0.6, + "content": "[ 1 , 2 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "weights and LASSO experiments use sparse transformations. We use tensorflow’s", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 162, + 312, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 194, + 176 + ], + "score": 1.0, + "content": "Adam optimizer with", + "type": "text" + }, + { + "bbox": [ + 194, + 163, + 240, + 173 + ], + "score": 0.87, + "content": "l r = . 0 0 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 162, + 312, + 176 + ], + "score": 1.0, + "content": "for 100000 steps.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 178, + 504, + 254 + ], + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 189 + ], + "score": 1.0, + "content": "Our approach of learning regularizer scale parameters can be viewed naturally through the lens of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 201 + ], + "score": 1.0, + "content": "hierarchical priors (Gelman et al., 2013). More specifically this approach is implicitly performing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "maximum a posteriori (MAP) inference on the prior’s scale with respect to a uniform prior on that", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "parameter. We note that though these methods for hyperparameter selection are common in the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 222, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 107, + 222, + 504, + 233 + ], + "score": 1.0, + "content": "Bayesian literature, they are not widely used in practice in the deep learning community. This work", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 504, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 504, + 244 + ], + "score": 1.0, + "content": "aims to bring these parameters back within the scope of deep learning where they can be easily", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 243, + 384, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 384, + 256 + ], + "score": 1.0, + "content": "expanded to more flexible forms such as our introduced Multi-Lasso.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 271, + 207, + 282 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 208, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 208, + 284 + ], + "score": 1.0, + "content": "5.3 RE-CALIBRATION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 291, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 504, + 305 + ], + "score": 1.0, + "content": "The work of (Guo et al., 2017) shows that modern networks are accurate, yet systematically over-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "confident, a phenomenon called mis-calibration. We investigate the role of optimizing likelihood", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "parameters to re-calibrate models. More specifically, we can fit likelihood parameter regressors on a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 325, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 506, + 337 + ], + "score": 1.0, + "content": "validation set to modify an existing model’s confidence to better align with the validation set. This", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "approach is a generalization of Guo et al. (2017)’s Temperature Scaling method, which we refer to", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 346, + 507, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 346, + 507, + 360 + ], + "score": 1.0, + "content": "as Global Scaling (GS) for notational consistency. Global Scaling re-calibrates classifiers with a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 357, + 335, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 210, + 371 + ], + "score": 1.0, + "content": "learned global parameter,", + "type": "text" + }, + { + "bbox": [ + 210, + 360, + 217, + 368 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 357, + 300, + 371 + ], + "score": 1.0, + "content": "in the loss function:", + "type": "text" + }, + { + "bbox": [ + 300, + 358, + 330, + 370 + ], + "score": 0.94, + "content": "\\sigma ( \\vec { x } , \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 357, + 335, + 371 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 448 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 504, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 504, + 386 + ], + "score": 1.0, + "content": "Fitting model-conditioned likelihood parameters to a validation set defines a broad class of re-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "calibration strategies. From these we introduce three new re-calibration methods. Linear Scaling", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 230, + 410 + ], + "score": 1.0, + "content": "(LS) learns a linear mapping,", + "type": "text" + }, + { + "bbox": [ + 230, + 397, + 235, + 407 + ], + "score": 0.59, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 397, + 428, + 410 + ], + "score": 1.0, + "content": ", to transform logits to a softmax temperature:", + "type": "text" + }, + { + "bbox": [ + 428, + 397, + 469, + 409 + ], + "score": 0.93, + "content": "\\sigma ( \\vec { x } , l ( \\vec { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 397, + 505, + 410 + ], + "score": 1.0, + "content": ". Linear", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 298, + 422 + ], + "score": 1.0, + "content": "Feature Scaling (LFS) learns a linear mapping,", + "type": "text" + }, + { + "bbox": [ + 298, + 410, + 303, + 420 + ], + "score": 0.64, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 409, + 483, + 422 + ], + "score": 1.0, + "content": ", to transform the features prior to the logits,", + "type": "text" + }, + { + "bbox": [ + 484, + 408, + 491, + 421 + ], + "score": 0.82, + "content": "\\bar { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 409, + 505, + 422 + ], + "score": 1.0, + "content": ", to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 204, + 435 + ], + "score": 1.0, + "content": "a softmax temperature:", + "type": "text" + }, + { + "bbox": [ + 204, + 421, + 245, + 435 + ], + "score": 0.94, + "content": "\\sigma ( \\vec { x } , l ( \\vec { f } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 422, + 505, + 435 + ], + "score": 1.0, + "content": ". Finally, we introduce Deep Scaling (DS) for regressors which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 460, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 217, + 449 + ], + "score": 1.0, + "content": "learns a nonlinear network,", + "type": "text" + }, + { + "bbox": [ + 217, + 436, + 227, + 446 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 435, + 319, + 449 + ], + "score": 1.0, + "content": ", to transform features,", + "type": "text" + }, + { + "bbox": [ + 320, + 434, + 327, + 448 + ], + "score": 0.81, + "content": "\\bar { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 435, + 408, + 449 + ], + "score": 1.0, + "content": ", into a temperature:", + "type": "text" + }, + { + "bbox": [ + 409, + 434, + 456, + 448 + ], + "score": 0.93, + "content": "\\sigma ( \\vec { x } , N ( \\vec { f } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 435, + 460, + 449 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 104, + 451, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 104, + 451, + 505, + 467 + ], + "score": 1.0, + "content": "In Table 3 we compare our recalibration approaches to the previous state of the art: Global Scaling.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 462, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 478 + ], + "score": 1.0, + "content": "We note that (Guo et al., 2017) have already shown that Global Scaling outperform Bayesian Binning", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 104, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "into Quantiles (Naeini et al., 2015), Histogram binning (Zadrozny & Elkan, 2001), and Isotonic", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "Regression. We recalibrate both ResNet50 (He et al., 2016) and DenseNet121 (Huang et al., 2017)", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "on a variety of vision datasets. We measure classifier miscalibration using the Expected Calibration", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "Error (ECE) (Guo et al., 2017) to align with prior art. We additionally evaluate Isotonic recalibration,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 484, + 532 + ], + "score": 1.0, + "content": "Platt Scaling (Platt, 1999), and Vector Scaling (VS) (Guo et al., 2017), which learns a vector,", + "type": "text" + }, + { + "bbox": [ + 484, + 519, + 491, + 529 + ], + "score": 0.75, + "content": "\\vec { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 519, + 506, + 532 + ], + "score": 1.0, + "content": ", to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 529, + 179, + 542 + ], + "score": 1.0, + "content": "re-weight logits:", + "type": "text" + }, + { + "bbox": [ + 180, + 530, + 214, + 542 + ], + "score": 0.93, + "content": "\\sigma ( \\vec { v } \\vec { x } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 529, + 506, + 542 + ], + "score": 1.0, + "content": ". LS and LFS tend to outperform other approaches like GS and VS,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "which demonstrates that richer likelihood parametrizations can improve calibration akin to how", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 261, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 261, + 564 + ], + "score": 1.0, + "content": "richer models can improve prediction.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 107, + 566, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 107, + 566, + 505, + 576 + ], + "score": 1.0, + "content": "Our experiments leverage Tensorflow’s Dataset APIs that include the SVHN, (Netzer et al., 2011),", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 107, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "ImageNet (Deng et al., 2009), CIFAR-100, CIFAR-10 (Krizhevsky, 2009) datasets. We use Keras", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "implementations of DenseNet-121 (Huang et al., 2017) and ResNet-50 (He et al., 2016) with default", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 312, + 610 + ], + "score": 1.0, + "content": "initializations. For optimization we use Adam with", + "type": "text" + }, + { + "bbox": [ + 312, + 598, + 362, + 609 + ], + "score": 0.82, + "content": "l r = 0 . 0 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 597, + 366, + 610 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 367, + 598, + 440, + 609 + ], + "score": 0.54, + "content": "\\beta _ { 1 } = . 9 , \\beta _ { 2 } = . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "(Kingma & Ba,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 608, + 327, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 327, + 621 + ], + "score": 1.0, + "content": "2015) and train for 300 epoch with a batch size of 512.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 725 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "For recalibrating regressors, we compare against the previous state of the art, Kuleshov et al. (2018),", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "who use an Isotonic regressor to correct a regressors’ confidence. We use the same experimental", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 107, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "setting as Kuleshov et al. (2018) including the UCI datasets (Dua & Graff, 2017), and regressor", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "calibration metric (CAL). Table 4 shows that our approaches can outperform this baseline as well as", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 107, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 107, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "the regression equivalent of Global Scaling. Inputs and targets are scaled to unit norm and variance", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 107, + 681, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 107, + 681, + 504, + 693 + ], + "score": 1.0, + "content": "prior to fitting for all regression experiments and missing values are imputed using scikit-learn’s", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 108, + 692, + 504, + 704 + ], + "spans": [ + { + "bbox": [ + 108, + 692, + 504, + 704 + ], + "score": 1.0, + "content": "“SimpleImputer” (Pedregosa et al., 2011). Experiments utilize Keras’ layers API with two hidden", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "score": 1.0, + "content": "rectified linear unit (ReLU) layers, Glorot uniform initialization (Dahl et al., 2013; Glorot & Bengio,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 713, + 444, + 726 + ], + "spans": [ + { + "bbox": [ + 106, + 713, + 249, + 726 + ], + "score": 1.0, + "content": "2010) and Adam optimization with", + "type": "text" + }, + { + "bbox": [ + 249, + 714, + 294, + 724 + ], + "score": 0.85, + "content": "l r = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 713, + 444, + 726 + ], + "score": 1.0, + "content": "for 3000 steps without minibatching.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 175 + ], + "lines": [], + "index": 3.5, + "bbox_fs": [ + 104, + 81, + 506, + 176 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 178, + 504, + 254 + ], + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 189 + ], + "score": 1.0, + "content": "Our approach of learning regularizer scale parameters can be viewed naturally through the lens of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 201 + ], + "score": 1.0, + "content": "hierarchical priors (Gelman et al., 2013). More specifically this approach is implicitly performing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "maximum a posteriori (MAP) inference on the prior’s scale with respect to a uniform prior on that", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "parameter. We note that though these methods for hyperparameter selection are common in the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 222, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 107, + 222, + 504, + 233 + ], + "score": 1.0, + "content": "Bayesian literature, they are not widely used in practice in the deep learning community. This work", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 504, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 504, + 244 + ], + "score": 1.0, + "content": "aims to bring these parameters back within the scope of deep learning where they can be easily", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 243, + 384, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 384, + 256 + ], + "score": 1.0, + "content": "expanded to more flexible forms such as our introduced Multi-Lasso.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 178, + 506, + 256 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 271, + 207, + 282 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 208, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 208, + 284 + ], + "score": 1.0, + "content": "5.3 RE-CALIBRATION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 291, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 504, + 305 + ], + "score": 1.0, + "content": "The work of (Guo et al., 2017) shows that modern networks are accurate, yet systematically over-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "confident, a phenomenon called mis-calibration. We investigate the role of optimizing likelihood", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "parameters to re-calibrate models. More specifically, we can fit likelihood parameter regressors on a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 325, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 506, + 337 + ], + "score": 1.0, + "content": "validation set to modify an existing model’s confidence to better align with the validation set. This", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "approach is a generalization of Guo et al. (2017)’s Temperature Scaling method, which we refer to", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 346, + 507, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 346, + 507, + 360 + ], + "score": 1.0, + "content": "as Global Scaling (GS) for notational consistency. Global Scaling re-calibrates classifiers with a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 357, + 335, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 210, + 371 + ], + "score": 1.0, + "content": "learned global parameter,", + "type": "text" + }, + { + "bbox": [ + 210, + 360, + 217, + 368 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 357, + 300, + 371 + ], + "score": 1.0, + "content": "in the loss function:", + "type": "text" + }, + { + "bbox": [ + 300, + 358, + 330, + 370 + ], + "score": 0.94, + "content": "\\sigma ( \\vec { x } , \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 357, + 335, + 371 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 291, + 507, + 371 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 448 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 504, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 504, + 386 + ], + "score": 1.0, + "content": "Fitting model-conditioned likelihood parameters to a validation set defines a broad class of re-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "calibration strategies. From these we introduce three new re-calibration methods. Linear Scaling", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 230, + 410 + ], + "score": 1.0, + "content": "(LS) learns a linear mapping,", + "type": "text" + }, + { + "bbox": [ + 230, + 397, + 235, + 407 + ], + "score": 0.59, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 397, + 428, + 410 + ], + "score": 1.0, + "content": ", to transform logits to a softmax temperature:", + "type": "text" + }, + { + "bbox": [ + 428, + 397, + 469, + 409 + ], + "score": 0.93, + "content": "\\sigma ( \\vec { x } , l ( \\vec { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 397, + 505, + 410 + ], + "score": 1.0, + "content": ". Linear", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 298, + 422 + ], + "score": 1.0, + "content": "Feature Scaling (LFS) learns a linear mapping,", + "type": "text" + }, + { + "bbox": [ + 298, + 410, + 303, + 420 + ], + "score": 0.64, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 409, + 483, + 422 + ], + "score": 1.0, + "content": ", to transform the features prior to the logits,", + "type": "text" + }, + { + "bbox": [ + 484, + 408, + 491, + 421 + ], + "score": 0.82, + "content": "\\bar { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 409, + 505, + 422 + ], + "score": 1.0, + "content": ", to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 204, + 435 + ], + "score": 1.0, + "content": "a softmax temperature:", + "type": "text" + }, + { + "bbox": [ + 204, + 421, + 245, + 435 + ], + "score": 0.94, + "content": "\\sigma ( \\vec { x } , l ( \\vec { f } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 422, + 505, + 435 + ], + "score": 1.0, + "content": ". Finally, we introduce Deep Scaling (DS) for regressors which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 460, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 217, + 449 + ], + "score": 1.0, + "content": "learns a nonlinear network,", + "type": "text" + }, + { + "bbox": [ + 217, + 436, + 227, + 446 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 435, + 319, + 449 + ], + "score": 1.0, + "content": ", to transform features,", + "type": "text" + }, + { + "bbox": [ + 320, + 434, + 327, + 448 + ], + "score": 0.81, + "content": "\\bar { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 435, + 408, + 449 + ], + "score": 1.0, + "content": ", into a temperature:", + "type": "text" + }, + { + "bbox": [ + 409, + 434, + 456, + 448 + ], + "score": 0.93, + "content": "\\sigma ( \\vec { x } , N ( \\vec { f } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 435, + 460, + 449 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 375, + 505, + 449 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 104, + 451, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 104, + 451, + 505, + 467 + ], + "score": 1.0, + "content": "In Table 3 we compare our recalibration approaches to the previous state of the art: Global Scaling.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 462, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 478 + ], + "score": 1.0, + "content": "We note that (Guo et al., 2017) have already shown that Global Scaling outperform Bayesian Binning", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 473, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 104, + 473, + 506, + 488 + ], + "score": 1.0, + "content": "into Quantiles (Naeini et al., 2015), Histogram binning (Zadrozny & Elkan, 2001), and Isotonic", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "Regression. We recalibrate both ResNet50 (He et al., 2016) and DenseNet121 (Huang et al., 2017)", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "on a variety of vision datasets. We measure classifier miscalibration using the Expected Calibration", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "Error (ECE) (Guo et al., 2017) to align with prior art. We additionally evaluate Isotonic recalibration,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 484, + 532 + ], + "score": 1.0, + "content": "Platt Scaling (Platt, 1999), and Vector Scaling (VS) (Guo et al., 2017), which learns a vector,", + "type": "text" + }, + { + "bbox": [ + 484, + 519, + 491, + 529 + ], + "score": 0.75, + "content": "\\vec { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 519, + 506, + 532 + ], + "score": 1.0, + "content": ", to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 529, + 179, + 542 + ], + "score": 1.0, + "content": "re-weight logits:", + "type": "text" + }, + { + "bbox": [ + 180, + 530, + 214, + 542 + ], + "score": 0.93, + "content": "\\sigma ( \\vec { v } \\vec { x } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 529, + 506, + 542 + ], + "score": 1.0, + "content": ". LS and LFS tend to outperform other approaches like GS and VS,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "which demonstrates that richer likelihood parametrizations can improve calibration akin to how", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 261, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 261, + 564 + ], + "score": 1.0, + "content": "richer models can improve prediction.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 451, + 506, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 107, + 566, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 107, + 566, + 505, + 576 + ], + "score": 1.0, + "content": "Our experiments leverage Tensorflow’s Dataset APIs that include the SVHN, (Netzer et al., 2011),", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 107, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "ImageNet (Deng et al., 2009), CIFAR-100, CIFAR-10 (Krizhevsky, 2009) datasets. We use Keras", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "implementations of DenseNet-121 (Huang et al., 2017) and ResNet-50 (He et al., 2016) with default", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 312, + 610 + ], + "score": 1.0, + "content": "initializations. For optimization we use Adam with", + "type": "text" + }, + { + "bbox": [ + 312, + 598, + 362, + 609 + ], + "score": 0.82, + "content": "l r = 0 . 0 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 597, + 366, + 610 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 367, + 598, + 440, + 609 + ], + "score": 0.54, + "content": "\\beta _ { 1 } = . 9 , \\beta _ { 2 } = . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "(Kingma & Ba,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 608, + 327, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 327, + 621 + ], + "score": 1.0, + "content": "2015) and train for 300 epoch with a batch size of 512.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 566, + 505, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 725 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "For recalibrating regressors, we compare against the previous state of the art, Kuleshov et al. (2018),", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "who use an Isotonic regressor to correct a regressors’ confidence. We use the same experimental", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 107, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "setting as Kuleshov et al. (2018) including the UCI datasets (Dua & Graff, 2017), and regressor", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "calibration metric (CAL). Table 4 shows that our approaches can outperform this baseline as well as", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 107, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 107, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "the regression equivalent of Global Scaling. Inputs and targets are scaled to unit norm and variance", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 107, + 681, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 107, + 681, + 504, + 693 + ], + "score": 1.0, + "content": "prior to fitting for all regression experiments and missing values are imputed using scikit-learn’s", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 108, + 692, + 504, + 704 + ], + "spans": [ + { + "bbox": [ + 108, + 692, + 504, + 704 + ], + "score": 1.0, + "content": "“SimpleImputer” (Pedregosa et al., 2011). Experiments utilize Keras’ layers API with two hidden", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "score": 1.0, + "content": "rectified linear unit (ReLU) layers, Glorot uniform initialization (Dahl et al., 2013; Glorot & Bengio,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 713, + 444, + 726 + ], + "spans": [ + { + "bbox": [ + 106, + 713, + 249, + 726 + ], + "score": 1.0, + "content": "2010) and Adam optimization with", + "type": "text" + }, + { + "bbox": [ + 249, + 714, + 294, + 724 + ], + "score": 0.85, + "content": "l r = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 713, + 444, + 726 + ], + "score": 1.0, + "content": "for 3000 steps without minibatching.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48, + "bbox_fs": [ + 106, + 627, + 505, + 726 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 131, + 466, + 232 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 80, + 504, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "score": 1.0, + "content": "Table 3: Comparison of calibration methods by ECE for ResNet-50 (RN50) and DenseNet-121", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 506, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 506, + 104 + ], + "score": 1.0, + "content": "(DN121) architectures on test data. Our predicted likelihood parameter methods: Linear Scaling", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 102, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 114 + ], + "score": 1.0, + "content": "(LS) and Linear Feature Scaling (LFS) outperform other approaches. In all cases our methods", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 365, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 365, + 125 + ], + "score": 1.0, + "content": "reduce miscalibration with comparable computation time as GS.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 144, + 131, + 466, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 144, + 131, + 466, + 232 + ], + "spans": [ + { + "bbox": [ + 144, + 131, + 466, + 232 + ], + "score": 0.976, + "html": "
ModelDatasetUncalibratedPlattIsotonicGSVSLSLFS
RN50CIFAR-10.250.034.053.046.037.018.018
RN50CIFAR-100.642.061.072.035.044.030.173
RN50SVHN.072.053.010.029.022.009.009
RN50ImageNet.430.018.070.019.023.026.015
DN121CIFAR-10.253.048.042.039.034.028.028
DN121CIFAR-100.537.049.067.024.024.014.031
DN121SVHN.079.018.010.022.017.011.010
DN121ImageNet.229.028.095.021.019.043.019
", + "type": "table", + "image_path": "369344b0ae6cfc25469ef3c2cd30adf97e0a11d84f88602e26aec2e434c106e2.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 144, + 131, + 466, + 164.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 144, + 164.66666666666666, + 466, + 198.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 144, + 198.33333333333331, + 466, + 231.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 188, + 279, + 422, + 381 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 250, + 504, + 273 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "Table 4: Comparison of regression calibration methods as evaluated by their calibration error as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "defined in (Kuleshov et al., 2018). Predicted likelihood parameters often outperform other methods.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "table_body", + "bbox": [ + 188, + 279, + 422, + 381 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 188, + 279, + 422, + 381 + ], + "spans": [ + { + "bbox": [ + 188, + 279, + 422, + 381 + ], + "score": 0.977, + "html": "
DatasetUncalibrated IsotonicGSLSDS
crime0.36240.34990.0693 0.0125 0.0310
kinematics0.01640.01030.0022 0.0021 0.0032
bank0.01220.00560.0027 0.0024 0.0020
wine0.00910.01080.0152 0.01310.0064
mpg0.21530.22000.1964 0.14830.0233
cpu0.08620.03400.3018 0.20780.1740
soil0.30830.30000.3130 0.3175 0.3137
fried0.00060.00020.00020.00020.0002
", + "type": "table", + "image_path": "ce104ddd599fdb776ec7eda3b07530a3b5dc8713e4fa687011f6a95cf6631b79.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 279, + 422, + 291.75 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 188, + 291.75, + 422, + 304.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 188, + 304.5, + 422, + 317.25 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 188, + 317.25, + 422, + 330.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 188, + 330.0, + 422, + 342.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 188, + 342.75, + 422, + 355.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 188, + 355.5, + 422, + 368.25 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 188, + 368.25, + 422, + 381.0 + ], + "spans": [], + "index": 16 + } + ] + } + ], + "index": 10.0 + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 255, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 257, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 257, + 429 + ], + "score": 1.0, + "content": "6 EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 439, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "We run all experiments on Ubuntu 16.04 Azure Standard NV24 virtual machines (24 CPUs, 224 Gb", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 161, + 462 + ], + "score": 1.0, + "content": "memory, and", + "type": "text" + }, + { + "bbox": [ + 161, + 451, + 226, + 462 + ], + "score": 0.31, + "content": "4 \\times \\mathrm { M } 6 0 \\ : \\mathrm { G P U s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "with Tensorflow 1.15 (Abadi et al., 2015) and PyTorch 1.17 (Paszke", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 107, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "et al., 2019). Many likelihood parameters have constrained domains, such as the normal variance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 473, + 504, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 153, + 484 + ], + "score": 0.92, + "content": "\\sigma \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 473, + 504, + 484 + ], + "score": 1.0, + "content": ". To evade the complexity of constrained optimization, we define unconstrained param-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 128, + 495 + ], + "score": 1.0, + "content": "eters", + "type": "text" + }, + { + "bbox": [ + 128, + 485, + 141, + 495 + ], + "score": 0.89, + "content": "\\phi _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 483, + 261, + 495 + ], + "score": 1.0, + "content": "and choose a transformation", + "type": "text" + }, + { + "bbox": [ + 261, + 484, + 276, + 496 + ], + "score": 0.89, + "content": "t ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 483, + 330, + 495 + ], + "score": 1.0, + "content": "with inverse", + "type": "text" + }, + { + "bbox": [ + 330, + 483, + 357, + 496 + ], + "score": 0.92, + "content": "t ^ { - 1 } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "to map to and from the constrained", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 114, + 506 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 494, + 505, + 506 + ], + "score": 1.0, + "content": ". For positivity, exp/log parameterization is standard (Kendall & Gal, 2017; Kendall et al., 2018;", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "Saxena et al., 2019). However, this parameterization can lead to instabilities and we use the softplus,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 214, + 529 + ], + "score": 0.91, + "content": "s ^ { + } ( x ) = \\log ( 1 + \\exp ( x ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 515, + 340, + 529 + ], + "score": 1.0, + "content": ", instead. Shifting the softplus,", + "type": "text" + }, + { + "bbox": [ + 340, + 516, + 502, + 529 + ], + "score": 0.91, + "content": "s _ { c } ^ { + } ( x ) = ( l n ( 1 + e ^ { x } ) + c ) / ( l n ( 2 ) + c ) ,", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 515, + 505, + 529 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 528, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 107, + 528, + 505, + 539 + ], + "score": 1.0, + "content": "further improves stability and we explore this effect in Figure 12 of the Appendix. We use an affine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 538, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 140, + 551 + ], + "score": 1.0, + "content": "softplus", + "type": "text" + }, + { + "bbox": [ + 141, + 538, + 157, + 551 + ], + "score": 0.91, + "content": "s _ { . 0 1 } ^ { \\mp }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 538, + 176, + 551 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 176, + 538, + 189, + 551 + ], + "score": 0.91, + "content": "s _ { . 2 } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 538, + 504, + 551 + ], + "score": 1.0, + "content": "respectively for adaptive scales and temperatures respectively. The one excep-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "tion is adaptive regularizer scales where we found exp led to faster convergence. For the constrained", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 559, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 140, + 576 + ], + "score": 1.0, + "content": "interval", + "type": "text" + }, + { + "bbox": [ + 140, + 560, + 160, + 572 + ], + "score": 0.92, + "content": "[ a , b ]", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 559, + 345, + 576 + ], + "score": 1.0, + "content": "we use affine transformations of the sigmoid", + "type": "text" + }, + { + "bbox": [ + 345, + 560, + 421, + 575 + ], + "score": 0.93, + "content": "\\begin{array} { r } { s ( x ) = \\frac { \\tilde { 1 } } { 1 + \\exp ( - x ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 559, + 504, + 576 + ], + "score": 1.0, + "content": "(Barron, 2019). We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 573, + 495, + 585 + ], + "spans": [ + { + "bbox": [ + 107, + 573, + 441, + 585 + ], + "score": 1.0, + "content": "initialize likelihood parameter biases to settings that yield MSE and Cross Entropy", + "type": "text" + }, + { + "bbox": [ + 441, + 573, + 487, + 584 + ], + "score": 0.85, + "content": "\\sigma = \\tau = 1 \\mathrm { { } } ", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 573, + 495, + 585 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 108, + 606, + 195, + 618 + ], + "lines": [ + { + "bbox": [ + 104, + 604, + 198, + 622 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 198, + 622 + ], + "score": 1.0, + "content": "7 CONCLUSION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "Optimizing the full likelihood can improve model quality by adapting losses and regularizers. Full", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "likelihoods are agnostic to the architecture, optimizer, and task, which makes them simple substitutes", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "for standard losses. Global, data, and predicted likelihood parameters offer different degrees of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "expressivity and efficiency. In particular, predicted parameters adapt the likelihood to each data", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "point during training and testing without significant time and space overhead. By including these", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "parameters in a loss function one can improve a model’s robustness and generalization ability and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "create new classes of outlier detectors and recalibrators that outperform baselines. More generally,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "we hope this work encourages joint optimization of model and likelihood parameters, and argue it", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 720, + 290, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 290, + 733 + ], + "score": 1.0, + "content": "is likely that your loss should be a likelihood.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 131, + 466, + 232 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 80, + 504, + 124 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "score": 1.0, + "content": "Table 3: Comparison of calibration methods by ECE for ResNet-50 (RN50) and DenseNet-121", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 506, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 506, + 104 + ], + "score": 1.0, + "content": "(DN121) architectures on test data. Our predicted likelihood parameter methods: Linear Scaling", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 102, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 114 + ], + "score": 1.0, + "content": "(LS) and Linear Feature Scaling (LFS) outperform other approaches. In all cases our methods", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 365, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 365, + 125 + ], + "score": 1.0, + "content": "reduce miscalibration with comparable computation time as GS.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 144, + 131, + 466, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 144, + 131, + 466, + 232 + ], + "spans": [ + { + "bbox": [ + 144, + 131, + 466, + 232 + ], + "score": 0.976, + "html": "
ModelDatasetUncalibratedPlattIsotonicGSVSLSLFS
RN50CIFAR-10.250.034.053.046.037.018.018
RN50CIFAR-100.642.061.072.035.044.030.173
RN50SVHN.072.053.010.029.022.009.009
RN50ImageNet.430.018.070.019.023.026.015
DN121CIFAR-10.253.048.042.039.034.028.028
DN121CIFAR-100.537.049.067.024.024.014.031
DN121SVHN.079.018.010.022.017.011.010
DN121ImageNet.229.028.095.021.019.043.019
", + "type": "table", + "image_path": "369344b0ae6cfc25469ef3c2cd30adf97e0a11d84f88602e26aec2e434c106e2.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 144, + 131, + 466, + 164.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 144, + 164.66666666666666, + 466, + 198.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 144, + 198.33333333333331, + 466, + 231.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 188, + 279, + 422, + 381 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 250, + 504, + 273 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "Table 4: Comparison of regression calibration methods as evaluated by their calibration error as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "defined in (Kuleshov et al., 2018). Predicted likelihood parameters often outperform other methods.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "table_body", + "bbox": [ + 188, + 279, + 422, + 381 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 188, + 279, + 422, + 381 + ], + "spans": [ + { + "bbox": [ + 188, + 279, + 422, + 381 + ], + "score": 0.977, + "html": "
DatasetUncalibrated IsotonicGSLSDS
crime0.36240.34990.0693 0.0125 0.0310
kinematics0.01640.01030.0022 0.0021 0.0032
bank0.01220.00560.0027 0.0024 0.0020
wine0.00910.01080.0152 0.01310.0064
mpg0.21530.22000.1964 0.14830.0233
cpu0.08620.03400.3018 0.20780.1740
soil0.30830.30000.3130 0.3175 0.3137
fried0.00060.00020.00020.00020.0002
", + "type": "table", + "image_path": "ce104ddd599fdb776ec7eda3b07530a3b5dc8713e4fa687011f6a95cf6631b79.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 279, + 422, + 291.75 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 188, + 291.75, + 422, + 304.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 188, + 304.5, + 422, + 317.25 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 188, + 317.25, + 422, + 330.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 188, + 330.0, + 422, + 342.75 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 188, + 342.75, + 422, + 355.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 188, + 355.5, + 422, + 368.25 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 188, + 368.25, + 422, + 381.0 + ], + "spans": [], + "index": 16 + } + ] + } + ], + "index": 10.0 + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 255, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 257, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 257, + 429 + ], + "score": 1.0, + "content": "6 EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 439, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "We run all experiments on Ubuntu 16.04 Azure Standard NV24 virtual machines (24 CPUs, 224 Gb", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 161, + 462 + ], + "score": 1.0, + "content": "memory, and", + "type": "text" + }, + { + "bbox": [ + 161, + 451, + 226, + 462 + ], + "score": 0.31, + "content": "4 \\times \\mathrm { M } 6 0 \\ : \\mathrm { G P U s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "with Tensorflow 1.15 (Abadi et al., 2015) and PyTorch 1.17 (Paszke", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 107, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "et al., 2019). Many likelihood parameters have constrained domains, such as the normal variance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 473, + 504, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 153, + 484 + ], + "score": 0.92, + "content": "\\sigma \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 473, + 504, + 484 + ], + "score": 1.0, + "content": ". To evade the complexity of constrained optimization, we define unconstrained param-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 128, + 495 + ], + "score": 1.0, + "content": "eters", + "type": "text" + }, + { + "bbox": [ + 128, + 485, + 141, + 495 + ], + "score": 0.89, + "content": "\\phi _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 483, + 261, + 495 + ], + "score": 1.0, + "content": "and choose a transformation", + "type": "text" + }, + { + "bbox": [ + 261, + 484, + 276, + 496 + ], + "score": 0.89, + "content": "t ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 483, + 330, + 495 + ], + "score": 1.0, + "content": "with inverse", + "type": "text" + }, + { + "bbox": [ + 330, + 483, + 357, + 496 + ], + "score": 0.92, + "content": "t ^ { - 1 } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "to map to and from the constrained", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 114, + 506 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 494, + 505, + 506 + ], + "score": 1.0, + "content": ". For positivity, exp/log parameterization is standard (Kendall & Gal, 2017; Kendall et al., 2018;", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "Saxena et al., 2019). However, this parameterization can lead to instabilities and we use the softplus,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 214, + 529 + ], + "score": 0.91, + "content": "s ^ { + } ( x ) = \\log ( 1 + \\exp ( x ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 515, + 340, + 529 + ], + "score": 1.0, + "content": ", instead. Shifting the softplus,", + "type": "text" + }, + { + "bbox": [ + 340, + 516, + 502, + 529 + ], + "score": 0.91, + "content": "s _ { c } ^ { + } ( x ) = ( l n ( 1 + e ^ { x } ) + c ) / ( l n ( 2 ) + c ) ,", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 515, + 505, + 529 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 528, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 107, + 528, + 505, + 539 + ], + "score": 1.0, + "content": "further improves stability and we explore this effect in Figure 12 of the Appendix. We use an affine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 538, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 140, + 551 + ], + "score": 1.0, + "content": "softplus", + "type": "text" + }, + { + "bbox": [ + 141, + 538, + 157, + 551 + ], + "score": 0.91, + "content": "s _ { . 0 1 } ^ { \\mp }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 538, + 176, + 551 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 176, + 538, + 189, + 551 + ], + "score": 0.91, + "content": "s _ { . 2 } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 538, + 504, + 551 + ], + "score": 1.0, + "content": "respectively for adaptive scales and temperatures respectively. The one excep-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "tion is adaptive regularizer scales where we found exp led to faster convergence. For the constrained", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 559, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 140, + 576 + ], + "score": 1.0, + "content": "interval", + "type": "text" + }, + { + "bbox": [ + 140, + 560, + 160, + 572 + ], + "score": 0.92, + "content": "[ a , b ]", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 559, + 345, + 576 + ], + "score": 1.0, + "content": "we use affine transformations of the sigmoid", + "type": "text" + }, + { + "bbox": [ + 345, + 560, + 421, + 575 + ], + "score": 0.93, + "content": "\\begin{array} { r } { s ( x ) = \\frac { \\tilde { 1 } } { 1 + \\exp ( - x ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 559, + 504, + 576 + ], + "score": 1.0, + "content": "(Barron, 2019). We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 573, + 495, + 585 + ], + "spans": [ + { + "bbox": [ + 107, + 573, + 441, + 585 + ], + "score": 1.0, + "content": "initialize likelihood parameter biases to settings that yield MSE and Cross Entropy", + "type": "text" + }, + { + "bbox": [ + 441, + 573, + 487, + 584 + ], + "score": 0.85, + "content": "\\sigma = \\tau = 1 \\mathrm { { } } ", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 573, + 495, + 585 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 439, + 505, + 585 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 606, + 195, + 618 + ], + "lines": [ + { + "bbox": [ + 104, + 604, + 198, + 622 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 198, + 622 + ], + "score": 1.0, + "content": "7 CONCLUSION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "Optimizing the full likelihood can improve model quality by adapting losses and regularizers. Full", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "likelihoods are agnostic to the architecture, optimizer, and task, which makes them simple substitutes", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "for standard losses. Global, data, and predicted likelihood parameters offer different degrees of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "expressivity and efficiency. In particular, predicted parameters adapt the likelihood to each data", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "point during training and testing without significant time and space overhead. By including these", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "parameters in a loss function one can improve a model’s robustness and generalization ability and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "create new classes of outlier detectors and recalibrators that outperform baselines. More generally,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "we hope this work encourages joint optimization of model and likelihood parameters, and argue it", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 720, + 290, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 290, + 733 + ], + "score": 1.0, + "content": "is likely that your loss should be a likelihood.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 633, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 110, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 116, + 110, + 505, + 124 + ], + "score": 1.0, + "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 157 + ], + "score": 1.0, + "content": "Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vin-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "cent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 164, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 505, + 180 + ], + "score": 1.0, + "content": "berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 186, + 178, + 202 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 178, + 202 + ], + "score": 1.0, + "content": "tensorflow.org.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 206, + 440, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 441, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 441, + 219 + ], + "score": 1.0, + "content": "Charu C Aggarwal. Outlier analysis. In Data mining, pp. 237–263. Springer, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 502, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 504, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 504, + 238 + ], + "score": 1.0, + "content": "Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 235, + 491, + 248 + ], + "spans": [ + { + "bbox": [ + 116, + 235, + 491, + 248 + ], + "score": 1.0, + "content": "conference on principles of data mining and knowledge discovery, pp. 15–27. Springer, 2002.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 430, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 431, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 431, + 267 + ], + "score": 1.0, + "content": "Jonathan T Barron. A general and adaptive robust loss function. In CVPR, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 273, + 334, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 272, + 335, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 335, + 286 + ], + "score": 1.0, + "content": "Christopher M Bishop. Mixture density networks. 1994.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "Christopher M Bishop et al. Pattern recognition and machine learning, volume 4. springer New", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 302, + 166, + 314 + ], + "spans": [ + { + "bbox": [ + 116, + 302, + 166, + 314 + ], + "score": 1.0, + "content": "York, 2006.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 108, + 321, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 317, + 335 + ], + "score": 1.0, + "content": "Markus M Breunig, Hans-Peter Kriegel, Raymond T", + "type": "text" + }, + { + "bbox": [ + 318, + 322, + 332, + 333 + ], + "score": 0.68, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 320, + 504, + 335 + ], + "score": 1.0, + "content": ", and Jorg Sander. Lof: identifying density- ¨", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 116, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 343, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 343, + 279, + 356 + ], + "score": 1.0, + "content": "Management of data, pp. 93–104, 2000.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "score": 1.0, + "content": "Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Deterministic edge- ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 374, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 116, + 374, + 505, + 385 + ], + "score": 1.0, + "content": "preserving regularization in computed imaging. IEEE Transactions on image processing, 6(2):", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 383, + 181, + 396 + ], + "spans": [ + { + "bbox": [ + 116, + 383, + 181, + 396 + ], + "score": 1.0, + "content": "298–311, 1997.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 502, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 504, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 504, + 415 + ], + "score": 1.0, + "content": "George E Dahl, Tara N Sainath, and Geoffrey E Hinton. Improving deep neural networks for lvcsr", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 413, + 441, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 413, + 441, + 426 + ], + "score": 1.0, + "content": "using rectified linear units and dropout. In ICASSP, pp. 8609–8613. IEEE, 2013.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 502, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 504, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 504, + 445 + ], + "score": 1.0, + "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 443, + 376, + 455 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 376, + 455 + ], + "score": 1.0, + "content": "hierarchical image database. In CVPR, pp. 248–255. Ieee, 2009.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 108, + 461, + 502, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 503, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 503, + 475 + ], + "score": 1.0, + "content": "John E Dennis Jr and Roy E Welsch. Techniques for nonlinear least squares and robust regression.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 472, + 439, + 485 + ], + "spans": [ + { + "bbox": [ + 117, + 472, + 439, + 485 + ], + "score": 1.0, + "content": "Communications in Statistics-simulation and Computation, 7(4):345–359, 1978.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 491, + 501, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 504, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 504, + 505 + ], + "score": 1.0, + "content": "Dheeru Dua and Casey Graff. UCI machine learning repository, 2017. URL http://archive.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 502, + 204, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 502, + 204, + 514 + ], + "score": 1.0, + "content": "ics.uci.edu/ml.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 502, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 534 + ], + "score": 1.0, + "content": "Bradley Efron, Trevor Hastie, Iain Johnstone, Robert Tibshirani, et al. Least angle regression. The", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 531, + 285, + 544 + ], + "spans": [ + { + "bbox": [ + 115, + 531, + 285, + 544 + ], + "score": 1.0, + "content": "Annals of statistics, 32(2):407–499, 2004.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 564 + ], + "score": 1.0, + "content": "Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "with applications to image analysis and automated cartography. Communications of the ACM, 24", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 572, + 196, + 584 + ], + "spans": [ + { + "bbox": [ + 115, + 572, + 196, + 584 + ], + "score": 1.0, + "content": "(6):381–395, 1981.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 503, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 602, + 324, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 602, + 324, + 614 + ], + "score": 1.0, + "content": "tomography. Amer. Statist. Assoc, pp. 12–18, 1985.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 620, + 502, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "score": 1.0, + "content": "Andrew Gelman, John B Carlin, Hal S Stern, David B Dunson, Aki Vehtari, and Donald B Rubin.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 632, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 288, + 644 + ], + "score": 1.0, + "content": "Bayesian data analysis. CRC press, 2013.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 661, + 290, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 661, + 290, + 673 + ], + "score": 1.0, + "content": "networks. In AISTATS, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 502, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 504, + 693 + ], + "score": 1.0, + "content": "Markus Goldstein and Andreas Dengel. Histogram-based outlier score (hbos): A fast unsupervised", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 691, + 446, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 446, + 704 + ], + "score": 1.0, + "content": "anomaly detection algorithm. KI-2012: Poster and Demo Track, pp. 59–63, 2012.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 721, + 223, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 223, + 731 + ], + "score": 1.0, + "content": "networks. In ICML, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 110, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 116, + 110, + 505, + 124 + ], + "score": 1.0, + "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 157 + ], + "score": 1.0, + "content": "Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vin-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "cent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 164, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 505, + 180 + ], + "score": 1.0, + "content": "berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 186, + 178, + 202 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 178, + 202 + ], + "score": 1.0, + "content": "tensorflow.org.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 100, + 505, + 202 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 206, + 440, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 441, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 441, + 219 + ], + "score": 1.0, + "content": "Charu C Aggarwal. Outlier analysis. In Data mining, pp. 237–263. Springer, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 205, + 441, + 219 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 502, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 504, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 504, + 238 + ], + "score": 1.0, + "content": "Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 235, + 491, + 248 + ], + "spans": [ + { + "bbox": [ + 116, + 235, + 491, + 248 + ], + "score": 1.0, + "content": "conference on principles of data mining and knowledge discovery, pp. 15–27. Springer, 2002.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 224, + 504, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 430, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 431, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 431, + 267 + ], + "score": 1.0, + "content": "Jonathan T Barron. A general and adaptive robust loss function. In CVPR, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 253, + 431, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 273, + 334, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 272, + 335, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 335, + 286 + ], + "score": 1.0, + "content": "Christopher M Bishop. Mixture density networks. 1994.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 272, + 335, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "Christopher M Bishop et al. Pattern recognition and machine learning, volume 4. springer New", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 302, + 166, + 314 + ], + "spans": [ + { + "bbox": [ + 116, + 302, + 166, + 314 + ], + "score": 1.0, + "content": "York, 2006.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 290, + 505, + 314 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 321, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 317, + 335 + ], + "score": 1.0, + "content": "Markus M Breunig, Hans-Peter Kriegel, Raymond T", + "type": "text" + }, + { + "bbox": [ + 318, + 322, + 332, + 333 + ], + "score": 0.68, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 320, + 504, + 335 + ], + "score": 1.0, + "content": ", and Jorg Sander. Lof: identifying density- ¨", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 116, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 343, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 343, + 279, + 356 + ], + "score": 1.0, + "content": "Management of data, pp. 93–104, 2000.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 320, + 505, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "score": 1.0, + "content": "Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Deterministic edge- ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 374, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 116, + 374, + 505, + 385 + ], + "score": 1.0, + "content": "preserving regularization in computed imaging. IEEE Transactions on image processing, 6(2):", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 383, + 181, + 396 + ], + "spans": [ + { + "bbox": [ + 116, + 383, + 181, + 396 + ], + "score": 1.0, + "content": "298–311, 1997.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 361, + 505, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 502, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 504, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 504, + 415 + ], + "score": 1.0, + "content": "George E Dahl, Tara N Sainath, and Geoffrey E Hinton. Improving deep neural networks for lvcsr", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 413, + 441, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 413, + 441, + 426 + ], + "score": 1.0, + "content": "using rectified linear units and dropout. In ICASSP, pp. 8609–8613. IEEE, 2013.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 402, + 504, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 502, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 504, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 504, + 445 + ], + "score": 1.0, + "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 443, + 376, + 455 + ], + "spans": [ + { + "bbox": [ + 115, + 443, + 376, + 455 + ], + "score": 1.0, + "content": "hierarchical image database. In CVPR, pp. 248–255. Ieee, 2009.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 431, + 504, + 455 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 461, + 502, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 503, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 503, + 475 + ], + "score": 1.0, + "content": "John E Dennis Jr and Roy E Welsch. Techniques for nonlinear least squares and robust regression.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 472, + 439, + 485 + ], + "spans": [ + { + "bbox": [ + 117, + 472, + 439, + 485 + ], + "score": 1.0, + "content": "Communications in Statistics-simulation and Computation, 7(4):345–359, 1978.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 462, + 503, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 491, + 501, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 504, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 504, + 505 + ], + "score": 1.0, + "content": "Dheeru Dua and Casey Graff. UCI machine learning repository, 2017. URL http://archive.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 502, + 204, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 502, + 204, + 514 + ], + "score": 1.0, + "content": "ics.uci.edu/ml.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 490, + 504, + 514 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 502, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 534 + ], + "score": 1.0, + "content": "Bradley Efron, Trevor Hastie, Iain Johnstone, Robert Tibshirani, et al. Least angle regression. The", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 531, + 285, + 544 + ], + "spans": [ + { + "bbox": [ + 115, + 531, + 285, + 544 + ], + "score": 1.0, + "content": "Annals of statistics, 32(2):407–499, 2004.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 519, + 505, + 544 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 564 + ], + "score": 1.0, + "content": "Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "with applications to image analysis and automated cartography. Communications of the ACM, 24", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 572, + 196, + 584 + ], + "spans": [ + { + "bbox": [ + 115, + 572, + 196, + 584 + ], + "score": 1.0, + "content": "(6):381–395, 1981.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 549, + 505, + 584 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 503, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 602, + 324, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 602, + 324, + 614 + ], + "score": 1.0, + "content": "tomography. Amer. Statist. Assoc, pp. 12–18, 1985.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 591, + 505, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 620, + 502, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 504, + 633 + ], + "score": 1.0, + "content": "Andrew Gelman, John B Carlin, Hal S Stern, David B Dunson, Aki Vehtari, and Donald B Rubin.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 632, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 288, + 644 + ], + "score": 1.0, + "content": "Bayesian data analysis. CRC press, 2013.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 620, + 504, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 661, + 290, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 661, + 290, + 673 + ], + "score": 1.0, + "content": "networks. In AISTATS, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 650, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 502, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 504, + 693 + ], + "score": 1.0, + "content": "Markus Goldstein and Andreas Dengel. Histogram-based outlier score (hbos): A fast unsupervised", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 691, + 446, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 446, + 704 + ], + "score": 1.0, + "content": "anomaly detection algorithm. KI-2012: Poster and Demo Track, pp. 59–63, 2012.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 679, + 504, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 721, + 223, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 223, + 731 + ], + "score": 1.0, + "content": "networks. In ICML, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 709, + 505, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Johanna Hardin and David M Rocke. Outlier detection in the multiple cluster setting using the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "minimum covariance determinant estimator. Computational Statistics & Data Analysis, 44(4):", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 182, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 182, + 116 + ], + "score": 1.0, + "content": "625–638, 2004.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 124, + 504, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning: data", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 429, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 429, + 147 + ], + "score": 1.0, + "content": "mining, inference, and prediction. Springer Science & Business Media, 2009.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 502, + 177 + ], + "lines": [ + { + "bbox": [ + 104, + 152, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 505, + 170 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 212, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 212, + 177 + ], + "score": 1.0, + "content": "nition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 503, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "Zengyou He, Xiaofei Xu, and Shengchun Deng. Discovering cluster-based local outliers. Pattern", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 196, + 312, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 312, + 208 + ], + "score": 1.0, + "content": "Recognition Letters, 24(9-10):1641–1650, 2003.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 228, + 254, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 228, + 254, + 238 + ], + "score": 1.0, + "content": "preprint arXiv:1503.02531, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 258, + 281, + 269 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 281, + 269 + ], + "score": 1.0, + "content": "convolutional networks. In CVPR, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 397, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 396, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 396, + 291 + ], + "score": 1.0, + "content": "Peter J Huber. Robust statistics, volume 523. John Wiley & Sons, 2004.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 108, + 297, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 311 + ], + "score": 1.0, + "content": "Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 308, + 289, + 320 + ], + "spans": [ + { + "bbox": [ + 117, + 308, + 289, + 320 + ], + "score": 1.0, + "content": "vision? In NeurIPS, pp. 5574–5584, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 341 + ], + "score": 1.0, + "content": "Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 339, + 388, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 388, + 351 + ], + "score": 1.0, + "content": "for scene geometry and semantics. In CVPR, pp. 7482–7491, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 489, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 489, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 489, + 371 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 105, + 378, + 504, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 392 + ], + "score": 1.0, + "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 389, + 214, + 400 + ], + "spans": [ + { + "bbox": [ + 115, + 389, + 214, + 400 + ], + "score": 1.0, + "content": "arXiv:1312.6114, 2013.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 104, + 409, + 504, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 419, + 270, + 432 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 270, + 432 + ], + "score": 1.0, + "content": "AIChE journal, 37(2):233–243, 1991.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 439, + 503, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 504, + 452 + ], + "score": 1.0, + "content": "Hans-Peter Kriegel, Matthias Schubert, and Arthur Zimek. Angle-based outlier detection in high-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 116, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "dimensional data. In Proceedings of the 14th ACM SIGKDD international conference on Knowl-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 462, + 329, + 474 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 329, + 474 + ], + "score": 1.0, + "content": "edge discovery and data mining, pp. 444–452, 2008.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 481, + 503, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 495 + ], + "score": 1.0, + "content": "Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 492, + 189, + 503 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 189, + 503 + ], + "score": 1.0, + "content": "of Toronto, 2009.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 511, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "Volodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learn-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 523, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 506, + 535 + ], + "score": 1.0, + "content": "ing using calibrated regression. In Jennifer Dy and Andreas Krause (eds.), ICML, volume 80", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "of Proceedings of Machine Learning Research, pp. 2801–2809, Stockholmsmassan, Stockholm ¨", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 545, + 251, + 557 + ], + "spans": [ + { + "bbox": [ + 116, + 545, + 251, + 557 + ], + "score": 1.0, + "content": "Sweden, 10–15 Jul 2018. PMLR.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 506, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "score": 1.0, + "content": "Aleksandar Lazarevic and Vipin Kumar. Feature bagging for outlier detection. In Proceedings of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 574, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 505, + 590 + ], + "score": 1.0, + "content": "the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, pp.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 587, + 181, + 598 + ], + "spans": [ + { + "bbox": [ + 117, + 587, + 181, + 598 + ], + "score": 1.0, + "content": "157–166, 2005.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 620 + ], + "score": 1.0, + "content": "Xuelong Li, Quanmao Lu, Yongsheng Dong, and Dacheng Tao. Robust subspace clustering by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "cauchy loss function. IEEE transactions on neural networks and learning systems, 30(7):2067–", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 628, + 167, + 640 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 167, + 640 + ], + "score": 1.0, + "content": "2078, 2018.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 104, + 647, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "score": 1.0, + "content": "Zheng Li, Yue Zhao, Nicola Botta, Cezar Ionescu, and Xiyang Hu. Copod: copula-based outlier", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 659, + 321, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 659, + 321, + 671 + ], + "score": 1.0, + "content": "detection. arXiv preprint arXiv:2009.09463, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "score": 1.0, + "content": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 Eighth IEEE Interna-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 689, + 365, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 689, + 365, + 702 + ], + "score": 1.0, + "content": "tional Conference on Data Mining, pp. 413–422. IEEE, 2008.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation-based anomaly detection. ACM Trans-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 721, + 401, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 401, + 732 + ], + "score": 1.0, + "content": "actions on Knowledge Discovery from Data (TKDD), 6(1):1–39, 2012.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Johanna Hardin and David M Rocke. Outlier detection in the multiple cluster setting using the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "minimum covariance determinant estimator. Computational Statistics & Data Analysis, 44(4):", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 182, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 182, + 116 + ], + "score": 1.0, + "content": "625–638, 2004.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 124, + 504, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning: data", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 429, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 429, + 147 + ], + "score": 1.0, + "content": "mining, inference, and prediction. Springer Science & Business Media, 2009.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 123, + 505, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 502, + 177 + ], + "lines": [ + { + "bbox": [ + 104, + 152, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 505, + 170 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 212, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 212, + 177 + ], + "score": 1.0, + "content": "nition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 152, + 505, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 503, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "Zengyou He, Xiaofei Xu, and Shengchun Deng. Discovering cluster-based local outliers. Pattern", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 196, + 312, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 312, + 208 + ], + "score": 1.0, + "content": "Recognition Letters, 24(9-10):1641–1650, 2003.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 185, + 505, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 228, + 254, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 228, + 254, + 238 + ], + "score": 1.0, + "content": "preprint arXiv:1503.02531, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 216, + 505, + 238 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 258, + 281, + 269 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 281, + 269 + ], + "score": 1.0, + "content": "convolutional networks. In CVPR, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 246, + 505, + 269 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 397, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 396, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 396, + 291 + ], + "score": 1.0, + "content": "Peter J Huber. Robust statistics, volume 523. John Wiley & Sons, 2004.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 277, + 396, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 297, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 311 + ], + "score": 1.0, + "content": "Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 308, + 289, + 320 + ], + "spans": [ + { + "bbox": [ + 117, + 308, + 289, + 320 + ], + "score": 1.0, + "content": "vision? In NeurIPS, pp. 5574–5584, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 296, + 505, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 341 + ], + "score": 1.0, + "content": "Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 339, + 388, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 388, + 351 + ], + "score": 1.0, + "content": "for scene geometry and semantics. In CVPR, pp. 7482–7491, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 327, + 505, + 351 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 489, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 489, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 489, + 371 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 358, + 489, + 371 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 378, + 504, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 392 + ], + "score": 1.0, + "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 389, + 214, + 400 + ], + "spans": [ + { + "bbox": [ + 115, + 389, + 214, + 400 + ], + "score": 1.0, + "content": "arXiv:1312.6114, 2013.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 376, + 506, + 400 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 409, + 504, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 419, + 270, + 432 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 270, + 432 + ], + "score": 1.0, + "content": "AIChE journal, 37(2):233–243, 1991.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 408, + 505, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 439, + 503, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 504, + 452 + ], + "score": 1.0, + "content": "Hans-Peter Kriegel, Matthias Schubert, and Arthur Zimek. Angle-based outlier detection in high-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 116, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "dimensional data. In Proceedings of the 14th ACM SIGKDD international conference on Knowl-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 462, + 329, + 474 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 329, + 474 + ], + "score": 1.0, + "content": "edge discovery and data mining, pp. 444–452, 2008.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 439, + 505, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 481, + 503, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 495 + ], + "score": 1.0, + "content": "Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 492, + 189, + 503 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 189, + 503 + ], + "score": 1.0, + "content": "of Toronto, 2009.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 479, + 505, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 511, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "Volodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learn-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 523, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 506, + 535 + ], + "score": 1.0, + "content": "ing using calibrated regression. In Jennifer Dy and Andreas Krause (eds.), ICML, volume 80", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "of Proceedings of Machine Learning Research, pp. 2801–2809, Stockholmsmassan, Stockholm ¨", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 545, + 251, + 557 + ], + "spans": [ + { + "bbox": [ + 116, + 545, + 251, + 557 + ], + "score": 1.0, + "content": "Sweden, 10–15 Jul 2018. PMLR.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 513, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 506, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "score": 1.0, + "content": "Aleksandar Lazarevic and Vipin Kumar. Feature bagging for outlier detection. In Proceedings of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 574, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 505, + 590 + ], + "score": 1.0, + "content": "the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, pp.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 587, + 181, + 598 + ], + "spans": [ + { + "bbox": [ + 117, + 587, + 181, + 598 + ], + "score": 1.0, + "content": "157–166, 2005.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 563, + 506, + 598 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 620 + ], + "score": 1.0, + "content": "Xuelong Li, Quanmao Lu, Yongsheng Dong, and Dacheng Tao. Robust subspace clustering by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "cauchy loss function. IEEE transactions on neural networks and learning systems, 30(7):2067–", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 628, + 167, + 640 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 167, + 640 + ], + "score": 1.0, + "content": "2078, 2018.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 605, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 647, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 505, + 661 + ], + "score": 1.0, + "content": "Zheng Li, Yue Zhao, Nicola Botta, Cezar Ionescu, and Xiyang Hu. Copod: copula-based outlier", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 659, + 321, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 659, + 321, + 671 + ], + "score": 1.0, + "content": "detection. arXiv preprint arXiv:2009.09463, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 647, + 505, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "score": 1.0, + "content": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 Eighth IEEE Interna-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 689, + 365, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 689, + 365, + 702 + ], + "score": 1.0, + "content": "tional Conference on Data Mining, pp. 413–422. IEEE, 2008.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 677, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation-based anomaly detection. ACM Trans-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 721, + 401, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 401, + 732 + ], + "score": 1.0, + "content": "actions on Knowledge Discovery from Data (TKDD), 6(1):1–39, 2012.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 709, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 223, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 223, + 105 + ], + "score": 1.0, + "content": "In ICCV, December 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 503, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "David JC MacKay and David JC Mac Kay. Information theory, inference and learning algorithms.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 121, + 256, + 134 + ], + "spans": [ + { + "bbox": [ + 117, + 121, + 256, + 134 + ], + "score": 1.0, + "content": "Cambridge university press, 2003.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 504, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "score": 1.0, + "content": "Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated prob-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "abilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 160, + 142, + 172 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 142, + 172 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 177, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "John Ashworth Nelder and Robert WM Wedderburn. Generalized linear models. Journal of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 189, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 116, + 189, + 394, + 201 + ], + "score": 1.0, + "content": "Royal Statistical Society: Series A (General), 135(3):370–384, 1972.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 108, + 205, + 504, + 229 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 220 + ], + "score": 1.0, + "content": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 217, + 382, + 230 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 382, + 230 + ], + "score": 1.0, + "content": "digits in natural images with unsupervised feature learning. 2011.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 503, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 504, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 504, + 247 + ], + "score": 1.0, + "content": "David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 245, + 346, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 245, + 346, + 258 + ], + "score": 1.0, + "content": "distribution. In ICNN, volume 1, pp. 55–60. IEEE, 1994.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 262, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 276 + ], + "score": 1.0, + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 116, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 282, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 505, + 298 + ], + "score": 1.0, + "content": "Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 8024–8035. Curran", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 329, + 209, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 329, + 209, + 339 + ], + "score": 1.0, + "content": "Associates, Inc., 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier ¨", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 356, + 504, + 369 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 504, + 369 + ], + "score": 1.0, + "content": "Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 367, + 497, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 367, + 497, + 380 + ], + "score": 1.0, + "content": "Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 503, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "John C Platt. Probabilistic outputs for support vector machines and comparisons to regularized", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 394, + 428, + 409 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 428, + 409 + ], + "score": 1.0, + "content": "likelihood methods. Advances in large margin classifiers, 10(3):61–74, 1999.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 412, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. Efficient algorithms for mining outliers", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 116, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "from large data sets. In Proceedings of the 2000 ACM SIGMOD international conference on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 434, + 283, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 434, + 283, + 446 + ], + "score": 1.0, + "content": "Management of data, pp. 427–438, 2000.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 366, + 463 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 367, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 367, + 465 + ], + "score": 1.0, + "content": "Shebuti Rayana. Odds library. Stony Brook,-2016, (2017), 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "Peter J Rousseeuw and Katrien Van Driessen. A fast algorithm for the minimum covariance deter-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 479, + 340, + 491 + ], + "spans": [ + { + "bbox": [ + 116, + 479, + 340, + 491 + ], + "score": 1.0, + "content": "minant estimator. Technometrics, 41(3):212–223, 1999.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 105, + 496, + 503, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "Shreyas Saxena, Oncel Tuzel, and Dennis DeCoste. Data parameters: A new family of parameters", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 508, + 428, + 520 + ], + "spans": [ + { + "bbox": [ + 116, + 508, + 428, + 520 + ], + "score": 1.0, + "content": "for learning a differentiable curriculum. In NeurIPS, pp. 11093–11103, 2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 503, + 558 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 504, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 504, + 537 + ], + "score": 1.0, + "content": "Bernhard Scholkopf, Robert C Williamson, Alex J Smola, John Shawe-Taylor, and John C Platt.¨", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 535, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 115, + 535, + 505, + 550 + ], + "score": 1.0, + "content": "Support vector method for novelty detection. In Advances in neural information processing sys-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 547, + 219, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 219, + 558 + ], + "score": 1.0, + "content": "tems, pp. 582–588, 2000.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and LiWu Chang. A novel anomaly de-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "tection scheme based on principal component classifier. Technical report, MIAMI UNIV CORAL", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 587, + 456, + 597 + ], + "spans": [ + { + "bbox": [ + 117, + 587, + 456, + 597 + ], + "score": 1.0, + "content": "GABLES FL DEPT OF ELECTRICAL AND COMPUTER ENGINEERING, 2003.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 504, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "Henri Theil. A rank-invariant method of linear and polynomial regression analysis. In Henri Theil’s", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 614, + 421, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 614, + 421, + 626 + ], + "score": 1.0, + "content": "contributions to economics and econometrics, pp. 345–381. Springer, 1992.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 645 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 116, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26–", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 653, + 157, + 665 + ], + "spans": [ + { + "bbox": [ + 115, + 653, + 157, + 665 + ], + "score": 1.0, + "content": "31, 2012.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 502, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 681, + 438, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 438, + 694 + ], + "score": 1.0, + "content": "and naive bayesian classifiers. In ICML, volume 1, pp. 609–616. Citeseer, 2001.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detec-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "tion. Journal of Machine Learning Research, 20(96):1–7, 2019. URL http://jmlr.org/", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 254, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 254, + 731 + ], + "score": 1.0, + "content": "papers/v20/19-011.html.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 223, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 223, + 105 + ], + "score": 1.0, + "content": "In ICCV, December 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 503, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "David JC MacKay and David JC Mac Kay. Information theory, inference and learning algorithms.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 121, + 256, + 134 + ], + "spans": [ + { + "bbox": [ + 117, + 121, + 256, + 134 + ], + "score": 1.0, + "content": "Cambridge university press, 2003.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 109, + 505, + 134 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 138, + 504, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "score": 1.0, + "content": "Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated prob-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "abilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 160, + 142, + 172 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 142, + 172 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 138, + 506, + 172 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 177, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "John Ashworth Nelder and Robert WM Wedderburn. Generalized linear models. Journal of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 189, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 116, + 189, + 394, + 201 + ], + "score": 1.0, + "content": "Royal Statistical Society: Series A (General), 135(3):370–384, 1972.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 177, + 505, + 201 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 205, + 504, + 229 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 220 + ], + "score": 1.0, + "content": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 217, + 382, + 230 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 382, + 230 + ], + "score": 1.0, + "content": "digits in natural images with unsupervised feature learning. 2011.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 204, + 505, + 230 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 503, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 504, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 504, + 247 + ], + "score": 1.0, + "content": "David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 245, + 346, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 245, + 346, + 258 + ], + "score": 1.0, + "content": "distribution. In ICNN, volume 1, pp. 55–60. IEEE, 1994.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 233, + 504, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 262, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 276 + ], + "score": 1.0, + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 116, + 273, + 505, + 286 + ], + "score": 1.0, + "content": "Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 282, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 282, + 505, + 298 + ], + "score": 1.0, + "content": "Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 8024–8035. Curran", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 329, + 209, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 329, + 209, + 339 + ], + "score": 1.0, + "content": "Associates, Inc., 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 261, + 506, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier ¨", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 356, + 504, + 369 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 504, + 369 + ], + "score": 1.0, + "content": "Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 367, + 497, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 367, + 497, + 380 + ], + "score": 1.0, + "content": "Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 345, + 506, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 503, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "John C Platt. Probabilistic outputs for support vector machines and comparisons to regularized", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 394, + 428, + 409 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 428, + 409 + ], + "score": 1.0, + "content": "likelihood methods. Advances in large margin classifiers, 10(3):61–74, 1999.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 384, + 505, + 409 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 412, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. Efficient algorithms for mining outliers", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 116, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "from large data sets. In Proceedings of the 2000 ACM SIGMOD international conference on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 434, + 283, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 434, + 283, + 446 + ], + "score": 1.0, + "content": "Management of data, pp. 427–438, 2000.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 412, + 505, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 366, + 463 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 367, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 367, + 465 + ], + "score": 1.0, + "content": "Shebuti Rayana. Odds library. Stony Brook,-2016, (2017), 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 450, + 367, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "Peter J Rousseeuw and Katrien Van Driessen. A fast algorithm for the minimum covariance deter-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 479, + 340, + 491 + ], + "spans": [ + { + "bbox": [ + 116, + 479, + 340, + 491 + ], + "score": 1.0, + "content": "minant estimator. Technometrics, 41(3):212–223, 1999.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 468, + 505, + 491 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 496, + 503, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "Shreyas Saxena, Oncel Tuzel, and Dennis DeCoste. Data parameters: A new family of parameters", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 508, + 428, + 520 + ], + "spans": [ + { + "bbox": [ + 116, + 508, + 428, + 520 + ], + "score": 1.0, + "content": "for learning a differentiable curriculum. In NeurIPS, pp. 11093–11103, 2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 497, + 505, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 503, + 558 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 504, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 504, + 537 + ], + "score": 1.0, + "content": "Bernhard Scholkopf, Robert C Williamson, Alex J Smola, John Shawe-Taylor, and John C Platt.¨", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 535, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 115, + 535, + 505, + 550 + ], + "score": 1.0, + "content": "Support vector method for novelty detection. In Advances in neural information processing sys-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 547, + 219, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 219, + 558 + ], + "score": 1.0, + "content": "tems, pp. 582–588, 2000.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 525, + 505, + 558 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and LiWu Chang. A novel anomaly de-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "tection scheme based on principal component classifier. Technical report, MIAMI UNIV CORAL", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 587, + 456, + 597 + ], + "spans": [ + { + "bbox": [ + 117, + 587, + 456, + 597 + ], + "score": 1.0, + "content": "GABLES FL DEPT OF ELECTRICAL AND COMPUTER ENGINEERING, 2003.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 564, + 505, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 504, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 616 + ], + "score": 1.0, + "content": "Henri Theil. A rank-invariant method of linear and polynomial regression analysis. In Henri Theil’s", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 614, + 421, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 614, + 421, + 626 + ], + "score": 1.0, + "content": "contributions to economics and econometrics, pp. 345–381. Springer, 1992.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 603, + 505, + 626 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 645 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 116, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26–", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 653, + 157, + 665 + ], + "spans": [ + { + "bbox": [ + 115, + 653, + 157, + 665 + ], + "score": 1.0, + "content": "31, 2012.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 630, + 505, + 665 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 502, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 681, + 438, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 438, + 694 + ], + "score": 1.0, + "content": "and naive bayesian classifiers. In ICML, volume 1, pp. 609–616. Citeseer, 2001.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 669, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detec-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "tion. Journal of Machine Learning Research, 20(96):1–7, 2019. URL http://jmlr.org/", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 254, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 254, + 731 + ], + "score": 1.0, + "content": "papers/v20/19-011.html.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 106, + 698, + 506, + 731 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_model.json b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_model.json new file mode 100644 index 0000000000000000000000000000000000000000..511958da3c34dbfa6e3b7bdc1ac28c5196de9cbb --- /dev/null +++ b/parse/train/KCzRX9N8BIH/KCzRX9N8BIH_model.json @@ -0,0 +1,17883 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 622, + 1302, + 622, + 1302, + 957, + 398, + 957 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1844, + 1402, + 1844, + 1402, + 1969, + 298, + 1969 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1231, + 1404, + 1231, + 1404, + 1507, + 298, + 1507 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1092, + 1403, + 1092, + 1403, + 1216, + 299, + 1216 + ], + "score": 0.97 + }, + { + "category_id": 0, + "poly": [ + 298, + 219, + 945, + 219, + 945, + 324, + 298, + 324 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 362, + 1537, + 1404, + 1537, + 1404, + 1728, + 362, + 1728 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 657, + 375, + 980, + 375, + 980, + 468, + 657, + 468 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 689, + 1986, + 1015, + 1986, + 1015, + 2042, + 689, + 2042 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 313, + 375, + 549, + 375, + 549, + 468, + 313, + 468 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 1085, + 376, + 1304, + 376, + 1304, + 469, + 1085, + 469 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 299, + 1772, + 558, + 1772, + 558, + 1809, + 299, + 1809 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 302, + 1021, + 573, + 1021, + 573, + 1056, + 302, + 1056 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 773, + 549, + 927, + 549, + 927, + 582, + 773, + 582 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1991, + 1400, + 1991, + 1400, + 2021, + 1366, + 2021 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.761 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.754 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 855, + 75, + 855, + 105, + 298, + 105 + ], + "score": 0.148 + }, + { + "category_id": 13, + "poly": [ + 660, + 1907, + 788, + 1907, + 788, + 1940, + 660, + 1940 + ], + "score": 0.94, + "latex": "\\hat { y } _ { i } = f _ { \\theta } ( x )" + }, + { + "category_id": 14, + "poly": [ + 686, + 1986, + 1016, + 1986, + 1016, + 2042, + 686, + 2042 + ], + "score": 0.94, + "latex": "\\operatorname* { m i n } _ { \\theta } \\underset { ( x , y ) \\sim \\mathcal { D } } { \\mathbb { E } } L ( \\hat { y } = f _ { \\theta } ( x ) , y )" + }, + { + "category_id": 13, + "poly": [ + 885, + 1906, + 968, + 1906, + 968, + 1940, + 885, + 1940 + ], + "score": 0.93, + "latex": "L ( \\hat { y } , y )" + }, + { + "category_id": 13, + "poly": [ + 1126, + 1845, + 1302, + 1845, + 1302, + 1879, + 1126, + 1879 + ], + "score": 0.92, + "latex": "i \\in \\{ 1 , \\ldots , N \\}" + }, + { + "category_id": 13, + "poly": [ + 490, + 1186, + 524, + 1186, + 524, + 1216, + 490, + 1216 + ], + "score": 0.89, + "latex": "L _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1216, + 1156, + 1251, + 1156, + 1251, + 1185, + 1216, + 1185 + ], + "score": 0.88, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 563, + 896, + 598, + 896, + 598, + 926, + 563, + 926 + ], + "score": 0.88, + "latex": "L _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 476, + 896, + 511, + 896, + 511, + 926, + 476, + 926 + ], + "score": 0.87, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1324, + 1878, + 1345, + 1878, + 1345, + 1909, + 1324, + 1909 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 964, + 1851, + 989, + 1851, + 989, + 1877, + 964, + 1877 + ], + "score": 0.85, + "latex": "y _ { i }" + }, + { + "category_id": 13, + "poly": [ + 803, + 1852, + 830, + 1852, + 830, + 1876, + 803, + 1876 + ], + "score": 0.85, + "latex": "x _ { i }" + }, + { + "category_id": 13, + "poly": [ + 666, + 1847, + 692, + 1847, + 692, + 1873, + 666, + 1873 + ], + "score": 0.82, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 516, + 1450, + 536, + 1450, + 536, + 1478, + 516, + 1478 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 845, + 1665, + 881, + 1665, + 881, + 1692, + 845, + 1692 + ], + "score": 0.81, + "latex": "L 2" + }, + { + "category_id": 13, + "poly": [ + 761, + 1665, + 796, + 1665, + 796, + 1692, + 761, + 1692 + ], + "score": 0.8, + "latex": "L 1" + }, + { + "category_id": 13, + "poly": [ + 427, + 1908, + 444, + 1908, + 444, + 1934, + 427, + 1934 + ], + "score": 0.78, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1132, + 1575, + 1151, + 1575, + 1151, + 1602, + 1132, + 1602 + ], + "score": 0.71, + "latex": "\\rho" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 215.0, + 949.0, + 215.0, + 949.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 271.0, + 861.0, + 271.0, + 861.0, + 329.0, + 295.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1769.0, + 561.0, + 1769.0, + 561.0, + 1816.0, + 291.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1018.0, + 579.0, + 1018.0, + 579.0, + 1065.0, + 294.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 545.0, + 934.0, + 545.0, + 934.0, + 588.0, + 768.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 855.0, + 73.0, + 855.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 621.0, + 1304.0, + 621.0, + 1304.0, + 656.0, + 394.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 649.0, + 1304.0, + 649.0, + 1304.0, + 687.0, + 393.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 682.0, + 1303.0, + 682.0, + 1303.0, + 715.0, + 394.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 712.0, + 1307.0, + 712.0, + 1307.0, + 748.0, + 394.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 743.0, + 1306.0, + 743.0, + 1306.0, + 776.0, + 393.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 774.0, + 1307.0, + 774.0, + 1307.0, + 806.0, + 393.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 803.0, + 1304.0, + 803.0, + 1304.0, + 838.0, + 393.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 833.0, + 1306.0, + 833.0, + 1306.0, + 869.0, + 393.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 865.0, + 1304.0, + 865.0, + 1304.0, + 899.0, + 394.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 897.0, + 475.0, + 897.0, + 475.0, + 930.0, + 395.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 897.0, + 562.0, + 897.0, + 562.0, + 930.0, + 512.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 897.0, + 1305.0, + 897.0, + 1305.0, + 930.0, + 599.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 924.0, + 1080.0, + 924.0, + 1080.0, + 962.0, + 392.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1843.0, + 665.0, + 1843.0, + 665.0, + 1883.0, + 292.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1843.0, + 802.0, + 1843.0, + 802.0, + 1883.0, + 693.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1843.0, + 963.0, + 1843.0, + 963.0, + 1883.0, + 831.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1843.0, + 1125.0, + 1843.0, + 1125.0, + 1883.0, + 990.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 1843.0, + 1407.0, + 1843.0, + 1407.0, + 1883.0, + 1303.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1876.0, + 1323.0, + 1876.0, + 1323.0, + 1909.0, + 293.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1876.0, + 1404.0, + 1876.0, + 1404.0, + 1909.0, + 1346.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1906.0, + 426.0, + 1906.0, + 426.0, + 1943.0, + 293.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1906.0, + 659.0, + 1906.0, + 659.0, + 1943.0, + 445.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1906.0, + 884.0, + 1906.0, + 884.0, + 1943.0, + 789.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1906.0, + 1406.0, + 1906.0, + 1406.0, + 1943.0, + 969.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1938.0, + 1249.0, + 1938.0, + 1249.0, + 1973.0, + 293.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1233.0, + 1402.0, + 1233.0, + 1402.0, + 1262.0, + 297.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1259.0, + 1406.0, + 1259.0, + 1406.0, + 1300.0, + 293.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1291.0, + 1406.0, + 1291.0, + 1406.0, + 1329.0, + 293.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1322.0, + 1406.0, + 1322.0, + 1406.0, + 1359.0, + 292.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1353.0, + 1406.0, + 1353.0, + 1406.0, + 1390.0, + 292.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1385.0, + 1406.0, + 1385.0, + 1406.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1416.0, + 1405.0, + 1416.0, + 1405.0, + 1449.0, + 294.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1443.0, + 515.0, + 1443.0, + 515.0, + 1483.0, + 293.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1443.0, + 1406.0, + 1443.0, + 1406.0, + 1483.0, + 537.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1477.0, + 864.0, + 1477.0, + 864.0, + 1510.0, + 297.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1093.0, + 1404.0, + 1093.0, + 1404.0, + 1126.0, + 295.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1125.0, + 1405.0, + 1125.0, + 1405.0, + 1157.0, + 295.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1156.0, + 1215.0, + 1156.0, + 1215.0, + 1189.0, + 295.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1156.0, + 1405.0, + 1156.0, + 1405.0, + 1189.0, + 1252.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1187.0, + 489.0, + 1187.0, + 489.0, + 1219.0, + 295.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1187.0, + 768.0, + 1187.0, + 768.0, + 1219.0, + 525.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1575.0, + 360.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1568.0, + 1131.0, + 1568.0, + 1131.0, + 1606.0, + 393.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1568.0, + 1158.0, + 1568.0, + 1158.0, + 1606.0, + 1152.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1597.0, + 1406.0, + 1597.0, + 1406.0, + 1639.0, + 355.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1634.0, + 603.0, + 1634.0, + 603.0, + 1663.0, + 395.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1661.0, + 760.0, + 1661.0, + 760.0, + 1700.0, + 357.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1661.0, + 844.0, + 1661.0, + 844.0, + 1700.0, + 797.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1661.0, + 1406.0, + 1661.0, + 1406.0, + 1700.0, + 882.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1695.0, + 809.0, + 1695.0, + 809.0, + 1730.0, + 393.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 377.0, + 860.0, + 377.0, + 860.0, + 406.0, + 657.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 404.0, + 910.0, + 404.0, + 910.0, + 438.0, + 655.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 439.0, + 981.0, + 439.0, + 981.0, + 469.0, + 658.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 373.0, + 508.0, + 373.0, + 508.0, + 408.0, + 309.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 406.0, + 494.0, + 406.0, + 494.0, + 436.0, + 311.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 440.0, + 552.0, + 440.0, + 552.0, + 468.0, + 311.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 373.0, + 1299.0, + 373.0, + 1299.0, + 409.0, + 1084.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 404.0, + 1239.0, + 404.0, + 1239.0, + 440.0, + 1083.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 439.0, + 1307.0, + 439.0, + 1307.0, + 468.0, + 1084.0, + 468.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1034, + 1405, + 1034, + 1405, + 1289, + 296, + 1289 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 866, + 1404, + 866, + 1404, + 1021, + 297, + 1021 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 684, + 1405, + 684, + 1405, + 779, + 297, + 779 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1691, + 1405, + 1691, + 1405, + 1787, + 298, + 1787 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1445, + 1404, + 1445, + 1404, + 1540, + 299, + 1540 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 307, + 232, + 1387, + 232, + 1387, + 529, + 307, + 529 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 586, + 1350, + 1111, + 1350, + 1111, + 1427, + 586, + 1427 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 681, + 794, + 1025, + 794, + 1025, + 850, + 681, + 850 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 564, + 1848, + 1170, + 1848, + 1170, + 1954, + 564, + 1954 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 526, + 1962, + 1084, + 1962, + 1084, + 2029, + 526, + 2029 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 655, + 1603, + 1041, + 1603, + 1041, + 1678, + 655, + 1678 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 300, + 1800, + 733, + 1800, + 733, + 1832, + 300, + 1832 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 297, + 1302, + 1150, + 1302, + 1150, + 1336, + 297, + 1336 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 297, + 1553, + 1232, + 1553, + 1232, + 1587, + 297, + 1587 + ], + "score": 0.923 + }, + { + "category_id": 4, + "poly": [ + 290, + 562, + 1404, + 562, + 1404, + 625, + 290, + 625 + ], + "score": 0.92 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1374, + 1400, + 1374, + 1400, + 1405, + 1366, + 1405 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1623, + 1400, + 1623, + 1400, + 1652, + 1366, + 1652 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1366, + 799, + 1400, + 799, + 1400, + 829, + 1366, + 829 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1978, + 1401, + 1978, + 1401, + 2008, + 1365, + 2008 + ], + "score": 0.875 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1888, + 1401, + 1888, + 1401, + 1918, + 1365, + 1918 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.747 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 854, + 76, + 854, + 104, + 298, + 104 + ], + "score": 0.741 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 852, + 76, + 852, + 104, + 299, + 104 + ], + "score": 0.37 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.14 + }, + { + "category_id": 14, + "poly": [ + 531, + 1846, + 1172, + 1846, + 1172, + 2033, + 531, + 2033 + ], + "score": 0.95, + "latex": "{ \\begin{array} { r l } & { \\rho \\left( x , \\alpha , \\sigma \\right) = { \\frac { \\left| \\alpha - 2 \\right| } { \\alpha } } \\left( \\left( { \\frac { \\left( x / \\sigma \\right) ^ { 2 } } { \\left| \\alpha - 2 \\right| } } + 1 \\right) ^ { \\alpha / 2 } - 1 \\right) { \\mathrm { a n d } } } \\\\ & { p \\left( { \\hat { y } } \\mid y , \\alpha , \\sigma \\right) = { \\frac { 1 } { \\sigma Z \\left( \\alpha \\right) } } \\exp \\left( - \\rho \\left( { \\hat { y } } - y , \\alpha , \\sigma \\right) \\right) , } \\end{array} }" + }, + { + "category_id": 14, + "poly": [ + 586, + 1350, + 1110, + 1350, + 1110, + 1428, + 586, + 1428 + ], + "score": 0.94, + "latex": "\\mathcal { N } ( \\hat { y } | y , \\sigma ) = ( 2 \\pi \\sigma ^ { 2 } ) ^ { - \\frac 1 2 } \\exp \\left( - \\frac 1 2 \\frac { ( \\hat { y } - y ) ^ { 2 } } { \\sigma ^ { 2 } } \\right) ," + }, + { + "category_id": 14, + "poly": [ + 679, + 794, + 1026, + 794, + 1026, + 850, + 679, + 850 + ], + "score": 0.94, + "latex": "\\operatorname* { m i n } _ { \\theta , \\phi } \\quad \\mathbb { E } _ { \\mathbf { \\Phi } ( x , y ) \\sim \\mathcal { D } } \\ell ( \\hat { y } = f _ { \\theta } ( x ) | y , \\phi )" + }, + { + "category_id": 14, + "poly": [ + 657, + 1601, + 1039, + 1601, + 1039, + 1677, + 657, + 1677 + ], + "score": 0.93, + "latex": "\\mathrm { s o f t m a x } ( \\hat { y } = y | z , \\tau ) = \\frac { e ^ { z _ { y } \\tau } } { \\sum _ { c } e ^ { z _ { c } \\tau } } ," + }, + { + "category_id": 13, + "poly": [ + 1105, + 1444, + 1395, + 1444, + 1395, + 1483, + 1105, + 1483 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\ell _ { N } = \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } - y ) ^ { 2 } + \\log \\sigma , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 450, + 686, + 559, + 686, + 559, + 720, + 450, + 720 + ], + "score": 0.93, + "latex": "\\mathcal { L } ( \\hat { y } | y , \\phi )" + }, + { + "category_id": 13, + "poly": [ + 353, + 1446, + 481, + 1446, + 481, + 1480, + 353, + 1480 + ], + "score": 0.93, + "latex": "\\sigma \\in ( 0 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 542, + 1693, + 669, + 1693, + 669, + 1725, + 542, + 1725 + ], + "score": 0.92, + "latex": "\\tau \\in ( 0 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 838, + 1305, + 905, + 1305, + 905, + 1336, + 838, + 1336 + ], + "score": 0.91, + "latex": "{ \\hat { y } } - y" + }, + { + "category_id": 13, + "poly": [ + 586, + 1724, + 756, + 1724, + 756, + 1757, + 586, + 1757 + ], + "score": 0.9, + "latex": "- \\log p ( \\hat { y } = y )" + }, + { + "category_id": 13, + "poly": [ + 1341, + 961, + 1402, + 961, + 1402, + 992, + 1341, + 992 + ], + "score": 0.9, + "latex": "{ \\hat { y } } - y" + }, + { + "category_id": 13, + "poly": [ + 937, + 1725, + 1006, + 1725, + 1006, + 1752, + 937, + 1752 + ], + "score": 0.9, + "latex": "\\tau = 1" + }, + { + "category_id": 13, + "poly": [ + 529, + 1509, + 599, + 1509, + 599, + 1536, + 529, + 1536 + ], + "score": 0.89, + "latex": "\\sigma = 1" + }, + { + "category_id": 13, + "poly": [ + 363, + 961, + 384, + 961, + 384, + 991, + 363, + 991 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1204, + 869, + 1223, + 869, + 1223, + 900, + 1204, + 900 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1109, + 961, + 1129, + 961, + 1129, + 992, + 1109, + 992 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 679, + 718, + 698, + 718, + 698, + 748, + 679, + 748 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1267, + 688, + 1286, + 688, + 1286, + 718, + 1267, + 718 + ], + "score": 0.82, + "latex": "\\hat { y }" + }, + { + "category_id": 13, + "poly": [ + 473, + 1807, + 491, + 1807, + 491, + 1833, + 473, + 1833 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 927, + 1759, + 946, + 1759, + 946, + 1781, + 927, + 1781 + ], + "score": 0.8, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1311, + 1116, + 1311, + 1116, + 1331, + 1096, + 1331 + ], + "score": 0.79, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 366, + 722, + 385, + 722, + 385, + 748, + 366, + 748 + ], + "score": 0.79, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 654, + 1101, + 672, + 1101, + 672, + 1130, + 654, + 1130 + ], + "score": 0.79, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1009, + 1562, + 1026, + 1562, + 1026, + 1581, + 1009, + 1581 + ], + "score": 0.79, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 857, + 1759, + 876, + 1759, + 876, + 1781, + 857, + 1781 + ], + "score": 0.78, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 297, + 960, + 315, + 960, + 315, + 987, + 297, + 987 + ], + "score": 0.78, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1103, + 1121, + 1103, + 1121, + 1125, + 1100, + 1125 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1188, + 1562, + 1204, + 1562, + 1204, + 1581, + 1188, + 1581 + ], + "score": 0.77, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 1218, + 1073, + 1238, + 1073, + 1238, + 1095, + 1218, + 1095 + ], + "score": 0.76, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 969, + 1103, + 991, + 1103, + 991, + 1125, + 969, + 1125 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 520, + 935, + 537, + 935, + 537, + 960, + 520, + 960 + ], + "score": 0.72, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1380, + 1043, + 1401, + 1043, + 1401, + 1065, + 1380, + 1065 + ], + "score": 0.7, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1096, + 718, + 1112, + 718, + 1112, + 745, + 1096, + 745 + ], + "score": 0.68, + "latex": "\\ell" + }, + { + "category_id": 13, + "poly": [ + 1267, + 246, + 1329, + 246, + 1329, + 265, + 1267, + 265 + ], + "score": 0.27, + "latex": "\\sigma \\left( \\tau = 1 \\right)" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 230.0, + 346.0, + 230.0, + 346.0, + 259.0, + 311.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 239.0, + 602.0, + 239.0, + 602.0, + 250.0, + 591.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 251.0, + 891.0, + 251.0, + 891.0, + 275.0, + 862.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 248.0, + 1144.0, + 248.0, + 1144.0, + 259.0, + 1135.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 242.0, + 1266.0, + 242.0, + 1266.0, + 270.0, + 1259.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 242.0, + 1336.0, + 242.0, + 1336.0, + 270.0, + 1330.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 273.0, + 346.0, + 273.0, + 346.0, + 300.0, + 311.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 284.0, + 601.0, + 284.0, + 601.0, + 290.0, + 595.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 290.0, + 891.0, + 290.0, + 891.0, + 314.0, + 862.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 282.0, + 1144.0, + 282.0, + 1144.0, + 294.0, + 1133.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 261.0, + 1369.0, + 261.0, + 1369.0, + 310.0, + 1259.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 314.0, + 346.0, + 314.0, + 346.0, + 342.0, + 311.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 316.0, + 1143.0, + 316.0, + 1143.0, + 328.0, + 1133.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 306.0, + 1384.0, + 306.0, + 1384.0, + 330.0, + 1261.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 321.0, + 602.0, + 321.0, + 602.0, + 333.0, + 591.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 328.0, + 891.0, + 328.0, + 891.0, + 353.0, + 861.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 357.0, + 343.0, + 357.0, + 343.0, + 382.0, + 312.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 363.0, + 602.0, + 363.0, + 602.0, + 376.0, + 591.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 368.0, + 891.0, + 368.0, + 891.0, + 393.0, + 861.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 382.0, + 1146.0, + 382.0, + 1146.0, + 398.0, + 1132.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 399.0, + 342.0, + 399.0, + 342.0, + 424.0, + 312.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 413.0, + 731.0, + 413.0, + 731.0, + 433.0, + 637.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 408.0, + 845.0, + 408.0, + 845.0, + 433.0, + 767.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 407.0, + 891.0, + 407.0, + 891.0, + 431.0, + 861.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 430.0, + 745.0, + 430.0, + 745.0, + 450.0, + 637.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 428.0, + 845.0, + 428.0, + 845.0, + 452.0, + 766.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 463.0, + 490.0, + 463.0, + 490.0, + 498.0, + 429.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 465.0, + 754.0, + 465.0, + 754.0, + 498.0, + 691.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 446.0, + 891.0, + 446.0, + 891.0, + 469.0, + 861.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 456.0, + 924.0, + 456.0, + 924.0, + 475.0, + 896.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 459.0, + 967.0, + 459.0, + 967.0, + 474.0, + 950.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 454.0, + 1020.0, + 454.0, + 1020.0, + 493.0, + 993.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 456.0, + 1067.0, + 456.0, + 1067.0, + 477.0, + 1046.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 455.0, + 1120.0, + 455.0, + 1120.0, + 479.0, + 1092.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 450.0, + 1146.0, + 450.0, + 1146.0, + 466.0, + 1130.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 455.0, + 1183.0, + 455.0, + 1183.0, + 477.0, + 1147.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 456.0, + 1227.0, + 456.0, + 1227.0, + 476.0, + 1201.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 454.0, + 1278.0, + 454.0, + 1278.0, + 492.0, + 1250.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 457.0, + 1321.0, + 457.0, + 1321.0, + 475.0, + 1305.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 456.0, + 1376.0, + 456.0, + 1376.0, + 476.0, + 1351.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 496.0, + 740.0, + 496.0, + 740.0, + 524.0, + 422.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 496.0, + 1288.0, + 496.0, + 1288.0, + 524.0, + 960.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 424.5, + 523.0, + 424.5, + 523.0, + 446.5, + 509.0, + 446.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 439.5, + 594.0, + 439.5, + 594.0, + 480.5, + 307.0, + 480.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 440.0, + 855.0, + 440.0, + 855.0, + 478.5, + 581.0, + 478.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 440.5, + 1289.0, + 440.5, + 1289.0, + 454.5, + 1273.0, + 454.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 558.0, + 1405.0, + 558.0, + 1405.0, + 601.0, + 292.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 591.0, + 678.0, + 591.0, + 678.0, + 629.0, + 292.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1031.0, + 1379.0, + 1031.0, + 1379.0, + 1073.0, + 291.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1031.0, + 1406.0, + 1031.0, + 1406.0, + 1073.0, + 1402.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1063.0, + 1217.0, + 1063.0, + 1217.0, + 1102.0, + 293.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1063.0, + 1406.0, + 1063.0, + 1406.0, + 1102.0, + 1239.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1094.0, + 653.0, + 1094.0, + 653.0, + 1132.0, + 294.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1094.0, + 968.0, + 1094.0, + 968.0, + 1132.0, + 673.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1094.0, + 1099.0, + 1094.0, + 1099.0, + 1132.0, + 992.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1094.0, + 1407.0, + 1094.0, + 1407.0, + 1132.0, + 1122.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1127.0, + 1403.0, + 1127.0, + 1403.0, + 1162.0, + 295.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1159.0, + 1403.0, + 1159.0, + 1403.0, + 1194.0, + 295.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1189.0, + 1403.0, + 1189.0, + 1403.0, + 1234.0, + 296.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 1405.0, + 1224.0, + 1405.0, + 1259.0, + 293.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1253.0, + 859.0, + 1253.0, + 859.0, + 1290.0, + 293.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 868.0, + 1203.0, + 868.0, + 1203.0, + 901.0, + 296.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 868.0, + 1404.0, + 868.0, + 1404.0, + 901.0, + 1224.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 898.0, + 1405.0, + 898.0, + 1405.0, + 932.0, + 295.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 924.0, + 519.0, + 924.0, + 519.0, + 967.0, + 290.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 924.0, + 1408.0, + 924.0, + 1408.0, + 967.0, + 538.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 953.0, + 296.0, + 953.0, + 296.0, + 1001.0, + 291.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 953.0, + 362.0, + 953.0, + 362.0, + 1001.0, + 316.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 953.0, + 1108.0, + 953.0, + 1108.0, + 1001.0, + 385.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 953.0, + 1340.0, + 953.0, + 1340.0, + 1001.0, + 1130.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 953.0, + 1408.0, + 953.0, + 1408.0, + 1001.0, + 1403.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 989.0, + 808.0, + 989.0, + 808.0, + 1024.0, + 292.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 685.0, + 449.0, + 685.0, + 449.0, + 723.0, + 294.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 685.0, + 1266.0, + 685.0, + 1266.0, + 723.0, + 560.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 685.0, + 1405.0, + 685.0, + 1405.0, + 723.0, + 1287.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 716.0, + 365.0, + 716.0, + 365.0, + 751.0, + 292.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 716.0, + 678.0, + 716.0, + 678.0, + 751.0, + 386.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 716.0, + 1095.0, + 716.0, + 1095.0, + 751.0, + 699.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 716.0, + 1406.0, + 716.0, + 1406.0, + 751.0, + 1113.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 749.0, + 1332.0, + 749.0, + 1332.0, + 783.0, + 295.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1690.0, + 541.0, + 1690.0, + 541.0, + 1728.0, + 293.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1690.0, + 1403.0, + 1690.0, + 1403.0, + 1728.0, + 670.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1723.0, + 585.0, + 1723.0, + 585.0, + 1757.0, + 293.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1723.0, + 936.0, + 1723.0, + 936.0, + 1757.0, + 757.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1723.0, + 1405.0, + 1723.0, + 1405.0, + 1757.0, + 1007.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1752.0, + 856.0, + 1752.0, + 856.0, + 1789.0, + 293.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1752.0, + 926.0, + 1752.0, + 926.0, + 1789.0, + 877.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1752.0, + 1308.0, + 1752.0, + 1308.0, + 1789.0, + 947.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1435.0, + 352.0, + 1435.0, + 352.0, + 1493.0, + 288.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1435.0, + 1104.0, + 1435.0, + 1104.0, + 1493.0, + 482.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1435.0, + 1409.0, + 1435.0, + 1409.0, + 1493.0, + 1396.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1477.0, + 1404.0, + 1477.0, + 1404.0, + 1511.0, + 295.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1508.0, + 528.0, + 1508.0, + 528.0, + 1542.0, + 295.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1508.0, + 612.0, + 1508.0, + 612.0, + 1542.0, + 600.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1798.0, + 472.0, + 1798.0, + 472.0, + 1835.0, + 297.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1798.0, + 733.0, + 1798.0, + 733.0, + 1835.0, + 492.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1299.0, + 837.0, + 1299.0, + 837.0, + 1340.0, + 296.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1299.0, + 1095.0, + 1299.0, + 1095.0, + 1340.0, + 906.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1299.0, + 1149.0, + 1299.0, + 1149.0, + 1340.0, + 1117.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1550.0, + 1008.0, + 1550.0, + 1008.0, + 1591.0, + 294.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1550.0, + 1187.0, + 1550.0, + 1187.0, + 1591.0, + 1027.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1550.0, + 1238.0, + 1550.0, + 1238.0, + 1591.0, + 1205.0, + 1591.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 743, + 1404, + 743, + 1404, + 1169, + 298, + 1169 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1619, + 1403, + 1619, + 1403, + 1897, + 298, + 1897 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1185, + 1403, + 1185, + 1403, + 1340, + 299, + 1340 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 543, + 1404, + 543, + 1404, + 728, + 298, + 728 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 227, + 1403, + 227, + 1403, + 434, + 298, + 434 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1451, + 1403, + 1451, + 1403, + 1605, + 299, + 1605 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1401, + 1911, + 1401, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 0, + "poly": [ + 301, + 474, + 587, + 474, + 587, + 510, + 301, + 510 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 301, + 1382, + 820, + 1382, + 820, + 1419, + 301, + 1419 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 104, + 298, + 104 + ], + "score": 0.835 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.599 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.454 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 852, + 76, + 852, + 104, + 299, + 104 + ], + "score": 0.176 + }, + { + "category_id": 13, + "poly": [ + 1186, + 1742, + 1350, + 1742, + 1350, + 1776, + 1186, + 1776 + ], + "score": 0.94, + "latex": "\\grave { \\phi ( x ) = { g _ { \\eta } ( x ) } }" + }, + { + "category_id": 13, + "poly": [ + 637, + 230, + 771, + 230, + 771, + 264, + 637, + 264 + ], + "score": 0.93, + "latex": "\\sigma \\in ( 0 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 427, + 230, + 559, + 230, + 559, + 264, + 427, + 264 + ], + "score": 0.92, + "latex": "\\alpha \\in [ 0 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 644, + 1652, + 719, + 1652, + 719, + 1684, + 644, + 1684 + ], + "score": 0.92, + "latex": "\\phi = c" + }, + { + "category_id": 13, + "poly": [ + 1097, + 230, + 1164, + 230, + 1164, + 264, + 1097, + 264 + ], + "score": 0.91, + "latex": "Z \\left( \\alpha \\right)" + }, + { + "category_id": 13, + "poly": [ + 923, + 1683, + 952, + 1683, + 952, + 1714, + 923, + 1714 + ], + "score": 0.88, + "latex": "\\phi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 803, + 366, + 975, + 366, + 975, + 399, + 803, + 399 + ], + "score": 0.87, + "latex": "( a l p h a = - \\infty )" + }, + { + "category_id": 13, + "poly": [ + 1371, + 1713, + 1400, + 1713, + 1400, + 1744, + 1371, + 1744 + ], + "score": 0.86, + "latex": "\\phi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1283, + 332, + 1393, + 332, + 1393, + 365, + 1283, + 365 + ], + "score": 0.85, + "latex": "( \\alpha = - 2 )" + }, + { + "category_id": 13, + "poly": [ + 358, + 1716, + 386, + 1716, + 386, + 1743, + 358, + 1743 + ], + "score": 0.84, + "latex": "x _ { i }" + }, + { + "category_id": 13, + "poly": [ + 705, + 299, + 783, + 299, + 783, + 329, + 705, + 329 + ], + "score": 0.84, + "latex": "( \\alpha = 2 )" + }, + { + "category_id": 13, + "poly": [ + 627, + 332, + 715, + 332, + 715, + 365, + 627, + 365 + ], + "score": 0.83, + "latex": "( \\alpha = 0 )" + }, + { + "category_id": 13, + "poly": [ + 792, + 1779, + 811, + 1779, + 811, + 1805, + 792, + 1805 + ], + "score": 0.78, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 713, + 932, + 731, + 932, + 731, + 959, + 713, + 959 + ], + "score": 0.73, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 573, + 1779, + 591, + 1779, + 591, + 1805, + 573, + 1805 + ], + "score": 0.73, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 1306, + 1684, + 1319, + 1684, + 1319, + 1709, + 1306, + 1709 + ], + "score": 0.62, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1311, + 299, + 1393, + 299, + 1393, + 330, + 1311, + 330 + ], + "score": 0.52, + "latex": "( \\alpha = 1 )" + }, + { + "category_id": 13, + "poly": [ + 1375, + 238, + 1396, + 238, + 1396, + 263, + 1375, + 263 + ], + "score": 0.47, + "latex": "\\rho _ { ; }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1805, + 314, + 1805, + 314, + 1832, + 298, + 1832 + ], + "score": 0.43, + "latex": "\\theta" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 471.0, + 594.0, + 471.0, + 594.0, + 518.0, + 291.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1383.0, + 824.0, + 1383.0, + 824.0, + 1421.0, + 295.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 745.0, + 1404.0, + 745.0, + 1404.0, + 776.0, + 296.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 771.0, + 1407.0, + 771.0, + 1407.0, + 809.0, + 292.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 802.0, + 1405.0, + 802.0, + 1405.0, + 840.0, + 293.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 832.0, + 1405.0, + 832.0, + 1405.0, + 870.0, + 293.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 864.0, + 1406.0, + 864.0, + 1406.0, + 902.0, + 293.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 895.0, + 1405.0, + 895.0, + 1405.0, + 931.0, + 294.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 926.0, + 712.0, + 926.0, + 712.0, + 962.0, + 291.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 926.0, + 1407.0, + 926.0, + 1407.0, + 962.0, + 732.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 957.0, + 1406.0, + 957.0, + 1406.0, + 990.0, + 293.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 988.0, + 1406.0, + 988.0, + 1406.0, + 1023.0, + 294.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1018.0, + 1405.0, + 1018.0, + 1405.0, + 1053.0, + 296.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 1405.0, + 1048.0, + 1405.0, + 1084.0, + 294.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1076.0, + 1404.0, + 1076.0, + 1404.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1109.0, + 1404.0, + 1109.0, + 1404.0, + 1143.0, + 293.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1139.0, + 1264.0, + 1139.0, + 1264.0, + 1171.0, + 297.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1618.0, + 1405.0, + 1618.0, + 1405.0, + 1656.0, + 296.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 643.0, + 1650.0, + 643.0, + 1687.0, + 294.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1650.0, + 1407.0, + 1650.0, + 1407.0, + 1687.0, + 720.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 922.0, + 1681.0, + 922.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1681.0, + 1305.0, + 1681.0, + 1305.0, + 1717.0, + 953.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 1681.0, + 1407.0, + 1681.0, + 1407.0, + 1717.0, + 1320.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1709.0, + 357.0, + 1709.0, + 357.0, + 1749.0, + 293.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1709.0, + 1370.0, + 1709.0, + 1370.0, + 1749.0, + 387.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1709.0, + 1404.0, + 1709.0, + 1404.0, + 1749.0, + 1401.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 1185.0, + 1740.0, + 1185.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1740.0, + 1407.0, + 1740.0, + 1407.0, + 1779.0, + 1351.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 572.0, + 1771.0, + 572.0, + 1812.0, + 293.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1771.0, + 791.0, + 1771.0, + 791.0, + 1812.0, + 592.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1812.0, + 812.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1801.0, + 297.0, + 1801.0, + 297.0, + 1841.0, + 293.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1801.0, + 1405.0, + 1801.0, + 1405.0, + 1841.0, + 315.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1833.0, + 1407.0, + 1833.0, + 1407.0, + 1870.0, + 293.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1863.0, + 1277.0, + 1863.0, + 1277.0, + 1902.0, + 292.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1187.0, + 1404.0, + 1187.0, + 1404.0, + 1220.0, + 295.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1213.0, + 1405.0, + 1213.0, + 1405.0, + 1252.0, + 294.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1243.0, + 1404.0, + 1243.0, + 1404.0, + 1288.0, + 293.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1276.0, + 1405.0, + 1276.0, + 1405.0, + 1313.0, + 295.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1309.0, + 1309.0, + 1309.0, + 1309.0, + 1343.0, + 294.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 541.0, + 1405.0, + 541.0, + 1405.0, + 580.0, + 292.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 576.0, + 1405.0, + 576.0, + 1405.0, + 608.0, + 294.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 606.0, + 1404.0, + 606.0, + 1404.0, + 638.0, + 296.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 633.0, + 1405.0, + 633.0, + 1405.0, + 672.0, + 293.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 1406.0, + 666.0, + 1406.0, + 702.0, + 294.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 698.0, + 477.0, + 698.0, + 477.0, + 731.0, + 296.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 225.0, + 426.0, + 225.0, + 426.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 225.0, + 636.0, + 225.0, + 636.0, + 268.0, + 560.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 225.0, + 1096.0, + 225.0, + 1096.0, + 268.0, + 772.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 225.0, + 1374.0, + 225.0, + 1374.0, + 268.0, + 1165.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 225.0, + 1406.0, + 225.0, + 1406.0, + 268.0, + 1397.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 299.0, + 293.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 297.0, + 704.0, + 297.0, + 704.0, + 334.0, + 294.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 297.0, + 1310.0, + 297.0, + 1310.0, + 334.0, + 784.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 297.0, + 1404.0, + 297.0, + 1404.0, + 334.0, + 1394.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 329.0, + 626.0, + 329.0, + 626.0, + 370.0, + 294.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 329.0, + 1282.0, + 329.0, + 1282.0, + 370.0, + 716.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 329.0, + 1404.0, + 329.0, + 1404.0, + 370.0, + 1394.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 364.0, + 802.0, + 364.0, + 802.0, + 402.0, + 294.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 364.0, + 1405.0, + 364.0, + 1405.0, + 402.0, + 976.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 398.0, + 887.0, + 398.0, + 887.0, + 438.0, + 293.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1451.0, + 1403.0, + 1451.0, + 1403.0, + 1488.0, + 297.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1480.0, + 1403.0, + 1480.0, + 1403.0, + 1519.0, + 295.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1512.0, + 1405.0, + 1512.0, + 1405.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1543.0, + 1408.0, + 1543.0, + 1408.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1572.0, + 1183.0, + 1572.0, + 1183.0, + 1611.0, + 293.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1405.0, + 1940.0, + 1405.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2003.0, + 1406.0, + 2003.0, + 1406.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1469, + 1406, + 1469, + 1406, + 1781, + 296, + 1781 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1222, + 1404, + 1222, + 1404, + 1456, + 297, + 1456 + ], + "score": 0.979 + }, + { + "category_id": 4, + "poly": [ + 296, + 1034, + 1404, + 1034, + 1404, + 1166, + 296, + 1166 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 298, + 653, + 1403, + 653, + 1403, + 1026, + 298, + 1026 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 858, + 478, + 1394, + 478, + 1394, + 608, + 858, + 608 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 305, + 453, + 839, + 453, + 839, + 618, + 305, + 618 + ], + "score": 0.957 + }, + { + "category_id": 3, + "poly": [ + 312, + 233, + 832, + 233, + 832, + 425, + 312, + 425 + ], + "score": 0.956 + }, + { + "category_id": 3, + "poly": [ + 860, + 236, + 1384, + 236, + 1384, + 450, + 860, + 450 + ], + "score": 0.947 + }, + { + "category_id": 0, + "poly": [ + 300, + 1825, + 563, + 1825, + 563, + 1859, + 300, + 1859 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 852, + 76, + 852, + 104, + 299, + 104 + ], + "score": 0.864 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2111, + 841, + 2111 + ], + "score": 0.751 + }, + { + "category_id": 13, + "poly": [ + 786, + 1532, + 889, + 1532, + 889, + 1564, + 786, + 1564 + ], + "score": 0.92, + "latex": "| \\phi | = C" + }, + { + "category_id": 13, + "poly": [ + 630, + 1623, + 843, + 1623, + 843, + 1652, + 630, + 1652 + ], + "score": 0.92, + "latex": "N \\times W \\times H \\times C" + }, + { + "category_id": 13, + "poly": [ + 298, + 1623, + 449, + 1623, + 449, + 1652, + 298, + 1652 + ], + "score": 0.92, + "latex": "W \\times H \\times C" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1531, + 1367, + 1531, + 1367, + 1565, + 1133, + 1565 + ], + "score": 0.91, + "latex": "| \\phi | = \\bar { W \\times H \\times C }" + }, + { + "category_id": 13, + "poly": [ + 435, + 1532, + 531, + 1532, + 531, + 1564, + 435, + 1564 + ], + "score": 0.91, + "latex": "| \\phi | = 1" + }, + { + "category_id": 13, + "poly": [ + 1333, + 1592, + 1404, + 1592, + 1404, + 1626, + 1333, + 1626 + ], + "score": 0.9, + "latex": "| \\phi | \\overset { \\cdot } { = }" + }, + { + "category_id": 13, + "poly": [ + 721, + 1470, + 755, + 1470, + 755, + 1503, + 721, + 1503 + ], + "score": 0.89, + "latex": "| \\phi |" + }, + { + "category_id": 13, + "poly": [ + 527, + 1503, + 544, + 1503, + 544, + 1534, + 527, + 1534 + ], + "score": 0.81, + "latex": "\\hat { y }" + }, + { + "category_id": 13, + "poly": [ + 1179, + 1624, + 1208, + 1624, + 1208, + 1650, + 1179, + 1650 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1144, + 1230, + 1162, + 1230, + 1162, + 1255, + 1144, + 1255 + ], + "score": 0.74, + "latex": "\\rho" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1030.0, + 1407.0, + 1030.0, + 1407.0, + 1070.0, + 293.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1060.0, + 1406.0, + 1060.0, + 1406.0, + 1099.0, + 293.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1094.0, + 1407.0, + 1094.0, + 1407.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 573.0, + 1130.0, + 573.0, + 1169.0, + 294.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 657.0, + 647.0, + 657.0, + 647.0, + 684.0, + 333.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 655.0, + 1010.0, + 655.0, + 1010.0, + 686.0, + 723.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 653.0, + 1341.0, + 653.0, + 1341.0, + 687.0, + 1117.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 702.0, + 670.0, + 702.0, + 670.0, + 736.0, + 612.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 690.0, + 1043.0, + 690.0, + 1043.0, + 721.0, + 989.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 693.0, + 1404.0, + 693.0, + 1404.0, + 863.0, + 1359.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 734.0, + 668.0, + 734.0, + 668.0, + 768.0, + 608.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 727.0, + 1043.0, + 727.0, + 1043.0, + 758.0, + 987.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 764.0, + 668.0, + 764.0, + 668.0, + 799.0, + 608.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 763.0, + 1044.0, + 763.0, + 1044.0, + 799.0, + 985.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 793.0, + 670.0, + 793.0, + 670.0, + 833.0, + 609.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 799.0, + 1044.0, + 799.0, + 1044.0, + 835.0, + 985.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 835.0, + 321.0, + 835.0, + 321.0, + 860.0, + 301.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 828.0, + 671.0, + 828.0, + 671.0, + 863.0, + 611.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 833.0, + 699.0, + 833.0, + 699.0, + 862.0, + 672.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 836.0, + 1044.0, + 836.0, + 1044.0, + 872.0, + 985.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 828.0, + 1078.0, + 828.0, + 1078.0, + 868.0, + 1048.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 859.0, + 670.0, + 859.0, + 670.0, + 894.0, + 612.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 858.0, + 1399.0, + 858.0, + 1399.0, + 897.0, + 1364.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 891.0, + 668.0, + 891.0, + 668.0, + 926.0, + 611.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 874.0, + 1044.0, + 874.0, + 1044.0, + 910.0, + 985.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 888.0, + 1399.0, + 888.0, + 1399.0, + 927.0, + 1364.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 924.0, + 667.0, + 924.0, + 667.0, + 955.0, + 612.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 911.0, + 1044.0, + 911.0, + 1044.0, + 947.0, + 985.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 923.0, + 1398.0, + 923.0, + 1398.0, + 954.0, + 1365.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 955.0, + 667.0, + 955.0, + 667.0, + 986.0, + 613.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 957.0, + 722.0, + 957.0, + 722.0, + 965.0, + 712.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 948.0, + 1044.0, + 948.0, + 1044.0, + 982.0, + 986.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 958.0, + 1386.0, + 958.0, + 1386.0, + 984.0, + 1366.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 987.0, + 667.0, + 987.0, + 667.0, + 1018.0, + 613.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 987.0, + 1043.0, + 987.0, + 1043.0, + 1017.0, + 989.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 991.0, + 1384.0, + 991.0, + 1384.0, + 1011.0, + 1369.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1004.0, + 470.0, + 1004.0, + 470.0, + 1033.0, + 438.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1005.0, + 843.0, + 1005.0, + 843.0, + 1033.0, + 810.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1004.0, + 1219.0, + 1004.0, + 1219.0, + 1033.0, + 1188.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 476.0, + 1397.0, + 476.0, + 1397.0, + 512.0, + 859.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 505.0, + 1396.0, + 505.0, + 1396.0, + 542.0, + 858.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 536.0, + 1398.0, + 536.0, + 1398.0, + 576.0, + 858.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 572.0, + 1395.0, + 572.0, + 1395.0, + 609.0, + 859.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 453.0, + 841.0, + 453.0, + 841.0, + 488.0, + 303.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 484.0, + 842.0, + 484.0, + 842.0, + 520.0, + 304.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 512.0, + 843.0, + 512.0, + 843.0, + 551.0, + 303.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 546.0, + 843.0, + 546.0, + 843.0, + 585.0, + 301.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 583.0, + 546.0, + 583.0, + 546.0, + 619.0, + 304.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 230.0, + 415.0, + 230.0, + 415.0, + 269.0, + 306.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 251.0, + 662.0, + 251.0, + 662.0, + 289.0, + 552.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 345.0, + 772.0, + 345.0, + 772.0, + 389.0, + 619.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 390.0, + 618.0, + 390.0, + 618.0, + 411.0, + 595.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 389.0, + 836.0, + 389.0, + 836.0, + 426.0, + 748.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 271.0, + 958.0, + 271.0, + 958.0, + 299.0, + 930.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 309.0, + 1152.0, + 309.0, + 1152.0, + 361.0, + 1003.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 358.0, + 973.0, + 358.0, + 973.0, + 400.0, + 931.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 362.0, + 1383.0, + 362.0, + 1383.0, + 393.0, + 1147.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1820.0, + 568.0, + 1820.0, + 568.0, + 1867.0, + 292.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 720.0, + 1467.0, + 720.0, + 1504.0, + 294.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1467.0, + 1407.0, + 1467.0, + 1407.0, + 1504.0, + 756.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1499.0, + 526.0, + 1499.0, + 526.0, + 1535.0, + 295.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1499.0, + 1404.0, + 1499.0, + 1404.0, + 1535.0, + 545.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1528.0, + 434.0, + 1528.0, + 434.0, + 1568.0, + 293.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1528.0, + 785.0, + 1528.0, + 785.0, + 1568.0, + 532.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1528.0, + 1132.0, + 1528.0, + 1132.0, + 1568.0, + 890.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1528.0, + 1408.0, + 1528.0, + 1408.0, + 1568.0, + 1368.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1557.0, + 1406.0, + 1557.0, + 1406.0, + 1599.0, + 291.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1332.0, + 1590.0, + 1332.0, + 1626.0, + 294.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1620.0, + 629.0, + 1620.0, + 629.0, + 1655.0, + 450.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1620.0, + 1178.0, + 1620.0, + 1178.0, + 1655.0, + 844.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1620.0, + 1404.0, + 1620.0, + 1404.0, + 1655.0, + 1209.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1653.0, + 1404.0, + 1653.0, + 1404.0, + 1686.0, + 296.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1407.0, + 1682.0, + 1407.0, + 1718.0, + 294.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1710.0, + 1406.0, + 1710.0, + 1406.0, + 1750.0, + 294.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1746.0, + 696.0, + 1746.0, + 696.0, + 1782.0, + 296.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1222.0, + 1143.0, + 1222.0, + 1143.0, + 1258.0, + 295.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1222.0, + 1406.0, + 1222.0, + 1406.0, + 1258.0, + 1163.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1253.0, + 1406.0, + 1253.0, + 1406.0, + 1289.0, + 294.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1284.0, + 1406.0, + 1284.0, + 1406.0, + 1325.0, + 296.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1319.0, + 1407.0, + 1319.0, + 1407.0, + 1357.0, + 292.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1352.0, + 1406.0, + 1352.0, + 1406.0, + 1392.0, + 294.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1387.0, + 1404.0, + 1387.0, + 1404.0, + 1424.0, + 295.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1419.0, + 591.0, + 1419.0, + 591.0, + 1464.0, + 291.0, + 1464.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 999, + 1404, + 999, + 1404, + 1365, + 297, + 1365 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 677, + 1403, + 677, + 1403, + 985, + 297, + 985 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1381, + 1403, + 1381, + 1403, + 1722, + 298, + 1722 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1737, + 1403, + 1737, + 1403, + 1986, + 297, + 1986 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 563, + 340, + 1132, + 340, + 1132, + 568, + 563, + 568 + ], + "score": 0.974, + "html": "
Param.DimMSETimeMem
Global1x1x1225.81.04×<1KB
Data1x1x1244.22.70×0.6GB
Pred.1x1x1228.51.04×<1MB
GlobalHxWxC231.11.08×<1MB
DataHxWxC252.69.42×4.4GB
Pred.HxWxC222.31.08×<1MB
" + }, + { + "category_id": 6, + "poly": [ + 296, + 224, + 1407, + 224, + 1407, + 317, + 296, + 317 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.754 + }, + { + "category_id": 0, + "poly": [ + 300, + 620, + 845, + 620, + 845, + 652, + 300, + 652 + ], + "score": 0.554 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 854, + 75, + 854, + 105, + 297, + 105 + ], + "score": 0.54 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.423 + }, + { + "category_id": 13, + "poly": [ + 979, + 1658, + 1112, + 1658, + 1112, + 1694, + 979, + 1694 + ], + "score": 0.95, + "latex": "\\bar { \\nabla } _ { \\phi } / \\| \\nabla _ { \\phi } \\| _ { 2 } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 625, + 1564, + 792, + 1564, + 792, + 1598, + 625, + 1598 + ], + "score": 0.93, + "latex": "c \\in [ 1 0 ^ { - } 8 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 1290, + 1533, + 1402, + 1533, + 1402, + 1567, + 1290, + 1567 + ], + "score": 0.92, + "latex": "\\alpha \\in [ 0 , 3 ]" + }, + { + "category_id": 13, + "poly": [ + 712, + 1658, + 771, + 1658, + 771, + 1692, + 712, + 1692 + ], + "score": 0.9, + "latex": "\\| \\phi \\| _ { 2 } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 568, + 1384, + 638, + 1384, + 638, + 1410, + 568, + 1410 + ], + "score": 0.85, + "latex": "1 \\times 1" + }, + { + "category_id": 13, + "poly": [ + 297, + 1694, + 432, + 1694, + 432, + 1723, + 297, + 1723 + ], + "score": 0.85, + "latex": "\\alpha _ { \\phi } = \\alpha \\cdot m" + }, + { + "category_id": 13, + "poly": [ + 621, + 1696, + 642, + 1696, + 642, + 1716, + 621, + 1716 + ], + "score": 0.74, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1256, + 260, + 1274, + 260, + 1274, + 287, + 1256, + 287 + ], + "score": 0.72, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 482, + 1098, + 501, + 1098, + 501, + 1124, + 482, + 1124 + ], + "score": 0.71, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 808, + 1695, + 836, + 1695, + 836, + 1716, + 808, + 1716 + ], + "score": 0.69, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 713, + 376, + 783, + 376, + 783, + 403, + 713, + 403 + ], + "score": 0.33, + "latex": "\\overline { { 1 { \\times } 1 { \\times } 1 } }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 221.0, + 1405.0, + 221.0, + 1405.0, + 260.0, + 293.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 255.0, + 1255.0, + 255.0, + 1255.0, + 289.0, + 295.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 255.0, + 1405.0, + 255.0, + 1405.0, + 289.0, + 1275.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 283.0, + 1031.0, + 283.0, + 1031.0, + 321.0, + 293.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 616.0, + 851.0, + 616.0, + 851.0, + 656.0, + 294.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1000.0, + 1404.0, + 1000.0, + 1404.0, + 1030.0, + 297.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1063.0, + 296.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1057.0, + 1405.0, + 1057.0, + 1405.0, + 1096.0, + 292.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1091.0, + 481.0, + 1091.0, + 481.0, + 1125.0, + 295.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1091.0, + 1402.0, + 1091.0, + 1402.0, + 1125.0, + 502.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1119.0, + 1405.0, + 1119.0, + 1405.0, + 1155.0, + 295.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1152.0, + 1406.0, + 1152.0, + 1406.0, + 1186.0, + 295.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1181.0, + 1406.0, + 1181.0, + 1406.0, + 1215.0, + 295.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1211.0, + 1404.0, + 1211.0, + 1404.0, + 1245.0, + 296.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1243.0, + 1404.0, + 1243.0, + 1404.0, + 1277.0, + 295.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1268.0, + 1406.0, + 1268.0, + 1406.0, + 1311.0, + 292.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1305.0, + 1405.0, + 1305.0, + 1405.0, + 1335.0, + 297.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1334.0, + 1335.0, + 1334.0, + 1335.0, + 1368.0, + 297.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 674.0, + 1401.0, + 674.0, + 1401.0, + 710.0, + 296.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 707.0, + 1405.0, + 707.0, + 1405.0, + 743.0, + 293.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 738.0, + 1403.0, + 738.0, + 1403.0, + 774.0, + 293.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 766.0, + 1407.0, + 766.0, + 1407.0, + 806.0, + 294.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 795.0, + 1405.0, + 795.0, + 1405.0, + 837.0, + 292.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 830.0, + 1407.0, + 830.0, + 1407.0, + 865.0, + 293.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 856.0, + 1407.0, + 856.0, + 1407.0, + 897.0, + 291.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 1405.0, + 887.0, + 1405.0, + 927.0, + 295.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 918.0, + 1407.0, + 918.0, + 1407.0, + 959.0, + 292.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 948.0, + 507.0, + 948.0, + 507.0, + 989.0, + 293.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1385.0, + 567.0, + 1385.0, + 567.0, + 1412.0, + 298.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1385.0, + 1403.0, + 1385.0, + 1403.0, + 1412.0, + 639.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1411.0, + 1405.0, + 1411.0, + 1405.0, + 1445.0, + 294.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1441.0, + 1402.0, + 1441.0, + 1402.0, + 1475.0, + 296.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1474.0, + 1404.0, + 1474.0, + 1404.0, + 1505.0, + 298.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1501.0, + 1405.0, + 1501.0, + 1405.0, + 1536.0, + 296.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1533.0, + 1289.0, + 1533.0, + 1289.0, + 1567.0, + 297.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1565.0, + 624.0, + 1565.0, + 624.0, + 1597.0, + 298.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1565.0, + 1400.0, + 1565.0, + 1400.0, + 1597.0, + 793.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1597.0, + 1403.0, + 1597.0, + 1403.0, + 1630.0, + 298.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1629.0, + 1402.0, + 1629.0, + 1402.0, + 1660.0, + 298.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1657.0, + 711.0, + 1657.0, + 711.0, + 1694.0, + 297.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1657.0, + 978.0, + 1657.0, + 978.0, + 1694.0, + 772.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1657.0, + 1403.0, + 1657.0, + 1403.0, + 1694.0, + 1113.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1689.0, + 620.0, + 1689.0, + 620.0, + 1723.0, + 433.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1689.0, + 807.0, + 1689.0, + 807.0, + 1723.0, + 643.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1689.0, + 1068.0, + 1689.0, + 1068.0, + 1723.0, + 837.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1740.0, + 1403.0, + 1740.0, + 1403.0, + 1771.0, + 299.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1769.0, + 1403.0, + 1769.0, + 1403.0, + 1800.0, + 297.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1798.0, + 1404.0, + 1798.0, + 1404.0, + 1832.0, + 296.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1831.0, + 1403.0, + 1831.0, + 1403.0, + 1861.0, + 297.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1859.0, + 1404.0, + 1859.0, + 1404.0, + 1894.0, + 293.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1891.0, + 1405.0, + 1891.0, + 1405.0, + 1925.0, + 296.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1920.0, + 1405.0, + 1920.0, + 1405.0, + 1956.0, + 295.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1951.0, + 1405.0, + 1951.0, + 1405.0, + 1986.0, + 293.0, + 1986.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1294, + 1405, + 1294, + 1405, + 1886, + 296, + 1886 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 862, + 642, + 1393, + 642, + 1393, + 1072, + 862, + 1072 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 855, + 1103, + 1404, + 1103, + 1404, + 1228, + 855, + 1228 + ], + "score": 0.96 + }, + { + "category_id": 3, + "poly": [ + 322, + 228, + 1383, + 228, + 1383, + 506, + 322, + 506 + ], + "score": 0.943 + }, + { + "category_id": 4, + "poly": [ + 297, + 540, + 1404, + 540, + 1404, + 606, + 297, + 606 + ], + "score": 0.937 + }, + { + "category_id": 5, + "poly": [ + 392, + 746, + 744, + 746, + 744, + 1241, + 392, + 1241 + ], + "score": 0.937, + "html": "
MethodMedian AUC
LOF.669
FB.702
ABOD.727
AE.737
VAE.792
COPOD.799
PCA.808
OCSVM.814
MCD.820
KNN.822
HBOS.822
IF.823
CBLOF.836
AE+S (Ours).846
PCA+S (Ours).868
" + }, + { + "category_id": 6, + "poly": [ + 295, + 635, + 842, + 635, + 842, + 727, + 295, + 727 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 301, + 1900, + 1400, + 1900, + 1400, + 1995, + 301, + 1995 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 854, + 76, + 854, + 104, + 298, + 104 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.789 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 853, + 76, + 853, + 104, + 299, + 104 + ], + "score": 0.113 + }, + { + "category_id": 13, + "poly": [ + 1006, + 1165, + 1078, + 1165, + 1078, + 1194, + 1006, + 1194 + ], + "score": 0.84, + "latex": "_ \\mathrm { A E + S }" + }, + { + "category_id": 13, + "poly": [ + 857, + 1165, + 948, + 1165, + 948, + 1195, + 857, + 1195 + ], + "score": 0.81, + "latex": "\\mathrm { P C A } { + } \\mathrm { S }" + }, + { + "category_id": 13, + "poly": [ + 1135, + 1364, + 1235, + 1364, + 1235, + 1394, + 1135, + 1394 + ], + "score": 0.78, + "latex": "( \\mathrm { P C A } { + } \\mathrm { S } )" + }, + { + "category_id": 13, + "poly": [ + 924, + 1363, + 1010, + 1363, + 1010, + 1394, + 924, + 1394 + ], + "score": 0.66, + "latex": "\\left( \\mathrm { A E } { + } \\mathrm { S } \\right)" + }, + { + "category_id": 13, + "poly": [ + 413, + 1177, + 486, + 1177, + 486, + 1205, + 413, + 1205 + ], + "score": 0.35, + "latex": "_ \\mathrm { A E + S }" + }, + { + "category_id": 13, + "poly": [ + 400, + 1207, + 494, + 1207, + 494, + 1238, + 400, + 1238 + ], + "score": 0.35, + "latex": "\\mathbf { P C A + S }" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 649.0, + 925.0, + 649.0, + 925.0, + 680.0, + 886.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 670.0, + 956.0, + 670.0, + 956.0, + 703.0, + 934.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 665.0, + 1047.0, + 665.0, + 1047.0, + 687.0, + 1028.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 665.0, + 1076.0, + 665.0, + 1076.0, + 716.0, + 1060.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 664.0, + 1105.0, + 664.0, + 1105.0, + 682.0, + 1088.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 689.0, + 1138.0, + 689.0, + 1138.0, + 737.0, + 1123.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 696.0, + 1167.0, + 696.0, + 1167.0, + 716.0, + 1151.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 667.0, + 1200.0, + 667.0, + 1200.0, + 720.0, + 1181.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 687.0, + 1228.0, + 687.0, + 1228.0, + 703.0, + 1217.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 664.0, + 1263.0, + 664.0, + 1263.0, + 727.0, + 1243.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 662.0, + 1295.0, + 662.0, + 1295.0, + 751.0, + 1269.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 688.0, + 1386.0, + 688.0, + 1386.0, + 723.0, + 1364.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 712.0, + 925.0, + 712.0, + 925.0, + 741.0, + 886.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 711.0, + 949.0, + 711.0, + 949.0, + 721.0, + 941.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 716.0, + 980.0, + 716.0, + 980.0, + 724.0, + 972.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 712.0, + 1231.0, + 712.0, + 1231.0, + 767.0, + 1214.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 728.0, + 954.0, + 728.0, + 954.0, + 797.0, + 934.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 730.0, + 1041.0, + 730.0, + 1041.0, + 735.0, + 1036.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 726.0, + 1262.0, + 726.0, + 1262.0, + 772.0, + 1241.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 737.0, + 982.0, + 737.0, + 982.0, + 758.0, + 966.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 760.0, + 1073.0, + 760.0, + 1073.0, + 769.0, + 1064.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 753.0, + 1169.0, + 753.0, + 1169.0, + 776.0, + 1154.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 769.0, + 980.0, + 769.0, + 980.0, + 781.0, + 972.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 771.0, + 1133.0, + 771.0, + 1133.0, + 781.0, + 1126.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 773.0, + 926.0, + 773.0, + 926.0, + 807.0, + 885.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 778.0, + 1044.0, + 778.0, + 1044.0, + 791.0, + 1032.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 783.0, + 892.0, + 783.0, + 892.0, + 838.0, + 860.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 835.0, + 926.0, + 835.0, + 926.0, + 865.0, + 886.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 840.0, + 1319.0, + 840.0, + 1319.0, + 847.0, + 1311.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 873.0, + 1261.0, + 873.0, + 1261.0, + 883.0, + 1246.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 896.0, + 926.0, + 896.0, + 926.0, + 929.0, + 885.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 897.0, + 1319.0, + 897.0, + 1319.0, + 908.0, + 1311.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 959.0, + 925.0, + 959.0, + 925.0, + 988.0, + 886.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 983.0, + 959.0, + 983.0, + 959.0, + 1026.0, + 931.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 983.0, + 988.0, + 983.0, + 988.0, + 1016.0, + 963.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 981.0, + 1020.0, + 981.0, + 1020.0, + 1042.0, + 990.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 982.0, + 1049.0, + 982.0, + 1049.0, + 1016.0, + 1023.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 981.0, + 1081.0, + 981.0, + 1081.0, + 1027.0, + 1053.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 982.0, + 1114.0, + 982.0, + 1114.0, + 1053.0, + 1084.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 981.0, + 1174.0, + 981.0, + 1174.0, + 1056.0, + 1115.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 982.0, + 1205.0, + 982.0, + 1205.0, + 1035.0, + 1178.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 983.0, + 1236.0, + 983.0, + 1236.0, + 1031.0, + 1208.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 982.0, + 1268.0, + 982.0, + 1268.0, + 1042.0, + 1238.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 983.0, + 1296.0, + 983.0, + 1296.0, + 1016.0, + 1270.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 982.0, + 1327.0, + 982.0, + 1327.0, + 1047.0, + 1300.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 980.0, + 1359.0, + 980.0, + 1359.0, + 1040.0, + 1329.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 981.0, + 1391.0, + 981.0, + 1391.0, + 1052.0, + 1361.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1046.0, + 1195.0, + 1046.0, + 1195.0, + 1075.0, + 1128.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1102.0, + 1404.0, + 1102.0, + 1404.0, + 1136.0, + 855.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1129.0, + 1405.0, + 1129.0, + 1405.0, + 1172.0, + 853.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1161.0, + 1005.0, + 1161.0, + 1005.0, + 1199.0, + 949.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1199.0, + 1079.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1192.0, + 1160.0, + 1192.0, + 1160.0, + 1231.0, + 856.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 230.0, + 1391.0, + 230.0, + 1391.0, + 380.0, + 316.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 380.0, + 1371.0, + 380.0, + 1371.0, + 410.0, + 324.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 395.0, + 1080.0, + 395.0, + 1080.0, + 515.0, + 527.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 434.0, + 1281.0, + 434.0, + 1281.0, + 507.0, + 1181.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 451.0, + 1323.0, + 451.0, + 1323.0, + 483.0, + 1300.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 457.0, + 1369.0, + 457.0, + 1369.0, + 496.0, + 1329.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.75, + 423.5, + 540.75, + 423.5, + 540.75, + 484.0, + 442.75, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 541.0, + 1404.0, + 541.0, + 1404.0, + 577.0, + 296.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 572.0, + 1346.0, + 572.0, + 1346.0, + 608.0, + 294.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 632.0, + 844.0, + 632.0, + 844.0, + 670.0, + 296.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 665.0, + 843.0, + 665.0, + 843.0, + 698.0, + 297.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 696.0, + 447.0, + 696.0, + 447.0, + 728.0, + 296.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1296.0, + 1403.0, + 1296.0, + 1403.0, + 1332.0, + 294.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1333.0, + 1405.0, + 1333.0, + 1405.0, + 1365.0, + 296.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1361.0, + 923.0, + 1361.0, + 923.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1361.0, + 1134.0, + 1361.0, + 1134.0, + 1395.0, + 1011.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1361.0, + 1405.0, + 1361.0, + 1405.0, + 1395.0, + 1236.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1392.0, + 1405.0, + 1392.0, + 1405.0, + 1427.0, + 298.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1422.0, + 1406.0, + 1422.0, + 1406.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1487.0, + 295.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1478.0, + 1405.0, + 1478.0, + 1405.0, + 1519.0, + 295.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1514.0, + 1405.0, + 1514.0, + 1405.0, + 1547.0, + 295.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1545.0, + 1403.0, + 1545.0, + 1403.0, + 1577.0, + 296.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1573.0, + 1403.0, + 1573.0, + 1403.0, + 1609.0, + 294.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1607.0, + 1405.0, + 1607.0, + 1405.0, + 1643.0, + 293.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1636.0, + 1402.0, + 1636.0, + 1402.0, + 1678.0, + 294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 1405.0, + 1668.0, + 1405.0, + 1706.0, + 294.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 1402.0, + 1698.0, + 1402.0, + 1735.0, + 295.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1730.0, + 1402.0, + 1730.0, + 1402.0, + 1766.0, + 296.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1402.0, + 1760.0, + 1402.0, + 1795.0, + 294.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1792.0, + 1401.0, + 1792.0, + 1401.0, + 1824.0, + 298.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1822.0, + 1402.0, + 1822.0, + 1402.0, + 1857.0, + 296.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1853.0, + 1124.0, + 1853.0, + 1124.0, + 1885.0, + 298.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1902.0, + 1404.0, + 1902.0, + 1404.0, + 1933.0, + 298.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1927.0, + 1404.0, + 1927.0, + 1404.0, + 1965.0, + 297.0, + 1965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1956.0, + 1406.0, + 1956.0, + 1406.0, + 1999.0, + 296.0, + 1999.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1248, + 1406, + 1248, + 1406, + 1493, + 296, + 1493 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 900, + 1403, + 900, + 1403, + 1146, + 298, + 1146 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1594, + 1404, + 1594, + 1404, + 1720, + 298, + 1720 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1879, + 1403, + 1879, + 1403, + 2036, + 298, + 2036 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 305, + 227, + 1369, + 227, + 1369, + 548, + 305, + 548 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 613, + 1493, + 1088, + 1493, + 1088, + 1583, + 613, + 1583 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 524, + 1759, + 1173, + 1759, + 1173, + 1854, + 524, + 1854 + ], + "score": 0.956 + }, + { + "category_id": 4, + "poly": [ + 295, + 574, + 1405, + 574, + 1405, + 760, + 295, + 760 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 300, + 818, + 1397, + 818, + 1397, + 884, + 300, + 884 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 297, + 1190, + 1034, + 1190, + 1034, + 1224, + 297, + 1224 + ], + "score": 0.906 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1520, + 1400, + 1520, + 1400, + 1550, + 1366, + 1550 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1790, + 1400, + 1790, + 1400, + 1820, + 1366, + 1820 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.691 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.221 + }, + { + "category_id": 14, + "poly": [ + 610, + 1487, + 1091, + 1487, + 1091, + 1586, + 610, + 1586 + ], + "score": 0.94, + "latex": "\\operatorname* { m i n } _ { \\theta } \\sum _ { i } ^ { N } ( \\hat { y _ { i } } : = f _ { \\theta } ( x _ { i } ) - y _ { i } ) ^ { 2 } + \\lambda \\sum _ { j } ^ { P } | \\theta _ { j } | ." + }, + { + "category_id": 14, + "poly": [ + 525, + 1757, + 1177, + 1757, + 1177, + 1856, + 525, + 1856 + ], + "score": 0.94, + "latex": "\\operatorname* { m i n } _ { \\theta , \\sigma , b } \\sum _ { i } ^ { N } \\left( \\frac { 1 } { 2 \\sigma ^ { 2 } } ( \\hat { y } _ { i } - y _ { i } ) ^ { 2 } + \\log \\sigma \\right) + \\sum _ { j } ^ { P } \\left( \\frac { | \\theta _ { j } | } { b } + \\log b \\right)" + }, + { + "category_id": 13, + "poly": [ + 624, + 1433, + 748, + 1433, + 748, + 1466, + 624, + 1466 + ], + "score": 0.92, + "latex": "\\lambda \\in [ 0 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 760, + 1343, + 819, + 1343, + 819, + 1375, + 760, + 1375 + ], + "score": 0.91, + "latex": "\\phi _ { \\mathrm { p r i o r } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1626, + 367, + 1626, + 367, + 1654, + 298, + 1654 + ], + "score": 0.88, + "latex": "\\lambda = 0" + }, + { + "category_id": 13, + "poly": [ + 1337, + 1912, + 1393, + 1912, + 1393, + 1944, + 1337, + 1944 + ], + "score": 0.87, + "latex": "\\log b" + }, + { + "category_id": 13, + "poly": [ + 601, + 902, + 674, + 902, + 674, + 931, + 601, + 931 + ], + "score": 0.86, + "latex": "{ \\mathrm { A E } } { + } S _ { \\cdot }" + }, + { + "category_id": 13, + "poly": [ + 599, + 824, + 654, + 824, + 654, + 851, + 599, + 851 + ], + "score": 0.86, + "latex": "20 \\%" + }, + { + "category_id": 13, + "poly": [ + 892, + 608, + 928, + 608, + 928, + 635, + 892, + 635 + ], + "score": 0.86, + "latex": "w ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 956, + 1912, + 975, + 1912, + 975, + 1948, + 956, + 1948 + ], + "score": 0.85, + "latex": "\\frac { 1 } { b }" + }, + { + "category_id": 13, + "poly": [ + 679, + 1657, + 699, + 1657, + 699, + 1683, + 679, + 1683 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 664, + 1597, + 684, + 1597, + 684, + 1623, + 664, + 1623 + ], + "score": 0.8, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 459, + 902, + 549, + 902, + 549, + 931, + 459, + 931 + ], + "score": 0.8, + "latex": "\\mathrm { P C A } { + } \\mathrm { S }" + }, + { + "category_id": 13, + "poly": [ + 1332, + 1377, + 1352, + 1377, + 1352, + 1399, + 1332, + 1399 + ], + "score": 0.75, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 591, + 576, + 628, + 576, + 628, + 604, + 591, + 604 + ], + "score": 0.72, + "latex": "L 2" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1312, + 1112, + 1312, + 1112, + 1338, + 1095, + 1338 + ], + "score": 0.7, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 298, + 1913, + 312, + 1913, + 312, + 1940, + 298, + 1940 + ], + "score": 0.67, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 298, + 1405, + 312, + 1405, + 312, + 1430, + 298, + 1430 + ], + "score": 0.65, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 744, + 576, + 781, + 576, + 781, + 604, + 744, + 604 + ], + "score": 0.64, + "latex": "L 1" + }, + { + "category_id": 13, + "poly": [ + 775, + 698, + 812, + 698, + 812, + 726, + 775, + 726 + ], + "score": 0.63, + "latex": "L 1" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 236.0, + 979.0, + 236.0, + 979.0, + 252.0, + 944.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 248.0, + 644.0, + 248.0, + 644.0, + 272.0, + 580.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 262.0, + 725.0, + 262.0, + 725.0, + 280.0, + 696.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 245.0, + 987.0, + 245.0, + 987.0, + 274.0, + 926.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 254.0, + 1099.0, + 254.0, + 1099.0, + 271.0, + 1071.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 273.0, + 702.0, + 273.0, + 702.0, + 327.0, + 672.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 289.0, + 725.0, + 289.0, + 725.0, + 307.0, + 693.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 298.0, + 1098.0, + 298.0, + 1098.0, + 316.0, + 1067.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 309.0, + 357.0, + 309.0, + 357.0, + 327.0, + 331.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 317.0, + 727.0, + 317.0, + 727.0, + 340.0, + 675.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 339.0, + 332.0, + 339.0, + 332.0, + 414.0, + 301.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 344.0, + 727.0, + 344.0, + 727.0, + 366.0, + 695.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 364.0, + 631.0, + 364.0, + 631.0, + 383.0, + 559.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 353.0, + 1076.0, + 353.0, + 1076.0, + 379.0, + 1053.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 375.0, + 725.0, + 375.0, + 725.0, + 392.0, + 693.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 374.0, + 1072.0, + 374.0, + 1072.0, + 389.0, + 1058.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 383.0, + 775.0, + 383.0, + 775.0, + 400.0, + 744.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 383.0, + 820.0, + 383.0, + 820.0, + 402.0, + 798.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 383.0, + 869.0, + 383.0, + 869.0, + 402.0, + 847.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 383.0, + 916.0, + 383.0, + 916.0, + 402.0, + 894.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 383.0, + 966.0, + 383.0, + 966.0, + 401.0, + 942.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 396.0, + 867.0, + 396.0, + 867.0, + 415.0, + 848.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 419.0, + 933.0, + 419.0, + 933.0, + 438.0, + 782.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 444.0, + 724.0, + 444.0, + 724.0, + 462.0, + 701.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 438.0, + 990.0, + 438.0, + 990.0, + 466.0, + 923.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 432.0, + 1100.0, + 432.0, + 1100.0, + 453.0, + 1070.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 462.0, + 1360.0, + 462.0, + 1360.0, + 481.0, + 1305.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 478.0, + 1099.0, + 478.0, + 1099.0, + 496.0, + 1066.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 474.0, + 1349.0, + 474.0, + 1349.0, + 501.0, + 1306.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 492.0, + 353.0, + 492.0, + 353.0, + 507.0, + 325.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 500.0, + 728.0, + 500.0, + 728.0, + 518.0, + 701.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 495.0, + 1349.0, + 495.0, + 1349.0, + 512.0, + 1306.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 519.0, + 410.0, + 519.0, + 410.0, + 535.0, + 378.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 517.0, + 513.0, + 517.0, + 513.0, + 539.0, + 487.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 519.0, + 564.0, + 519.0, + 564.0, + 535.0, + 538.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 518.0, + 618.0, + 518.0, + 618.0, + 537.0, + 593.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 512.0, + 776.0, + 512.0, + 776.0, + 529.0, + 747.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 513.0, + 819.0, + 513.0, + 819.0, + 527.0, + 798.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 512.0, + 869.0, + 512.0, + 869.0, + 543.0, + 845.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 511.0, + 917.0, + 511.0, + 917.0, + 529.0, + 894.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 512.0, + 966.0, + 512.0, + 966.0, + 529.0, + 941.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 515.0, + 1116.0, + 515.0, + 1116.0, + 531.0, + 1103.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 514.0, + 1161.0, + 514.0, + 1161.0, + 532.0, + 1140.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 514.0, + 1204.0, + 514.0, + 1204.0, + 532.0, + 1180.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 514.0, + 1246.0, + 514.0, + 1246.0, + 532.0, + 1220.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 514.0, + 1287.0, + 514.0, + 1287.0, + 532.0, + 1261.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 514.0, + 1328.0, + 514.0, + 1328.0, + 532.0, + 1303.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 514.0, + 1369.0, + 514.0, + 1369.0, + 532.0, + 1345.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 535.0, + 508.0, + 535.0, + 508.0, + 550.0, + 491.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.25, + 240.5, + 631.25, + 240.5, + 631.25, + 257.0, + 599.25, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 232.0, + 728.0, + 232.0, + 728.0, + 250.0, + 699.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.75, + 340.5, + 1100.75, + 340.5, + 1100.75, + 364.5, + 1050.75, + 364.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 383.0, + 1099.0, + 383.0, + 1099.0, + 408.5, + 1052.0, + 408.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.75, + 470.0, + 726.75, + 470.0, + 726.75, + 489.0, + 698.75, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 518.0, + 461.0, + 518.0, + 461.0, + 536.0, + 432.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.75, + 525.5, + 1262.75, + 525.5, + 1262.75, + 549.5, + 1194.75, + 549.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 576.0, + 590.0, + 576.0, + 590.0, + 608.0, + 295.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 576.0, + 743.0, + 576.0, + 743.0, + 608.0, + 629.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 576.0, + 1402.0, + 576.0, + 1402.0, + 608.0, + 782.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 606.0, + 891.0, + 606.0, + 891.0, + 643.0, + 293.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 606.0, + 1403.0, + 606.0, + 1403.0, + 643.0, + 929.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 636.0, + 1405.0, + 636.0, + 1405.0, + 672.0, + 293.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 667.0, + 1406.0, + 667.0, + 1406.0, + 703.0, + 294.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 696.0, + 774.0, + 696.0, + 774.0, + 731.0, + 293.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 696.0, + 1406.0, + 696.0, + 1406.0, + 731.0, + 813.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 1276.0, + 728.0, + 1276.0, + 763.0, + 294.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1190.0, + 1038.0, + 1190.0, + 1038.0, + 1227.0, + 294.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1248.0, + 1406.0, + 1248.0, + 1406.0, + 1287.0, + 293.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1282.0, + 1404.0, + 1282.0, + 1404.0, + 1313.0, + 294.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 1094.0, + 1312.0, + 1094.0, + 1345.0, + 294.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1312.0, + 1404.0, + 1312.0, + 1404.0, + 1345.0, + 1113.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 759.0, + 1339.0, + 759.0, + 1378.0, + 294.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1339.0, + 1406.0, + 1339.0, + 1406.0, + 1378.0, + 820.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1371.0, + 1331.0, + 1371.0, + 1331.0, + 1406.0, + 293.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1371.0, + 1407.0, + 1371.0, + 1407.0, + 1406.0, + 1353.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 297.0, + 1403.0, + 297.0, + 1436.0, + 294.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1403.0, + 1403.0, + 1403.0, + 1403.0, + 1436.0, + 313.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1428.0, + 623.0, + 1428.0, + 623.0, + 1472.0, + 291.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1428.0, + 1407.0, + 1428.0, + 1407.0, + 1472.0, + 749.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1461.0, + 405.0, + 1461.0, + 405.0, + 1498.0, + 294.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 899.0, + 458.0, + 899.0, + 458.0, + 934.0, + 294.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 899.0, + 600.0, + 899.0, + 600.0, + 934.0, + 550.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 899.0, + 1405.0, + 899.0, + 1405.0, + 934.0, + 675.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 929.0, + 1404.0, + 929.0, + 1404.0, + 963.0, + 294.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 961.0, + 1405.0, + 961.0, + 1405.0, + 995.0, + 294.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 992.0, + 1403.0, + 992.0, + 1403.0, + 1025.0, + 296.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1023.0, + 1405.0, + 1023.0, + 1405.0, + 1057.0, + 297.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1053.0, + 1405.0, + 1053.0, + 1405.0, + 1087.0, + 297.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1084.0, + 1403.0, + 1084.0, + 1403.0, + 1118.0, + 296.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1116.0, + 1222.0, + 1116.0, + 1222.0, + 1146.0, + 297.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1593.0, + 663.0, + 1593.0, + 663.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1630.0, + 685.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1626.0, + 297.0, + 1626.0, + 297.0, + 1659.0, + 294.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1626.0, + 1404.0, + 1626.0, + 1404.0, + 1659.0, + 368.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1655.0, + 678.0, + 1655.0, + 678.0, + 1691.0, + 293.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1655.0, + 1405.0, + 1655.0, + 1405.0, + 1691.0, + 700.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1688.0, + 1374.0, + 1688.0, + 1374.0, + 1721.0, + 296.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1913.0, + 297.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1908.0, + 297.0, + 1908.0, + 297.0, + 1953.0, + 291.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1908.0, + 955.0, + 1908.0, + 955.0, + 1953.0, + 313.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1908.0, + 1336.0, + 1908.0, + 1336.0, + 1953.0, + 976.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1908.0, + 1407.0, + 1908.0, + 1407.0, + 1953.0, + 1394.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1407.0, + 2002.0, + 1407.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 813.0, + 598.0, + 813.0, + 598.0, + 860.0, + 295.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 813.0, + 1402.0, + 813.0, + 1402.0, + 860.0, + 655.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 844.0, + 725.0, + 844.0, + 725.0, + 889.0, + 294.0, + 889.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1258, + 1404, + 1258, + 1404, + 1565, + 297, + 1565 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1742, + 1403, + 1742, + 1403, + 2016, + 299, + 2016 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1042, + 1404, + 1042, + 1404, + 1245, + 297, + 1245 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 811, + 1404, + 811, + 1404, + 1028, + 297, + 1028 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1402, + 228, + 1402, + 488, + 297, + 488 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1571, + 1401, + 1571, + 1401, + 1723, + 299, + 1723 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 495, + 1402, + 495, + 1402, + 708, + 299, + 708 + ], + "score": 0.975 + }, + { + "category_id": 0, + "poly": [ + 300, + 754, + 575, + 754, + 575, + 785, + 300, + 785 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.831 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.795 + }, + { + "category_id": 13, + "poly": [ + 568, + 1171, + 683, + 1171, + 683, + 1209, + 568, + 1209 + ], + "score": 0.94, + "latex": "\\sigma ( \\vec { x } , l ( \\vec { f } ) )" + }, + { + "category_id": 13, + "poly": [ + 835, + 996, + 919, + 996, + 919, + 1029, + 835, + 1029 + ], + "score": 0.94, + "latex": "\\sigma ( \\vec { x } , \\tau )" + }, + { + "category_id": 13, + "poly": [ + 1137, + 1207, + 1268, + 1207, + 1268, + 1247, + 1137, + 1247 + ], + "score": 0.93, + "latex": "\\sigma ( \\vec { x } , N ( \\vec { f } ) )" + }, + { + "category_id": 13, + "poly": [ + 1190, + 1103, + 1304, + 1103, + 1304, + 1137, + 1190, + 1137 + ], + "score": 0.93, + "latex": "\\sigma ( \\vec { x } , l ( \\vec { x } ) )" + }, + { + "category_id": 13, + "poly": [ + 500, + 1473, + 597, + 1473, + 597, + 1506, + 500, + 1506 + ], + "score": 0.93, + "latex": "\\sigma ( \\vec { v } \\vec { x } , 1 )" + }, + { + "category_id": 13, + "poly": [ + 420, + 293, + 448, + 293, + 448, + 325, + 420, + 325 + ], + "score": 0.88, + "latex": "\\theta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 540, + 454, + 667, + 454, + 667, + 483, + 540, + 483 + ], + "score": 0.87, + "latex": "l r = . 0 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1201, + 263, + 1228, + 263, + 1228, + 295, + 1201, + 295 + ], + "score": 0.87, + "latex": "b _ { j }" + }, + { + "category_id": 13, + "poly": [ + 693, + 1985, + 818, + 1985, + 818, + 2013, + 693, + 2013 + ], + "score": 0.85, + "latex": "l r = 0 . 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 1137, + 265, + 1168, + 265, + 1168, + 295, + 1137, + 295 + ], + "score": 0.85, + "latex": "\\sigma _ { j }" + }, + { + "category_id": 13, + "poly": [ + 869, + 1662, + 1008, + 1662, + 1008, + 1692, + 869, + 1692 + ], + "score": 0.82, + "latex": "l r = 0 . 0 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1134, + 1365, + 1134, + 1365, + 1172, + 1345, + 1172 + ], + "score": 0.82, + "latex": "\\bar { f }" + }, + { + "category_id": 13, + "poly": [ + 605, + 1213, + 633, + 1213, + 633, + 1241, + 605, + 1241 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 889, + 1207, + 909, + 1207, + 909, + 1245, + 889, + 1245 + ], + "score": 0.81, + "latex": "\\bar { f }" + }, + { + "category_id": 13, + "poly": [ + 991, + 266, + 1012, + 266, + 1012, + 290, + 991, + 290 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1347, + 1444, + 1365, + 1444, + 1365, + 1471, + 1347, + 1471 + ], + "score": 0.75, + "latex": "\\vec { v }" + }, + { + "category_id": 13, + "poly": [ + 585, + 1001, + 604, + 1001, + 604, + 1023, + 585, + 1023 + ], + "score": 0.72, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1663, + 1110, + 1663, + 1110, + 1693, + 1019, + 1693 + ], + "score": 0.67, + "latex": "\\beta _ { 1 } = . 9" + }, + { + "category_id": 13, + "poly": [ + 1044, + 263, + 1060, + 263, + 1060, + 289, + 1044, + 289 + ], + "score": 0.64, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 830, + 1141, + 842, + 1141, + 842, + 1168, + 830, + 1168 + ], + "score": 0.64, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 438, + 420, + 494, + 420, + 494, + 454, + 438, + 454 + ], + "score": 0.6, + "latex": "[ 1 , 2 ]" + }, + { + "category_id": 13, + "poly": [ + 641, + 1105, + 653, + 1105, + 653, + 1132, + 641, + 1132 + ], + "score": 0.59, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1662, + 1223, + 1662, + 1223, + 1693, + 1020, + 1693 + ], + "score": 0.54, + "latex": "\\beta _ { 1 } = . 9 , \\beta _ { 2 } = . 9 9" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 751.0, + 579.0, + 751.0, + 579.0, + 789.0, + 294.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1254.0, + 1405.0, + 1254.0, + 1405.0, + 1298.0, + 291.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1329.0, + 292.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1316.0, + 1408.0, + 1316.0, + 1408.0, + 1358.0, + 291.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1350.0, + 1406.0, + 1350.0, + 1406.0, + 1386.0, + 295.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1381.0, + 1406.0, + 1381.0, + 1406.0, + 1417.0, + 294.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1412.0, + 1406.0, + 1412.0, + 1406.0, + 1447.0, + 295.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1442.0, + 1346.0, + 1442.0, + 1346.0, + 1478.0, + 295.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 1442.0, + 1406.0, + 1442.0, + 1406.0, + 1478.0, + 1366.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1470.0, + 499.0, + 1470.0, + 499.0, + 1508.0, + 291.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1470.0, + 1407.0, + 1470.0, + 1407.0, + 1508.0, + 598.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1502.0, + 1405.0, + 1502.0, + 1405.0, + 1539.0, + 294.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1534.0, + 725.0, + 1534.0, + 725.0, + 1569.0, + 294.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1775.0, + 297.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1771.0, + 1403.0, + 1771.0, + 1403.0, + 1804.0, + 299.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1802.0, + 1404.0, + 1802.0, + 1404.0, + 1835.0, + 298.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1831.0, + 1405.0, + 1831.0, + 1405.0, + 1867.0, + 297.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1863.0, + 1404.0, + 1863.0, + 1404.0, + 1896.0, + 299.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1893.0, + 1402.0, + 1893.0, + 1402.0, + 1926.0, + 298.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1923.0, + 1402.0, + 1923.0, + 1402.0, + 1956.0, + 302.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1950.0, + 1403.0, + 1950.0, + 1403.0, + 1989.0, + 297.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1982.0, + 692.0, + 1982.0, + 692.0, + 2019.0, + 295.0, + 2019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1982.0, + 1236.0, + 1982.0, + 1236.0, + 2019.0, + 819.0, + 2019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1044.0, + 1401.0, + 1044.0, + 1401.0, + 1074.0, + 296.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1070.0, + 1405.0, + 1070.0, + 1405.0, + 1108.0, + 292.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1103.0, + 640.0, + 1103.0, + 640.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1103.0, + 1189.0, + 1103.0, + 1189.0, + 1140.0, + 654.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1103.0, + 1405.0, + 1103.0, + 1405.0, + 1140.0, + 1305.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1138.0, + 829.0, + 1138.0, + 829.0, + 1174.0, + 292.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1138.0, + 1344.0, + 1138.0, + 1344.0, + 1174.0, + 843.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 1138.0, + 1405.0, + 1138.0, + 1405.0, + 1174.0, + 1366.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1174.0, + 567.0, + 1174.0, + 567.0, + 1211.0, + 292.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1174.0, + 1405.0, + 1174.0, + 1405.0, + 1211.0, + 684.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1209.0, + 604.0, + 1209.0, + 604.0, + 1249.0, + 292.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1209.0, + 888.0, + 1209.0, + 888.0, + 1249.0, + 634.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1209.0, + 1136.0, + 1209.0, + 1136.0, + 1249.0, + 910.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1209.0, + 1279.0, + 1209.0, + 1279.0, + 1249.0, + 1269.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 811.0, + 1402.0, + 811.0, + 1402.0, + 848.0, + 294.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 844.0, + 1404.0, + 844.0, + 1404.0, + 879.0, + 294.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 872.0, + 1407.0, + 872.0, + 1407.0, + 911.0, + 292.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 903.0, + 1406.0, + 903.0, + 1406.0, + 938.0, + 295.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 933.0, + 1407.0, + 933.0, + 1407.0, + 970.0, + 292.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 962.0, + 1409.0, + 962.0, + 1409.0, + 1001.0, + 291.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 994.0, + 584.0, + 994.0, + 584.0, + 1032.0, + 294.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 994.0, + 834.0, + 994.0, + 834.0, + 1032.0, + 605.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 994.0, + 932.0, + 994.0, + 932.0, + 1032.0, + 920.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1407.0, + 227.0, + 1407.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 990.0, + 261.0, + 990.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 261.0, + 1043.0, + 261.0, + 1043.0, + 296.0, + 1013.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 261.0, + 1136.0, + 261.0, + 1136.0, + 296.0, + 1061.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 261.0, + 1200.0, + 261.0, + 1200.0, + 296.0, + 1169.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 296.0, + 1229.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 289.0, + 419.0, + 289.0, + 419.0, + 327.0, + 291.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 289.0, + 1406.0, + 289.0, + 1406.0, + 327.0, + 449.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 321.0, + 1406.0, + 321.0, + 1406.0, + 355.0, + 295.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 352.0, + 1406.0, + 352.0, + 1406.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 385.0, + 1406.0, + 385.0, + 1406.0, + 420.0, + 295.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 417.0, + 437.0, + 417.0, + 437.0, + 455.0, + 296.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 417.0, + 1404.0, + 417.0, + 1404.0, + 455.0, + 495.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 450.0, + 539.0, + 450.0, + 539.0, + 491.0, + 295.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 450.0, + 868.0, + 450.0, + 868.0, + 491.0, + 668.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1573.0, + 1403.0, + 1573.0, + 1403.0, + 1602.0, + 299.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1598.0, + 1405.0, + 1598.0, + 1405.0, + 1635.0, + 298.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1665.0, + 297.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1660.0, + 868.0, + 1660.0, + 868.0, + 1697.0, + 297.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1660.0, + 1018.0, + 1660.0, + 1018.0, + 1697.0, + 1009.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1660.0, + 1403.0, + 1660.0, + 1403.0, + 1697.0, + 1224.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1689.0, + 910.0, + 1689.0, + 910.0, + 1726.0, + 297.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 496.0, + 1404.0, + 496.0, + 1404.0, + 527.0, + 297.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1406.0, + 520.0, + 1406.0, + 560.0, + 294.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 555.0, + 1403.0, + 555.0, + 1403.0, + 588.0, + 295.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 586.0, + 1404.0, + 586.0, + 1404.0, + 617.0, + 295.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 617.0, + 1402.0, + 617.0, + 1402.0, + 648.0, + 298.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 644.0, + 1401.0, + 644.0, + 1401.0, + 680.0, + 297.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 675.0, + 1067.0, + 675.0, + 1067.0, + 712.0, + 295.0, + 712.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1405, + 1758, + 1405, + 2034, + 297, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1221, + 1403, + 1221, + 1403, + 1626, + 298, + 1626 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 524, + 777, + 1174, + 777, + 1174, + 1060, + 524, + 1060 + ], + "score": 0.977, + "html": "
DatasetUncalibrated IsotonicGSLSDS
crime0.36240.34990.0693 0.0125 0.0310
kinematics0.01640.01030.0022 0.0021 0.0032
bank0.01220.00560.0027 0.0024 0.0020
wine0.00910.01080.0152 0.01310.0064
mpg0.21530.22000.1964 0.14830.0233
cpu0.08620.03400.3018 0.20780.1740
soil0.30830.30000.3130 0.3175 0.3137
fried0.00060.00020.00020.00020.0002
" + }, + { + "category_id": 5, + "poly": [ + 402, + 365, + 1295, + 365, + 1295, + 647, + 402, + 647 + ], + "score": 0.976, + "html": "
ModelDatasetUncalibratedPlattIsotonicGSVSLSLFS
RN50CIFAR-10.250.034.053.046.037.018.018
RN50CIFAR-100.642.061.072.035.044.030.173
RN50SVHN.072.053.010.029.022.009.009
RN50ImageNet.430.018.070.019.023.026.015
DN121CIFAR-10.253.048.042.039.034.028.028
DN121CIFAR-100.537.049.067.024.024.014.031
DN121SVHN.079.018.010.022.017.011.010
DN121ImageNet.229.028.095.021.019.043.019
" + }, + { + "category_id": 6, + "poly": [ + 298, + 223, + 1402, + 223, + 1402, + 347, + 298, + 347 + ], + "score": 0.923 + }, + { + "category_id": 6, + "poly": [ + 295, + 696, + 1402, + 696, + 1402, + 760, + 295, + 760 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 300, + 1684, + 544, + 1684, + 544, + 1719, + 300, + 1719 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 302, + 1152, + 709, + 1152, + 709, + 1189, + 302, + 1189 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.797 + }, + { + "category_id": 13, + "poly": [ + 961, + 1556, + 1171, + 1556, + 1171, + 1598, + 961, + 1598 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { s ( x ) = \\frac { \\tilde { 1 } } { 1 + \\exp ( - x ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 390, + 1558, + 447, + 1558, + 447, + 1591, + 390, + 1591 + ], + "score": 0.92, + "latex": "[ a , b ]" + }, + { + "category_id": 13, + "poly": [ + 919, + 1343, + 992, + 1343, + 992, + 1378, + 919, + 1378 + ], + "score": 0.92, + "latex": "t ^ { - 1 } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1314, + 425, + 1314, + 425, + 1347, + 297, + 1347 + ], + "score": 0.92, + "latex": "\\sigma \\in [ 0 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 392, + 1496, + 438, + 1496, + 438, + 1531, + 392, + 1531 + ], + "score": 0.91, + "latex": "s _ { . 0 1 } ^ { \\mp }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1435, + 595, + 1435, + 595, + 1470, + 296, + 1470 + ], + "score": 0.91, + "latex": "s ^ { + } ( x ) = \\log ( 1 + \\exp ( x ) )" + }, + { + "category_id": 13, + "poly": [ + 490, + 1496, + 525, + 1496, + 525, + 1531, + 490, + 1531 + ], + "score": 0.91, + "latex": "s _ { . 2 } ^ { + }" + }, + { + "category_id": 13, + "poly": [ + 946, + 1436, + 1396, + 1436, + 1396, + 1470, + 946, + 1470 + ], + "score": 0.91, + "latex": "s _ { c } ^ { + } ( x ) = ( l n ( 1 + e ^ { x } ) + c ) / ( l n ( 2 ) + c ) ," + }, + { + "category_id": 13, + "poly": [ + 726, + 1345, + 769, + 1345, + 769, + 1378, + 726, + 1378 + ], + "score": 0.89, + "latex": "t ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 358, + 1348, + 393, + 1348, + 393, + 1376, + 358, + 1376 + ], + "score": 0.89, + "latex": "\\phi _ { u }" + }, + { + "category_id": 13, + "poly": [ + 1227, + 1594, + 1355, + 1594, + 1355, + 1624, + 1227, + 1624 + ], + "score": 0.85, + "latex": "\\sigma = \\tau = 1 \\mathrm { { } } " + }, + { + "category_id": 13, + "poly": [ + 296, + 1376, + 318, + 1376, + 318, + 1407, + 296, + 1407 + ], + "score": 0.83, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 450, + 1255, + 489, + 1255, + 489, + 1282, + 450, + 1282 + ], + "score": 0.51, + "latex": "4 \\times" + }, + { + "category_id": 13, + "poly": [ + 449, + 1254, + 629, + 1254, + 629, + 1285, + 449, + 1285 + ], + "score": 0.31, + "latex": "4 \\times \\mathrm { M } 6 0 \\ : \\mathrm { G P U s } )" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 224.0, + 1402.0, + 224.0, + 1402.0, + 257.0, + 294.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 252.0, + 1406.0, + 252.0, + 1406.0, + 290.0, + 293.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 286.0, + 1406.0, + 286.0, + 1406.0, + 319.0, + 296.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 315.0, + 1014.0, + 315.0, + 1014.0, + 348.0, + 293.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 695.0, + 1406.0, + 695.0, + 1406.0, + 733.0, + 294.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 725.0, + 1404.0, + 725.0, + 1404.0, + 763.0, + 294.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1678.0, + 550.0, + 1678.0, + 550.0, + 1729.0, + 291.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1147.0, + 714.0, + 1147.0, + 714.0, + 1194.0, + 296.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1759.0, + 1402.0, + 1759.0, + 1402.0, + 1792.0, + 297.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1786.0, + 1406.0, + 1786.0, + 1406.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1819.0, + 1405.0, + 1819.0, + 1405.0, + 1854.0, + 295.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1402.0, + 1942.0, + 1402.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 806.0, + 2002.0, + 806.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1221.0, + 1403.0, + 1221.0, + 1403.0, + 1255.0, + 297.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1252.0, + 448.0, + 1252.0, + 448.0, + 1285.0, + 294.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1252.0, + 1404.0, + 1252.0, + 1404.0, + 1285.0, + 630.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1284.0, + 1403.0, + 1284.0, + 1403.0, + 1314.0, + 298.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1315.0, + 1400.0, + 1315.0, + 1400.0, + 1346.0, + 426.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1343.0, + 357.0, + 1343.0, + 357.0, + 1377.0, + 297.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1343.0, + 725.0, + 1343.0, + 725.0, + 1377.0, + 394.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1343.0, + 918.0, + 1343.0, + 918.0, + 1377.0, + 770.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1343.0, + 1404.0, + 1343.0, + 1404.0, + 1377.0, + 993.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1374.0, + 1403.0, + 1374.0, + 1403.0, + 1408.0, + 319.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1403.0, + 1403.0, + 1403.0, + 1403.0, + 1438.0, + 297.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1432.0, + 945.0, + 1432.0, + 945.0, + 1472.0, + 596.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1432.0, + 1403.0, + 1432.0, + 1403.0, + 1472.0, + 1397.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1467.0, + 1403.0, + 1467.0, + 1403.0, + 1498.0, + 299.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1495.0, + 391.0, + 1495.0, + 391.0, + 1533.0, + 296.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1495.0, + 489.0, + 1495.0, + 489.0, + 1533.0, + 439.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1495.0, + 1402.0, + 1495.0, + 1402.0, + 1533.0, + 526.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1525.0, + 1403.0, + 1525.0, + 1403.0, + 1559.0, + 297.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1553.0, + 389.0, + 1553.0, + 389.0, + 1602.0, + 296.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1553.0, + 960.0, + 1553.0, + 960.0, + 1602.0, + 448.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1553.0, + 1402.0, + 1553.0, + 1402.0, + 1602.0, + 1172.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1592.0, + 1226.0, + 1592.0, + 1226.0, + 1626.0, + 298.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1592.0, + 1375.0, + 1592.0, + 1375.0, + 1626.0, + 1356.0, + 1626.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 300, + 279, + 1404, + 279, + 1404, + 553, + 300, + 553 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1006, + 1404, + 1006, + 1404, + 1098, + 299, + 1098 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 300, + 893, + 1401, + 893, + 1401, + 987, + 300, + 987 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 295, + 812, + 1401, + 812, + 1401, + 873, + 295, + 873 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 299, + 625, + 1396, + 625, + 1396, + 688, + 299, + 688 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 297, + 1529, + 1404, + 1529, + 1404, + 1623, + 297, + 1623 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 296, + 1447, + 1397, + 1447, + 1397, + 1511, + 296, + 1511 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 297, + 1118, + 1396, + 1118, + 1396, + 1182, + 297, + 1182 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 298, + 1199, + 1396, + 1199, + 1396, + 1265, + 298, + 1265 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 295, + 1642, + 1399, + 1642, + 1399, + 1706, + 295, + 1706 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 298, + 760, + 930, + 760, + 930, + 791, + 298, + 791 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 296, + 707, + 1197, + 707, + 1197, + 739, + 296, + 739 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 298, + 1365, + 1394, + 1365, + 1394, + 1428, + 298, + 1428 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 300, + 1283, + 1395, + 1283, + 1395, + 1346, + 300, + 1346 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 297, + 1724, + 1397, + 1724, + 1397, + 1788, + 297, + 1788 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 854, + 76, + 854, + 104, + 298, + 104 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 298, + 1807, + 1399, + 1807, + 1399, + 1871, + 298, + 1871 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1398, + 1971, + 1398, + 2034, + 297, + 2034 + ], + "score": 0.825 + }, + { + "category_id": 1, + "poly": [ + 297, + 1889, + 1397, + 1889, + 1397, + 1953, + 297, + 1953 + ], + "score": 0.821 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.82 + }, + { + "category_id": 1, + "poly": [ + 300, + 573, + 1223, + 573, + 1223, + 607, + 300, + 607 + ], + "score": 0.68 + }, + { + "category_id": 13, + "poly": [ + 884, + 895, + 923, + 895, + 923, + 926, + 884, + 926 + ], + "score": 0.68, + "latex": "\\mathrm { N g }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 278.0, + 1404.0, + 278.0, + 1404.0, + 315.0, + 294.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 308.0, + 1405.0, + 308.0, + 1405.0, + 345.0, + 323.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 339.0, + 1405.0, + 339.0, + 1405.0, + 376.0, + 323.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 369.0, + 1405.0, + 369.0, + 1405.0, + 406.0, + 321.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 437.0, + 321.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 428.0, + 1405.0, + 428.0, + 1405.0, + 469.0, + 319.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 458.0, + 1405.0, + 458.0, + 1405.0, + 500.0, + 320.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 493.0, + 1404.0, + 493.0, + 1404.0, + 529.0, + 323.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 518.0, + 496.0, + 518.0, + 496.0, + 562.0, + 321.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1003.0, + 1403.0, + 1003.0, + 1403.0, + 1044.0, + 293.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1039.0, + 1403.0, + 1039.0, + 1403.0, + 1072.0, + 324.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1064.0, + 505.0, + 1064.0, + 505.0, + 1101.0, + 323.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 889.0, + 883.0, + 889.0, + 883.0, + 932.0, + 293.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 889.0, + 1401.0, + 889.0, + 1401.0, + 932.0, + 924.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 923.0, + 1405.0, + 923.0, + 1405.0, + 959.0, + 323.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 953.0, + 775.0, + 953.0, + 775.0, + 991.0, + 323.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 807.0, + 1405.0, + 807.0, + 1405.0, + 849.0, + 294.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 841.0, + 462.0, + 841.0, + 462.0, + 874.0, + 324.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 623.0, + 1402.0, + 623.0, + 1402.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 655.0, + 1366.0, + 655.0, + 1366.0, + 691.0, + 323.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1526.0, + 1405.0, + 1526.0, + 1405.0, + 1568.0, + 292.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1561.0, + 1404.0, + 1561.0, + 1404.0, + 1595.0, + 324.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1591.0, + 546.0, + 1591.0, + 546.0, + 1623.0, + 322.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1443.0, + 1403.0, + 1443.0, + 1403.0, + 1485.0, + 294.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1476.0, + 794.0, + 1476.0, + 794.0, + 1512.0, + 322.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1118.0, + 1401.0, + 1118.0, + 1401.0, + 1154.0, + 297.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1149.0, + 1226.0, + 1149.0, + 1226.0, + 1185.0, + 322.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 1401.0, + 1198.0, + 1401.0, + 1237.0, + 293.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1231.0, + 1045.0, + 1231.0, + 1045.0, + 1266.0, + 322.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1642.0, + 1404.0, + 1642.0, + 1404.0, + 1678.0, + 296.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1673.0, + 900.0, + 1673.0, + 900.0, + 1707.0, + 321.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 758.0, + 931.0, + 758.0, + 931.0, + 796.0, + 296.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 705.0, + 1199.0, + 705.0, + 1199.0, + 744.0, + 294.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1363.0, + 1400.0, + 1363.0, + 1400.0, + 1403.0, + 294.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1397.0, + 569.0, + 1397.0, + 569.0, + 1428.0, + 324.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1284.0, + 1398.0, + 1284.0, + 1398.0, + 1320.0, + 295.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1313.0, + 1220.0, + 1313.0, + 1220.0, + 1348.0, + 325.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1724.0, + 1402.0, + 1724.0, + 1402.0, + 1760.0, + 296.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1757.0, + 800.0, + 1757.0, + 800.0, + 1789.0, + 324.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1807.0, + 1403.0, + 1807.0, + 1403.0, + 1843.0, + 296.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1838.0, + 807.0, + 1838.0, + 807.0, + 1872.0, + 322.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 622.0, + 2003.0, + 622.0, + 2033.0, + 322.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 1402.0, + 1888.0, + 1402.0, + 1926.0, + 293.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1920.0, + 1239.0, + 1920.0, + 1239.0, + 1956.0, + 322.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 572.0, + 1227.0, + 572.0, + 1227.0, + 611.0, + 296.0, + 611.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.82 + }, + { + "category_id": 1, + "poly": [ + 300, + 825, + 1402, + 825, + 1402, + 890, + 300, + 890 + ], + "score": 0.806 + }, + { + "category_id": 1, + "poly": [ + 297, + 770, + 1103, + 770, + 1103, + 805, + 297, + 805 + ], + "score": 0.799 + }, + { + "category_id": 1, + "poly": [ + 294, + 1051, + 1402, + 1051, + 1402, + 1116, + 294, + 1116 + ], + "score": 0.79 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 1399, + 228, + 1399, + 323, + 295, + 323 + ], + "score": 0.782 + }, + { + "category_id": 1, + "poly": [ + 295, + 995, + 1359, + 995, + 1359, + 1031, + 295, + 1031 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 289, + 1137, + 1400, + 1137, + 1400, + 1201, + 289, + 1201 + ], + "score": 0.777 + }, + { + "category_id": 1, + "poly": [ + 297, + 1337, + 1399, + 1337, + 1399, + 1402, + 297, + 1402 + ], + "score": 0.772 + }, + { + "category_id": 1, + "poly": [ + 298, + 686, + 1402, + 686, + 1402, + 750, + 298, + 750 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 296, + 514, + 1398, + 514, + 1398, + 580, + 296, + 580 + ], + "score": 0.758 + }, + { + "category_id": 1, + "poly": [ + 296, + 910, + 1400, + 910, + 1400, + 977, + 296, + 977 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 297, + 1684, + 1403, + 1684, + 1403, + 1779, + 297, + 1779 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 298, + 1222, + 1399, + 1222, + 1399, + 1318, + 298, + 1318 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 296, + 600, + 1402, + 600, + 1402, + 665, + 296, + 665 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 296, + 1568, + 1407, + 1568, + 1407, + 1664, + 296, + 1664 + ], + "score": 0.72 + }, + { + "category_id": 1, + "poly": [ + 296, + 429, + 1397, + 429, + 1397, + 494, + 296, + 494 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 297, + 1422, + 1405, + 1422, + 1405, + 1547, + 297, + 1547 + ], + "score": 0.689 + }, + { + "category_id": 1, + "poly": [ + 294, + 345, + 1401, + 345, + 1401, + 409, + 294, + 409 + ], + "score": 0.649 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1398, + 1971, + 1398, + 2035, + 297, + 2035 + ], + "score": 0.642 + }, + { + "category_id": 1, + "poly": [ + 291, + 1799, + 1402, + 1799, + 1402, + 1865, + 291, + 1865 + ], + "score": 0.618 + }, + { + "category_id": 1, + "poly": [ + 296, + 1885, + 1399, + 1885, + 1399, + 1951, + 296, + 1951 + ], + "score": 0.615 + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 824.0, + 1403.0, + 824.0, + 1403.0, + 864.0, + 295.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 858.0, + 803.0, + 858.0, + 803.0, + 890.0, + 325.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 770.0, + 1102.0, + 770.0, + 1102.0, + 809.0, + 294.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1047.0, + 1407.0, + 1047.0, + 1407.0, + 1091.0, + 292.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1082.0, + 597.0, + 1082.0, + 597.0, + 1113.0, + 322.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1403.0, + 227.0, + 1403.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 262.0, + 1403.0, + 262.0, + 1403.0, + 296.0, + 323.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 291.0, + 506.0, + 291.0, + 506.0, + 323.0, + 324.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 995.0, + 1359.0, + 995.0, + 1359.0, + 1032.0, + 293.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1136.0, + 1403.0, + 1136.0, + 1403.0, + 1175.0, + 293.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1166.0, + 751.0, + 1166.0, + 751.0, + 1202.0, + 322.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1333.0, + 1403.0, + 1333.0, + 1403.0, + 1377.0, + 293.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1368.0, + 525.0, + 1368.0, + 525.0, + 1399.0, + 322.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 686.0, + 1404.0, + 686.0, + 1404.0, + 722.0, + 296.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 718.0, + 782.0, + 718.0, + 782.0, + 749.0, + 323.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 515.0, + 1404.0, + 515.0, + 1404.0, + 551.0, + 295.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 547.0, + 868.0, + 547.0, + 868.0, + 579.0, + 322.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 910.0, + 1404.0, + 910.0, + 1404.0, + 949.0, + 295.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 944.0, + 1079.0, + 944.0, + 1079.0, + 977.0, + 321.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1403.0, + 1682.0, + 1403.0, + 1723.0, + 294.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1714.0, + 1406.0, + 1714.0, + 1406.0, + 1751.0, + 320.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1745.0, + 464.0, + 1745.0, + 464.0, + 1780.0, + 321.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1221.0, + 1401.0, + 1221.0, + 1401.0, + 1258.0, + 294.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1254.0, + 1403.0, + 1254.0, + 1403.0, + 1288.0, + 324.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1285.0, + 914.0, + 1285.0, + 914.0, + 1319.0, + 324.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 600.0, + 1404.0, + 600.0, + 1404.0, + 636.0, + 295.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 634.0, + 707.0, + 634.0, + 707.0, + 663.0, + 320.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1566.0, + 1408.0, + 1566.0, + 1408.0, + 1607.0, + 294.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1639.0, + 321.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1631.0, + 505.0, + 1631.0, + 505.0, + 1663.0, + 325.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 423.0, + 1403.0, + 423.0, + 1403.0, + 473.0, + 291.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 460.0, + 590.0, + 460.0, + 590.0, + 494.0, + 321.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1425.0, + 1403.0, + 1425.0, + 1403.0, + 1458.0, + 297.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1454.0, + 1406.0, + 1454.0, + 1406.0, + 1488.0, + 320.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1484.0, + 1405.0, + 1484.0, + 1405.0, + 1520.0, + 323.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1515.0, + 699.0, + 1515.0, + 699.0, + 1548.0, + 323.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 344.0, + 1404.0, + 344.0, + 1404.0, + 381.0, + 294.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 376.0, + 1193.0, + 376.0, + 1193.0, + 410.0, + 321.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1402.0, + 1971.0, + 1402.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 1115.0, + 2003.0, + 1115.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1798.0, + 1404.0, + 1798.0, + 1404.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1831.0, + 892.0, + 1831.0, + 892.0, + 1866.0, + 322.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1925.0, + 292.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1915.0, + 1015.0, + 1915.0, + 1015.0, + 1952.0, + 321.0, + 1952.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.832 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1400, + 229, + 1400, + 294, + 294, + 294 + ], + "score": 0.81 + }, + { + "category_id": 1, + "poly": [ + 299, + 494, + 1404, + 494, + 1404, + 558, + 299, + 558 + ], + "score": 0.784 + }, + { + "category_id": 1, + "poly": [ + 297, + 1302, + 1400, + 1302, + 1400, + 1366, + 297, + 1366 + ], + "score": 0.778 + }, + { + "category_id": 1, + "poly": [ + 297, + 1253, + 1018, + 1253, + 1018, + 1288, + 297, + 1288 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 296, + 650, + 1398, + 650, + 1398, + 715, + 296, + 715 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 299, + 306, + 1398, + 306, + 1398, + 372, + 299, + 372 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 300, + 572, + 1400, + 572, + 1400, + 637, + 300, + 637 + ], + "score": 0.753 + }, + { + "category_id": 1, + "poly": [ + 297, + 1676, + 1400, + 1676, + 1400, + 1740, + 297, + 1740 + ], + "score": 0.739 + }, + { + "category_id": 1, + "poly": [ + 294, + 1380, + 1399, + 1380, + 1399, + 1446, + 294, + 1446 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 295, + 1068, + 1398, + 1068, + 1398, + 1133, + 295, + 1133 + ], + "score": 0.714 + }, + { + "category_id": 1, + "poly": [ + 295, + 1862, + 1397, + 1862, + 1397, + 1928, + 295, + 1928 + ], + "score": 0.698 + }, + { + "category_id": 1, + "poly": [ + 301, + 1942, + 1402, + 1942, + 1402, + 2034, + 301, + 2034 + ], + "score": 0.695 + }, + { + "category_id": 1, + "poly": [ + 299, + 384, + 1402, + 384, + 1402, + 478, + 299, + 478 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 297, + 1458, + 1399, + 1458, + 1399, + 1552, + 297, + 1552 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 298, + 729, + 1405, + 729, + 1405, + 945, + 298, + 945 + ], + "score": 0.664 + }, + { + "category_id": 1, + "poly": [ + 297, + 959, + 1404, + 959, + 1404, + 1054, + 297, + 1054 + ], + "score": 0.631 + }, + { + "category_id": 1, + "poly": [ + 297, + 1754, + 1405, + 1754, + 1405, + 1847, + 297, + 1847 + ], + "score": 0.616 + }, + { + "category_id": 1, + "poly": [ + 296, + 1567, + 1404, + 1567, + 1404, + 1661, + 296, + 1661 + ], + "score": 0.577 + }, + { + "category_id": 1, + "poly": [ + 296, + 1146, + 1405, + 1146, + 1405, + 1240, + 296, + 1240 + ], + "score": 0.548 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 262.0, + 622.0, + 262.0, + 622.0, + 292.0, + 322.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 494.0, + 1404.0, + 494.0, + 1404.0, + 530.0, + 294.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 525.0, + 1095.0, + 525.0, + 1095.0, + 561.0, + 324.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1302.0, + 1404.0, + 1302.0, + 1404.0, + 1338.0, + 295.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1333.0, + 947.0, + 1333.0, + 947.0, + 1365.0, + 323.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1252.0, + 1020.0, + 1252.0, + 1020.0, + 1294.0, + 293.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 648.0, + 1402.0, + 648.0, + 1402.0, + 688.0, + 294.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 681.0, + 962.0, + 681.0, + 962.0, + 717.0, + 324.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 305.0, + 1404.0, + 305.0, + 1404.0, + 345.0, + 294.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 338.0, + 712.0, + 338.0, + 712.0, + 374.0, + 325.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 567.0, + 1405.0, + 567.0, + 1405.0, + 612.0, + 295.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 603.0, + 1062.0, + 603.0, + 1062.0, + 639.0, + 322.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1676.0, + 1404.0, + 1676.0, + 1404.0, + 1712.0, + 295.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1707.0, + 1172.0, + 1707.0, + 1172.0, + 1741.0, + 322.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1381.0, + 1404.0, + 1381.0, + 1404.0, + 1417.0, + 295.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1413.0, + 1190.0, + 1413.0, + 1190.0, + 1445.0, + 323.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1067.0, + 1403.0, + 1067.0, + 1403.0, + 1104.0, + 293.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1097.0, + 1191.0, + 1097.0, + 1191.0, + 1137.0, + 321.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1861.0, + 1403.0, + 1861.0, + 1403.0, + 1899.0, + 294.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1894.0, + 1219.0, + 1894.0, + 1219.0, + 1929.0, + 321.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1977.0, + 296.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1969.0, + 1402.0, + 1969.0, + 1402.0, + 2010.0, + 320.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2004.0, + 706.0, + 2004.0, + 706.0, + 2033.0, + 321.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 386.0, + 1402.0, + 386.0, + 1402.0, + 420.0, + 295.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 414.0, + 1406.0, + 414.0, + 1406.0, + 454.0, + 320.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 446.0, + 396.0, + 446.0, + 396.0, + 480.0, + 320.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1459.0, + 1401.0, + 1459.0, + 1401.0, + 1493.0, + 295.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1488.0, + 1403.0, + 1488.0, + 1403.0, + 1528.0, + 321.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1521.0, + 611.0, + 1521.0, + 611.0, + 1552.0, + 322.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 726.0, + 1405.0, + 726.0, + 1405.0, + 768.0, + 292.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 760.0, + 1405.0, + 760.0, + 1405.0, + 795.0, + 323.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 786.0, + 1405.0, + 786.0, + 1405.0, + 828.0, + 321.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 820.0, + 1405.0, + 820.0, + 1405.0, + 857.0, + 321.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 852.0, + 1405.0, + 852.0, + 1405.0, + 886.0, + 322.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 881.0, + 1406.0, + 881.0, + 1406.0, + 920.0, + 320.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 914.0, + 581.0, + 914.0, + 581.0, + 944.0, + 323.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 959.0, + 1406.0, + 959.0, + 1406.0, + 997.0, + 294.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 989.0, + 1402.0, + 989.0, + 1402.0, + 1026.0, + 322.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1022.0, + 1381.0, + 1022.0, + 1381.0, + 1056.0, + 324.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1751.0, + 1405.0, + 1751.0, + 1405.0, + 1792.0, + 294.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1786.0, + 1403.0, + 1786.0, + 1403.0, + 1820.0, + 323.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1814.0, + 438.0, + 1814.0, + 438.0, + 1849.0, + 321.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1567.0, + 1404.0, + 1567.0, + 1404.0, + 1604.0, + 294.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1598.0, + 1405.0, + 1598.0, + 1405.0, + 1632.0, + 321.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1631.0, + 1268.0, + 1631.0, + 1268.0, + 1661.0, + 325.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1147.0, + 1403.0, + 1147.0, + 1403.0, + 1181.0, + 296.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1176.0, + 1405.0, + 1176.0, + 1405.0, + 1212.0, + 323.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1208.0, + 787.0, + 1208.0, + 787.0, + 1241.0, + 323.0, + 1241.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/S1lvm305YQ/images/0b3de429d2f94633aaffe960365e05c58edb90d552a707b2b674480e8e0c37d1.jpg b/parse/train/S1lvm305YQ/images/0b3de429d2f94633aaffe960365e05c58edb90d552a707b2b674480e8e0c37d1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5bb5e91b1a658675109746008ae5432093724cf2 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/0b3de429d2f94633aaffe960365e05c58edb90d552a707b2b674480e8e0c37d1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42c6cd088de7e72be5f2f13adf99309e293fc774f79fc55e249314d02d7d0976 +size 8284 diff --git a/parse/train/S1lvm305YQ/images/26e1473a6fb95c6c5b9dddb56a25a26f03182405a73147d1976579d5068a1309.jpg b/parse/train/S1lvm305YQ/images/26e1473a6fb95c6c5b9dddb56a25a26f03182405a73147d1976579d5068a1309.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5252b47907d4dca62efee677018a2f55947310b5 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/26e1473a6fb95c6c5b9dddb56a25a26f03182405a73147d1976579d5068a1309.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47b1fc2c628e8b4bb5a619ac3c93c8d932bb39e9c918275454497edd97d3cecd +size 6207 diff --git a/parse/train/S1lvm305YQ/images/41371be6b193eb7d6fb3c03ca64ee56a0036d7ca4387870175f5e01b7f578c32.jpg b/parse/train/S1lvm305YQ/images/41371be6b193eb7d6fb3c03ca64ee56a0036d7ca4387870175f5e01b7f578c32.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ac183f71f7182c143401457335b7d2897ad53d06 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/41371be6b193eb7d6fb3c03ca64ee56a0036d7ca4387870175f5e01b7f578c32.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:789a0d5aa10fc90443aa04f913fc23b7a1149bfebd215d40480e873b77dc3e35 +size 47165 diff --git a/parse/train/S1lvm305YQ/images/4d9cdac31dffec220d75296140c7bee749082676986a3aa1f54d2fe1fc008b7c.jpg b/parse/train/S1lvm305YQ/images/4d9cdac31dffec220d75296140c7bee749082676986a3aa1f54d2fe1fc008b7c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..22903109ed362d738bd07eef772e87e1c8befb9a --- /dev/null +++ b/parse/train/S1lvm305YQ/images/4d9cdac31dffec220d75296140c7bee749082676986a3aa1f54d2fe1fc008b7c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56c20a76d036049c842f5fed477c458938509cc097f3bf26eb97432ffc82f5a3 +size 125492 diff --git a/parse/train/S1lvm305YQ/images/588ca95db06e0cb4d2f5461ab179ebaad6cb0d8a3e1d8e21ddbd5f79f3a1f5a9.jpg b/parse/train/S1lvm305YQ/images/588ca95db06e0cb4d2f5461ab179ebaad6cb0d8a3e1d8e21ddbd5f79f3a1f5a9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..20a833050d6ff4f3916cda8b9897c4ecfd68575c --- /dev/null +++ b/parse/train/S1lvm305YQ/images/588ca95db06e0cb4d2f5461ab179ebaad6cb0d8a3e1d8e21ddbd5f79f3a1f5a9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df17e2a20834cfcd36a700a713c762f58d2833d850241145238d4be316163c8c +size 22999 diff --git a/parse/train/S1lvm305YQ/images/650c98590ee92220698cc8c588386fb04e4e3329c9d34666a908dac28eca9b96.jpg b/parse/train/S1lvm305YQ/images/650c98590ee92220698cc8c588386fb04e4e3329c9d34666a908dac28eca9b96.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2b79550c024ec7185bd97e7285b83f25d27e4989 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/650c98590ee92220698cc8c588386fb04e4e3329c9d34666a908dac28eca9b96.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee520ee3f69402d5693c3bdb1703fd10fa71c5f2adf64403a5bf8735a4d72985 +size 7912 diff --git a/parse/train/S1lvm305YQ/images/73562af20c5395070a563fbb1b5c00b015626f906f446a46d6eff3e271de2a9c.jpg b/parse/train/S1lvm305YQ/images/73562af20c5395070a563fbb1b5c00b015626f906f446a46d6eff3e271de2a9c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..895b0b67ce5b77e70c88fa7c467fa33bfee5755f --- /dev/null +++ b/parse/train/S1lvm305YQ/images/73562af20c5395070a563fbb1b5c00b015626f906f446a46d6eff3e271de2a9c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4a4ea7f1f215a36f1a85ce9145b31facf1463a6e554eaff3f02c618446d9f56 +size 8541 diff --git a/parse/train/S1lvm305YQ/images/a57fed0d78aca8db561770bf227181c481148ab08f6c4d404ef43aa4071cf60d.jpg b/parse/train/S1lvm305YQ/images/a57fed0d78aca8db561770bf227181c481148ab08f6c4d404ef43aa4071cf60d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ae6cfe4b80dc5f91e90b96f6eb29ca97bf75e050 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/a57fed0d78aca8db561770bf227181c481148ab08f6c4d404ef43aa4071cf60d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3af71f3a56524ed2e2cf5601c3e19d51ff2ed1fd38daefca63bcb58cef05fe8 +size 31211 diff --git a/parse/train/S1lvm305YQ/images/b196faa01a4f8f12c96ad25ed5a68594082cb3925703f32514f0eb139f5fce21.jpg b/parse/train/S1lvm305YQ/images/b196faa01a4f8f12c96ad25ed5a68594082cb3925703f32514f0eb139f5fce21.jpg new file mode 100644 index 0000000000000000000000000000000000000000..598d53493c039df290375879c0c964410daa7457 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/b196faa01a4f8f12c96ad25ed5a68594082cb3925703f32514f0eb139f5fce21.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:757c62650efbd9863a306cdf1cb530f6b5ddcf070b9f06886213b868eef45597 +size 22188 diff --git a/parse/train/S1lvm305YQ/images/b1f409115a04244b694817878bd6141c021e063d063f1f26e0d9c9e55d01559e.jpg b/parse/train/S1lvm305YQ/images/b1f409115a04244b694817878bd6141c021e063d063f1f26e0d9c9e55d01559e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..73ffefc2ec627b18773d9055739f669a45011452 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/b1f409115a04244b694817878bd6141c021e063d063f1f26e0d9c9e55d01559e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:919f1804663f4aece0a94f7ef37dd1bcfdc8b6339410fa7ad1747664dba4eae0 +size 25948 diff --git a/parse/train/S1lvm305YQ/images/baf5cae67422b2418efbf16466cad51c13df39e45c1778ff1515c43061fb772e.jpg b/parse/train/S1lvm305YQ/images/baf5cae67422b2418efbf16466cad51c13df39e45c1778ff1515c43061fb772e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9717292bd8e2abd497857bfc85c429d3b72a346e --- /dev/null +++ b/parse/train/S1lvm305YQ/images/baf5cae67422b2418efbf16466cad51c13df39e45c1778ff1515c43061fb772e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0b0c3ca620fbdea8a850daae86d2539cc07cf9feea2793fabbbe53536add422 +size 27650 diff --git a/parse/train/S1lvm305YQ/images/bbc4130d241fb032b73b300d307ab4c980870ea9128731e899ab48acaad1b118.jpg b/parse/train/S1lvm305YQ/images/bbc4130d241fb032b73b300d307ab4c980870ea9128731e899ab48acaad1b118.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2175159bf805c57537b0daed485a21663a484f50 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/bbc4130d241fb032b73b300d307ab4c980870ea9128731e899ab48acaad1b118.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2a1a860d3446d8f92f4519ef51e82d0421f201deb7c8cbedbcf1c0bb96d2b7a +size 15411 diff --git a/parse/train/S1lvm305YQ/images/dd96c0afbefe1b997c59d30b2df1db24e4e23ffa173c45adee3f5f942c80006f.jpg b/parse/train/S1lvm305YQ/images/dd96c0afbefe1b997c59d30b2df1db24e4e23ffa173c45adee3f5f942c80006f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..68e6e44c6355f8da0ab8f9ed47c6fe0a0001731a --- /dev/null +++ b/parse/train/S1lvm305YQ/images/dd96c0afbefe1b997c59d30b2df1db24e4e23ffa173c45adee3f5f942c80006f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1af24d207b09d5ed2927b979094eb3ee7e2e0d36840ff486681b38f6d2f6954 +size 7250 diff --git a/parse/train/S1lvm305YQ/images/de75b77444d1cb9df9e4c43890831d517f1634117d48305676474206e7c4530c.jpg b/parse/train/S1lvm305YQ/images/de75b77444d1cb9df9e4c43890831d517f1634117d48305676474206e7c4530c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1d248e35c3e4ba5a2f33b3d61e58eafc5abd08e9 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/de75b77444d1cb9df9e4c43890831d517f1634117d48305676474206e7c4530c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b06616d4895ad02799c8434214e54dc43e865870d980d5b7eda94871238deb49 +size 3141 diff --git a/parse/train/S1lvm305YQ/images/e442b9be8a1678a84bb23998a313c9a0c0c16bbf0976d414ba21b7f0dc9970a9.jpg b/parse/train/S1lvm305YQ/images/e442b9be8a1678a84bb23998a313c9a0c0c16bbf0976d414ba21b7f0dc9970a9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b875b0ad981d70dd6330fc13dc8aaa5200c44bb0 --- /dev/null +++ b/parse/train/S1lvm305YQ/images/e442b9be8a1678a84bb23998a313c9a0c0c16bbf0976d414ba21b7f0dc9970a9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4cac88ac44c83ce7666811d4fb4e6a2fe38bb986dd20ca00793a02b768cf319 +size 45641 diff --git a/parse/train/S1lvm305YQ/images/fd1942a5fb6327a5b8a66207cea9b6c2e1d89925b67093cccfa7f1f4346c11ab.jpg b/parse/train/S1lvm305YQ/images/fd1942a5fb6327a5b8a66207cea9b6c2e1d89925b67093cccfa7f1f4346c11ab.jpg new file mode 100644 index 0000000000000000000000000000000000000000..78e3764c443519469fca509a003674615a5eb09f --- /dev/null +++ b/parse/train/S1lvm305YQ/images/fd1942a5fb6327a5b8a66207cea9b6c2e1d89925b67093cccfa7f1f4346c11ab.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b55fd57775072e2ac8f801e3be81dccd8d1848e138537433b72aa4b10d7cd5c +size 44750 diff --git a/parse/train/SkVhlh09tX/images/13d2a5a765bcaf746bd19db6b98f82965f33cd9fe8fdcd167d82b6a2309204aa.jpg b/parse/train/SkVhlh09tX/images/13d2a5a765bcaf746bd19db6b98f82965f33cd9fe8fdcd167d82b6a2309204aa.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3993ee492258ba20e9554c99d192b24f4206b3c3 --- /dev/null +++ b/parse/train/SkVhlh09tX/images/13d2a5a765bcaf746bd19db6b98f82965f33cd9fe8fdcd167d82b6a2309204aa.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aadf284b927a43aa22a60a0e11bbafd17bb3bae409a0e49d9bc4dea33c05b287 +size 33221 diff --git a/parse/train/SkVhlh09tX/images/1ce22619be9bff7b8e64643f952c752389e9d0ea6b28d6ac7c3672feecc0b2d9.jpg b/parse/train/SkVhlh09tX/images/1ce22619be9bff7b8e64643f952c752389e9d0ea6b28d6ac7c3672feecc0b2d9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..04ec33d401b1b9b2ae292222411b76cfe3d052eb --- /dev/null +++ b/parse/train/SkVhlh09tX/images/1ce22619be9bff7b8e64643f952c752389e9d0ea6b28d6ac7c3672feecc0b2d9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d41ef5e751b4054a72634d671b1e898dfbeb564c1b270bc8710429af5c8e7ac6 +size 48361 diff --git a/parse/train/SkVhlh09tX/images/23c6ee7b02632313eccf9653c1faecd04a1b1716df2438cf798d24a44e8a4e67.jpg b/parse/train/SkVhlh09tX/images/23c6ee7b02632313eccf9653c1faecd04a1b1716df2438cf798d24a44e8a4e67.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b7677e96ad1e14eb7cbfd9cfe96f4305d2068a3b --- /dev/null +++ b/parse/train/SkVhlh09tX/images/23c6ee7b02632313eccf9653c1faecd04a1b1716df2438cf798d24a44e8a4e67.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f148af73a3d5f7db860d13a7914049afe7fe286b6a2fabecb7196de27198099d +size 8830 diff --git a/parse/train/SkVhlh09tX/images/449c272b5e8a0aab9b31a951380dccf832e0468078d57e9234e1cc6749cc8127.jpg b/parse/train/SkVhlh09tX/images/449c272b5e8a0aab9b31a951380dccf832e0468078d57e9234e1cc6749cc8127.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bd8db9f9b0ec612027049862f986acdb405ff00a --- /dev/null +++ b/parse/train/SkVhlh09tX/images/449c272b5e8a0aab9b31a951380dccf832e0468078d57e9234e1cc6749cc8127.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f7aa4b4229c00423cd331e3d09d3351c1ad76128d0625293051203e7d1f1c37 +size 32709 diff --git a/parse/train/SkVhlh09tX/images/464b1634f1028256b1ad237b5cb3a62a4bdac14298b7e56f191e573521bebf86.jpg b/parse/train/SkVhlh09tX/images/464b1634f1028256b1ad237b5cb3a62a4bdac14298b7e56f191e573521bebf86.jpg new file mode 100644 index 0000000000000000000000000000000000000000..86e1ec02aac957c2f9a928881599402252d987a8 --- /dev/null +++ b/parse/train/SkVhlh09tX/images/464b1634f1028256b1ad237b5cb3a62a4bdac14298b7e56f191e573521bebf86.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a3403d62d99cdbcc641223f1f984694f3627f76fff50c8f1f6a33748dd341c2 +size 6268 diff --git a/parse/train/SkVhlh09tX/images/4a5df697cf84116e73bc3c2ba7b36cd844e6271ee21e3f9159eec6c0b0d7663c.jpg b/parse/train/SkVhlh09tX/images/4a5df697cf84116e73bc3c2ba7b36cd844e6271ee21e3f9159eec6c0b0d7663c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..439d50899cdeec5681d0825d47ce4f41d769015c --- /dev/null +++ b/parse/train/SkVhlh09tX/images/4a5df697cf84116e73bc3c2ba7b36cd844e6271ee21e3f9159eec6c0b0d7663c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:199bd08486ed211063d3552cbb2c2172b0367e2ae435c70d4207bee25602ba29 +size 8685 diff --git a/parse/train/SkVhlh09tX/images/5077ffd6ca276bc95c9c23964c9b268792ca8c66e7d595515e146cafc44d3c24.jpg b/parse/train/SkVhlh09tX/images/5077ffd6ca276bc95c9c23964c9b268792ca8c66e7d595515e146cafc44d3c24.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bc1d83c1335643dd41a34e444aa7bd6883e72313 --- /dev/null +++ b/parse/train/SkVhlh09tX/images/5077ffd6ca276bc95c9c23964c9b268792ca8c66e7d595515e146cafc44d3c24.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29ec8c621924ae2c3ad77391f1c734346c052dfe21c1dd669f86e64c9578f83e +size 41439 diff --git a/parse/train/SkVhlh09tX/images/7705546c140d0c385860ce6b551f7e38f2b779e186a7940d567bdbf1a2f35cd9.jpg b/parse/train/SkVhlh09tX/images/7705546c140d0c385860ce6b551f7e38f2b779e186a7940d567bdbf1a2f35cd9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7aeb15ab5e70f31f09e6810260e917e4990b4d6d --- /dev/null +++ b/parse/train/SkVhlh09tX/images/7705546c140d0c385860ce6b551f7e38f2b779e186a7940d567bdbf1a2f35cd9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db6247a20e6abe34f56856b86419f0e73721073adc641b832ba0824d5d27970a +size 6662 diff --git a/parse/train/SkVhlh09tX/images/77ff323782e9e5fd0c425477a3ab972dfbe728e2a5c4c76a9726f71432b75390.jpg b/parse/train/SkVhlh09tX/images/77ff323782e9e5fd0c425477a3ab972dfbe728e2a5c4c76a9726f71432b75390.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7a265273e1b4b0d840a1449142147e1ab79f40df --- /dev/null +++ b/parse/train/SkVhlh09tX/images/77ff323782e9e5fd0c425477a3ab972dfbe728e2a5c4c76a9726f71432b75390.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c406ec270e9b3ebcf0e73e35cf5f29c1084b81274820d484ca0509e2e8f2301 +size 33379 diff --git a/parse/train/SkVhlh09tX/images/a2a2744ad9f83a0aba7443dec1f2d6953fde9e7dff0af5e369c965dabd8f9070.jpg b/parse/train/SkVhlh09tX/images/a2a2744ad9f83a0aba7443dec1f2d6953fde9e7dff0af5e369c965dabd8f9070.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b1cb31ec9ccb5c87464d04812fccff8de5e402df --- /dev/null +++ b/parse/train/SkVhlh09tX/images/a2a2744ad9f83a0aba7443dec1f2d6953fde9e7dff0af5e369c965dabd8f9070.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12074e63ccc598335cf8d73d4a98fd9a9acf2ff743607e2b1af78761e7f2d69e +size 108159 diff --git a/parse/train/SkVhlh09tX/images/a772bafb4e0921c9fc1ecd581618073af74848d1fcfe5446ef867e9f0e98af49.jpg b/parse/train/SkVhlh09tX/images/a772bafb4e0921c9fc1ecd581618073af74848d1fcfe5446ef867e9f0e98af49.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e0f9756887b96c1a421e96d51b53f9ae9fc7515d --- /dev/null +++ b/parse/train/SkVhlh09tX/images/a772bafb4e0921c9fc1ecd581618073af74848d1fcfe5446ef867e9f0e98af49.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c827ede1483c241e0e0e31322d7c0b64d7b31b3230a13ae2babf50c74f63527a +size 45993 diff --git a/parse/train/SkVhlh09tX/images/c0935bfa4ef13f7eb89c6f0d1111afa72263a783fdbd5ca4e3188662eec5c63b.jpg b/parse/train/SkVhlh09tX/images/c0935bfa4ef13f7eb89c6f0d1111afa72263a783fdbd5ca4e3188662eec5c63b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8ee1c81679a05eaa1c46718660720918047122dd --- /dev/null +++ b/parse/train/SkVhlh09tX/images/c0935bfa4ef13f7eb89c6f0d1111afa72263a783fdbd5ca4e3188662eec5c63b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd10fef91564cfa31a9999d0b284b4372edb3304fe6d5095966d57a7e03be39e +size 143289 diff --git a/parse/train/SkVhlh09tX/images/e32d355bf92cc67704985b54c492b910910adc6f316a21b8764091874495bdae.jpg b/parse/train/SkVhlh09tX/images/e32d355bf92cc67704985b54c492b910910adc6f316a21b8764091874495bdae.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4e9c1f4509c5ea90138605221450e09770a8fb0d --- /dev/null +++ b/parse/train/SkVhlh09tX/images/e32d355bf92cc67704985b54c492b910910adc6f316a21b8764091874495bdae.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:011ad14b32d59d3b867fd294806b122083d74ab13ca564815d421ff36a2bfc21 +size 45689 diff --git a/parse/train/SkVhlh09tX/images/f475991d5a01c3b151617eacd438bc8bbe92abbd2add512dbc282e4813ec0470.jpg b/parse/train/SkVhlh09tX/images/f475991d5a01c3b151617eacd438bc8bbe92abbd2add512dbc282e4813ec0470.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a8fc3207ce8bf620b75a0ed2039133d5891f90ff --- /dev/null +++ b/parse/train/SkVhlh09tX/images/f475991d5a01c3b151617eacd438bc8bbe92abbd2add512dbc282e4813ec0470.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c630933eb8c83437e039b1056cfebf1102ebf7bebea670a96e41bf9cea5b616a +size 6573 diff --git a/parse/train/Sy2fzU9gl/Sy2fzU9gl.md b/parse/train/Sy2fzU9gl/Sy2fzU9gl.md new file mode 100644 index 0000000000000000000000000000000000000000..588f458bca1c837430c32304f8b8a8a5be76d6b4 --- /dev/null +++ b/parse/train/Sy2fzU9gl/Sy2fzU9gl.md @@ -0,0 +1,320 @@ +# $\beta$ -VAE: LEARNING BASIC VISUAL CONCEPTS WITH A CONSTRAINED VARIATIONAL FRAMEWORK + +Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, Alexander Lerchner + +Google DeepMind {irinah,lmatthey,arkap,cpburgess,glorotx, botvinick,shakir,lerchner}@google.com + +# ABSTRACT + +Learning an interpretable factorised representation of the independent data generative factors of the world without supervision is an important precursor for the development of artificial intelligence that is able to learn and reason in the same way that humans do. We introduce $\beta$ -VAE, a new state-of-the-art framework for automated discovery of interpretable factorised latent representations from raw image data in a completely unsupervised manner. Our approach is a modification of the variational autoencoder (VAE) framework. We introduce an adjustable hyperparameter $\beta$ that balances latent channel capacity and independence constraints with reconstruction accuracy. We demonstrate that $\beta$ -VAE with appropriately tuned $\beta > 1$ qualitatively outperforms VAE $\begin{array} { r } { \beta = 1 , } \end{array}$ ), as well as state of the art unsupervised (InfoGAN) and semi-supervised (DC-IGN) approaches to disentangled factor learning on a variety of datasets (celebA, faces and chairs). Furthermore, we devise a protocol to quantitatively compare the degree of disentanglement learnt by different models, and show that our approach also significantly outperforms all baselines quantitatively. Unlike InfoGAN, $\beta$ -VAE is stable to train, makes few assumptions about the data and relies on tuning a single hyperparameter $\beta$ , which can be directly optimised through a hyperparameter search using weakly labelled data or through heuristic visual inspection for purely unsupervised data. + +# 1 INTRODUCTION + +The difficulty of learning a task for a given machine learning approach can vary significantly depending on the choice of the data representation. Having a representation that is well suited to the particular task and data domain can significantly improve the learning success and robustness of the chosen model (Bengio et al., 2013). It has been suggested that learning a disentangled representation of the generative factors in the data can be useful for a large variety of tasks and domains (Bengio et al., 2013; Ridgeway, 2016). A disentangled representation can be defined as one where single latent units are sensitive to changes in single generative factors, while being relatively invariant to changes in other factors (Bengio et al., 2013). For example, a model trained on a dataset of 3D objects might learn independent latent units sensitive to single independent data generative factors, such as object identity, position, scale, lighting or colour, thus acting as an inverse graphics model (Kulkarni et al., 2015). In a disentangled representation, knowledge about one factor can generalise to novel configurations of other factors. According to Lake et al. (2016), disentangled representations could boost the performance of state-of-the-art AI approaches in situations where they still struggle but where humans excel. Such scenarios include those which require knowledge transfer, where faster learning is achieved by reusing learnt representations for numerous tasks; zero-shot inference, where reasoning about new data is enabled by recombining previously learnt factors; or novelty detection. + +Unsupervised learning of a disentangled posterior distribution over the underlying generative factors of sensory data is a major challenge in AI research (Bengio et al., 2013; Lake et al., 2016). Most previous attempts required a priori knowledge of the number and/or nature of the data generative factors (Hinton et al., 2011; Rippel & Adams, 2013; Reed et al., 2014; Zhu et al., 2014; Yang et al., 2015; Goroshin et al., 2015; Kulkarni et al., 2015; Cheung et al., 2015; Whitney et al., 2016; Karaletsos et al., 2016). This is not always feasible in the real world, where the newly initialised learner may be exposed to complex data where no a priori knowledge of the generative factors exists, and little to no supervision for discovering the factors is available. Until recently purely unsupervised approaches to disentangled factor learning have not scaled well (Schmidhuber, 1992; Desjardins et al., 2012; Tang et al., 2013; Cohen & Welling, 2014; 2015). + +![](images/6bae38335f4edd4fe6fe641c6af928cf5435a2ee5710353b8d5ef567cae439cf.jpg) +Figure 1: Manipulating latent variables on celebA: Qualitative results comparing disentangling performance of $\beta$ -VAE $\beta = 2 5 0 $ ), VAE (Kingma & Welling, 2014) $( \beta = 1 )$ ) and InfoGAN (Chen et al., 2016). In all figures of latent code traversal each block corresponds to the traversal of a single latent variable while keeping others fixed to either their inferred ( $\beta$ -VAE, VAE and DC-IGN where applicable) or sampled (InfoGAN) values. Each row represents a different seed image used to infer the latent values in the VAE-based models, or a random sample of the noise variables in InfoGAN. $\beta$ -VAE and VAE traversal is over the $I { \cdot } 3 , \ 3 I$ range. InfoGAN traversal is over ten dimensional categorical latent variables. Only $\beta$ -VAE and InfoGAN learnt to disentangle factors like azimuth (a), emotion (b) and hair style (c), whereas VAE learnt an entangled representation (e.g. azimuth is entangled with emotion, presence of glasses and gender). InfoGAN images adapted from Chen et al. (2016). Reprinted with permission. + +Recently a scalable unsupervised approach for disentangled factor learning has been developed, called InfoGAN (Chen et al., 2016). InfoGAN extends the generative adversarial network (GAN) (Goodfellow et al., 2014) framework to additionally maximise the mutual information between a subset of the generating noise variables and the output of a recognition network. It has been reported to be capable of discovering at least a subset of data generative factors and of learning a disentangled representation of these factors. The reliance of InfoGAN on the GAN framework, however, comes at the cost of training instability and reduced sample diversity. Furthermore, InfoGAN requires some a priori knowledge of the data, since its performance is sensitive to the choice of the prior distribution and the number of the regularised noise variables. InfoGAN also lacks a principled inference network (although the recognition network can be used as one). The ability to infer the posterior latent distribution from sensory input is important when using the unsupervised model in transfer learning or zero-shot inference scenarios. Hence, while InfoGAN is an important step in the right direction, we believe that further improvements are necessary to achieve a principled way of using unsupervised learning for developing more human-like learning and reasoning in algorithms as described by Lake et al. (2016). + +Finally, there is currently no general method for quantifying the degree of learnt disentanglement. Therefore there is no way to quantitatively compare the degree of disentanglement achieved by different models or when optimising the hyperparameters of a single model. + +![](images/504fafaff72a9120d6f80158ae647468090601903bd9e60571b38eabc767fa49.jpg) +Figure 2: Manipulating latent variables on 3D chairs: Qualitative results comparing disentangling performance of $\beta$ -VAE $\beta = 5$ ), VAE (Kingma & Welling, 2014) $\mathcal { \beta } = 1 \mathrm { \check { \beta } } ,$ ), InfoGAN (Chen et al., 2016) and DC-IGN (Kulkarni et al., 2015). InfoGAN traversal is over the [-1, 1] range. VAE always learns an entangled representation (e.g. chair width is entangled with azimuth and leg style (b)). All models apart from VAE learnt to disentangle the labelled data generative factor, azimuth (a). InfoGAN and $\beta$ -VAE were also able to discover unlabelled factors in the dataset, such as chair width (b). Only $\beta$ -VAE, however, learnt about the unlabelled factor of chair leg style (c). InfoGAN and DC-IGN images adapted from Chen et al. (2016) and Kulkarni et al. (2015), respectively. Reprinted with permission. + +In this paper we attempt to address these issues. We propose $\beta$ -VAE, a deep unsupervised generative approach for disentangled factor learning that can automatically discover the independent latent factors of variation in unsupervised data. Our approach is based on the variational autoencoder (VAE) framework (Kingma & Welling, 2014; Rezende et al., 2014), which brings scalability and training stability. While the original VAE work has been shown to achieve limited disentangling performance on simple datasets, such as FreyFaces or MNIST (Kingma & Welling, 2014), disentangling performance does not scale to more complex datasets (e.g. Aubry et al., 2014; Paysan et al., 2009; Liu et al., 2015), prompting the development of more elaborate semi-supervised VAE-based approaches for learning disentangled factors (e.g. Kulkarni et al., 2015; Karaletsos et al., 2016). + +We propose augmenting the original VAE framework with a single hyperparameter $\beta$ that modulates the learning constraints applied to the model. These constraints impose a limit on the capacity of the latent information channel and control the emphasis on learning statistically independent latent factors. $\beta$ -VAE with $\beta = 1$ corresponds to the original VAE framework (Kingma & Welling, 2014; Rezende et al., 2014). With $\beta > 1$ the model is pushed to learn a more efficient latent representation of the data, which is disentangled if the data contains at least some underlying factors of variation that are independent. We show that this simple modification allows $\beta$ -VAE to significantly improve the degree of disentanglement in learnt latent representations compared to the unmodified VAE framework (Kingma & Welling, 2014; Rezende et al., 2014). Furthermore, we show that $\beta$ -VAE achieves state of the art disentangling performance against both the best unsupervised (InfoGAN: Chen et al., 2016) and semi-supervised (DC-IGN: Kulkarni et al., 2015) approaches for disentangled factor learning on a number of benchmark datasets, such as CelebA (Liu et al., 2015), chairs (Aubry et al., 2014) and faces (Paysan et al., 2009) using qualitative evaluation. Finally, to help quantify the differences, we develop a new measure of disentanglement and show that $\beta$ -VAE significantly outperforms all our baselines on this measure (ICA, PCA, VAE Kingma & Ba (2014), DC-IGN Kulkarni et al. (2015), and InfoGAN Chen et al. (2016)). + +Our main contributions are the following: 1) we propose $\beta$ -VAE, a new unsupervised approach for learning disentangled representations of independent visual data generative factors; 2) we devise a protocol to quantitatively compare the degree of disentanglement learnt by different models; 3) we demonstrate both qualitatively and quantitatively that our $\beta$ -VAE approach achieves state-of-the-art disentanglement performance compared to various baselines on a variety of complex datasets. + +![](images/4afdd3550ccd39290f23e0e238e9a039d6d3016ff06f8dad0ca88f63a1ea1b20.jpg) +Figure 3: Manipulating latent variables on 3D faces: Qualitative results comparing disentangling performance of $\beta$ -VAE $\beta = 2 0$ ), VAE (Kingma & Welling, 2014) $\begin{array} { r } { \beta = 1 , } \end{array}$ ), InfoGAN (Chen et al., 2016) and DC-IGN (Kulkarni et al., 2015). InfoGAN traversal is over the [-1, 1] range. All models learnt to disentangle lighting (b) and elevation (c). DC-IGN and VAE struggled to continuously interpolate between different azimuth angles (a), unlike $\beta$ -VAE, which additionally learnt to encode a wider range of azimuth angles than other models. InfoGAN and DC-IGN images adapted from Chen et al. (2016) and Kulkarni et al. (2015), respectively. Reprinted with permission. + +![](images/40660321d43bd855a3ea6abb7c3408b840648a9ba4a972704a38a38e1104972e.jpg) +Figure 4: Latent factors learnt by $\beta$ -VAE on celebA: traversal of individual latents demonstrates that $\beta$ -VAE discovered in an unsupervised manner factors that encode skin colour, transition from an elderly male to younger female, and image saturation. + +# 2 $\beta$ -VAE FRAMEWORK DERIVATION + +Let $\mathcal { D } = \{ X , V , W \}$ be the set that consists of images $\mathbf { x } \in \mathbb { R } ^ { N }$ and two sets of ground truth data generative factors: conditionally independent factors $\mathbf { \bar { v } } \in \mathbb { R } ^ { K }$ , where $\begin{array} { r } { \log p ( \mathbf { v } | \mathbf { x } ) = \sum _ { k } \log p ( v _ { k } | \mathbf { x } ) } \end{array}$ ; and conditionally dependent factors $\mathbf { w } \in \mathbb { R } ^ { H }$ . We assume that the images $\mathbf { x }$ are generated by the true world simulator using the corresponding ground truth data generative factors: $p ( \mathbf { x } | \mathbf { v } , \dot { \mathbf { w } } ) =$ $\mathbf { S i m } ( \mathbf { v } , \mathbf { w } )$ . + +We want to develop an unsupervised deep generative model that, using samples from $X$ only, can learn the joint distribution of the data $\mathbf { x }$ and a set of generative latent factors $\mathbf { z }$ ${ \bf \Psi } ( { \bf z } \in \mathbb { R } ^ { M }$ , where $M \geq K$ ) such that $\mathbf { z }$ can generate the observed data $\mathbf { x }$ ; that is, $p ( \mathbf { x } | \mathbf { z } ) \approx p ( \mathbf { x } | \mathbf { v } , \mathbf { w } ) = \mathbf { S i m } ( \mathbf { v } , \mathbf { w } )$ . Thus a suitable objective is to maximise the marginal (log-)likelihood of the observed data $\mathbf { x }$ in expectation over the whole distribution of latent factors $\mathbf { z }$ : + +$$ +\operatorname* { m a x } _ { \theta } \mathbb { E } _ { p _ { \theta } ( \mathbf { z } ) } [ p _ { \theta } ( \mathbf { x } | \mathbf { z } ) ] +$$ + +For a given observation $\mathbf { x }$ , we describe the inferred posterior configurations of the latent factors $\mathbf { z }$ by a probability distribution $q _ { \phi } ( { \bf z } | { \bf x } )$ . Our aim is to ensure that the inferred latent factors $q _ { \phi } ( { \bf z } | { \bf x } )$ capture the generative factors $\mathbf { v }$ in a disentangled manner. The conditionally dependent data generative factors w can remain entangled in a separate subset of $\mathbf { z }$ that is not used for representing $\mathbf { v }$ . In order to encourage this disentangling property in the inferred $q _ { \phi } ( { \bf z } | { \bf x } )$ , we introduce a constraint over it by trying to match it to a prior $p ( \mathbf { z } )$ that can both control the capacity of the latent information bottleneck, and embodies the desiderata of statistical independence mentioned above. This can be achieved if we set the prior to be an isotropic unit Gaussian $( p ( \mathbf { z } ) = \mathcal { N } ( \mathbf { 0 } , I ) )$ , hence arriving at the constrained optimisation problem in Eq. 2, where $\epsilon$ specifies the strength of the applied constraint. + +$$ +\operatorname* { m a x } _ { \phi , \theta } \mathbb { E } _ { x \sim \mathbf { D } } \left[ \mathbb { E } _ { q _ { \phi } ( \mathbf { z } | \mathbf { x } ) } [ \log p _ { \theta } ( \mathbf { x } | \mathbf { z } ) ] \right] \quad \mathrm { ~ s u b j e c t ~ t o ~ } D _ { K L } ( q _ { \phi } ( \mathbf { z } | \mathbf { x } ) | | p ( \mathbf { z } ) ) < \epsilon +$$ + +Re-writing Eq. 2 as a Lagrangian under the KKT conditions (Kuhn & Tucker, 1951; Karush, 1939), we obtain: + +$$ +\mathcal { F } ( \boldsymbol { \theta } , \boldsymbol { \phi } , \boldsymbol { \beta } ; \mathbf { x } , \mathbf { z } ) = \mathbb { E } _ { q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) } [ \log p _ { \boldsymbol { \theta } } ( \mathbf { x } | \mathbf { z } ) ] - \beta \left( D _ { K L } ( q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) | | p ( \mathbf { z } ) ) - \epsilon \right) +$$ + +where the KKT multiplier $\beta$ is the regularisation coefficient that constrains the capacity of the latent information channel $\mathbf { z }$ and puts implicit independence pressure on the learnt posterior due to the isotropic nature of the Gaussian prior $p ( \mathbf { z } )$ . Since $\beta , \epsilon \geq 0$ according to the complementary slackness KKT condition, Eq. 3 can be re-written to arrive at the $\beta$ -VAE formulation - as the familiar variational free energy objective function as described by Jordan et al. (1999), but with the addition of the $\beta$ coefficient: + +$$ +\mathcal { F } ( \boldsymbol { \theta } , \boldsymbol { \phi } , \boldsymbol { \beta } ; \mathbf { x } , \mathbf { z } ) \geq \mathcal { L } ( \boldsymbol { \theta } , \boldsymbol { \phi } ; \mathbf { x } , \mathbf { z } , \boldsymbol { \beta } ) = \mathbb { E } _ { q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) } [ \log p _ { \boldsymbol { \theta } } ( \mathbf { x } | \mathbf { z } ) ] - \beta D _ { K L } ( q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) | | p ( \mathbf { z } ) ) +$$ + +Varying $\beta$ changes the degree of applied learning pressure during training, thus encouraging different learnt representations. $\beta$ -VAE where $\beta = 1$ corresponds to the original VAE formulation of (Kingma & Welling, 2014). We postulate that in order to learn disentangled representations of the conditionally independent data generative factors $\mathbf { v }$ , it is important to set $\beta > 1$ , thus putting a stronger constraint on the latent bottleneck than in the original VAE formulation of Kingma & Welling (2014). These constraints limit the capacity of $\mathbf { z }$ , which, combined with the pressure to maximise the log likelihood of the training data $\mathbf { x }$ under the model, should encourage the model to learn the most efficient representation of the data. Since the data $\mathbf { x }$ is generated using at least some conditionally independent ground truth factors $\mathbf { v }$ , and the $D _ { K L }$ term of the $\beta$ -VAE objective function encourages conditional independence in $q _ { \phi } ( \mathbf { z } | \mathbf { x } )$ , we hypothesise that higher values of $\beta$ should encourage learning a disentangled representation of $\mathbf { v }$ . The extra pressures coming from high $\beta$ values, however, may create a trade-off between reconstruction fidelity and the quality of disentanglement within the learnt latent representations. Disentangled representations emerge when the right balance is found between information preservation (reconstruction cost as regularisation) and latent channel capacity restriction $( \beta > 1 )$ ). The latter can lead to poorer reconstructions due to the loss of high frequency details when passing through a constrained latent bottleneck. Hence, the log likelihood of the data under the learnt model is a poor metric for evaluating disentangling in $\beta$ -VAEs. Instead we propose a quantitative metric that directly measures the degree of learnt disentanglement in the latent representation. + +Since our proposed hyperparameter $\beta$ directly affects the degree of learnt disentanglement, we would like to estimate the optimal $\beta$ for learning a disentangled latent representation directly. However, it is not possible to do so. This is because the optimal $\beta$ will depend on the value of $\epsilon$ in Eq.2. Different datasets and different model architectures will require different optimal values of $\epsilon$ . However, when optimising $\beta$ in Eq. 4, we are indirectly also optimising $\epsilon$ for the best disentanglement (see Sec.A.7 for details), and while we can not learn the optimal value of $\beta$ directly, we can instead estimate it using either our proposed disentanglement metric (see Sec. 3) or through visual inspection heuristics. + +![](images/b2c9c3518517d6763a2eaa3691033afb73151e60d49d36e7012763a6045855bb.jpg) +Figure 5: Schematic of the proposed disentanglement metric: over a batch of $L$ samples, each pair of images has a fixed value for one target generative factor $y$ (here $y = s c a l e )$ and differs on all others. A linear classifier is then trained to identify the target factor using the average pairwise difference $\mathbf { z } _ { \mathrm { d i f f } } ^ { b }$ in the latent space over $L$ samples. + +# 3 DISENTANGLEMENT METRIC + +It is important to be able to quantify the level of disentanglement achieved by different models. Designing a metric for this, however, is not straightforward. We begin by defining the properties that we expect a disentangled representation to have. Then we describe our proposed solution for quantifying the presence of such properties in a learnt representation. + +As stated above, we assume that the data is generated by a ground truth simulation process which uses a number of data generative factors, some of which are conditionally independent, and we also assume that they are interpretable. For example, the simulator might sample independent factors corresponding to object shape, colour and size to generate an image of a small green apple. Because of the independence property, the simulator can also generate small red apples or big green apples. A representation of the data that is disentangled with respect to these generative factors, i.e. which encodes them in separate latents, would enable robust classification even using very simple linear classifiers (hence providing interpretability). For example, a classifier that learns a decision boundary that relies on object shape would perform as well when other data generative factors, such as size or colour, are varied. + +Note that a representation consisting of independent latents is not necessarily disentangled, according to our desiderata. Independence can readily be achieved by a variety of approaches (such as PCA or ICA) that learn to project the data onto independent bases. Representations learnt by such approaches do not in general align with the data generative factors and hence may lack interpretability. For this reason, a simple cross-correlation calculation between the inferred latents would not suffice as a disentanglement metric. + +Our proposed disentangling metric, therefore, measures both the independence and interpretability (due to the use of a simple classifier) of the inferred latents. To apply our metric, we run inference on a number of images that are generated by fixing the value of one data generative factor while randomly sampling all others. If the independence and interpretability properties hold for the inferred representations, there will be less variance in the inferred latents that correspond to the fixed generative factor. We use a low capacity linear classifier to identify this factor and report the accuracy value as the final disentanglement metric score. Smaller variance in the latents corresponding to the target factor will make the job of this classifier easier, resulting in a higher score under the metric. See Fig. 5 for a representation of the full process. + +More formally, we start from a dataset $\mathcal { D } = \{ X , V , W \}$ as described in Sec. 2, assumed to contain a balanced distribution of ground truth factors $( \mathbf { v } , \mathbf { w } )$ , where images data points are obtained using a ground truth simulator process $\mathbf { x } \sim \mathbf { S i m } ( \mathbf { v } , \mathbf { \dot { w } } )$ . We also assume we are given labels identifying a subset of the independent data generative factors $\mathbf { v } \in V$ for at least some instances. + +We then construct a batch of $B$ vectors $\mathbf { z } _ { \mathrm { d i f f } } ^ { b }$ , to be fed as inputs to a linear classifier as follows: + +1. Choose a factor $y \sim U n i f [ 1 . . . K ]$ (e.g. $y = s c a l e$ in Fig. 5). + +2. For a batch of $L$ samples: + +(a) Sample two sets of latent representations, $\mathbf { v } _ { 1 , l }$ and $\mathbf { v } _ { 2 , l }$ , enforcing $\begin{array} { r l } { [ \mathbf { v } _ { 1 , l } ] _ { k } } & { { } = } \end{array}$ $\left[ \mathbf { v } _ { 2 , l } \right] _ { k }$ if $k = y$ (so that the value of factor $k = y$ is kept fixed). +(b) Simulate image $\mathbf { x } _ { 1 , l } \sim \mathbf { S i m } ( \mathbf { v } _ { 1 , 1 } )$ , then infer ${ \bf z } _ { 1 , l } = \mu ( { \bf x } _ { 1 , l } )$ , using the encoder $q ( \mathbf { z } | \mathbf { x } ) \sim N \left( \mu ( \mathbf { x } ) , \sigma ( \mathbf { x } ) \right)$ . Repeat the process for $\mathbf { v } _ { 2 , l }$ . +(c) Compute the difference $\mathbf { z } _ { \mathrm { d i f f } } ^ { l } = | \mathbf { z } _ { 1 , l } - \mathbf { z } _ { 2 , l } |$ , the absolute linear difference between the inferred latent representations. + +3. Use the average $\begin{array} { r } { \mathbf { z } _ { \mathrm { d i f f } } ^ { b } = \frac { 1 } { L } \sum _ { l = 1 } ^ { L } \mathbf { z } _ { \mathrm { d i f f } } ^ { l } } \end{array}$ to predict $p ( \boldsymbol { y } | \mathbf { z } _ { \mathrm { d i f f } } ^ { b } )$ (again, $y = s c a l e$ in Fig. 5) and report the accuracy of this predictor as disentangement metric score. + +The classifier’s goal is to predict the index of the generative factor that was kept fixed for a given $\mathbf { z } _ { \mathrm { d i f f } } ^ { b }$ . The accuracy of this classifier over multiple batches is used as our disentanglement metric score. We choose a linear classifier with low VC-dimension in order to ensure it has no capacity to perform nonlinear disentangling by itself. We take differences of two inferred latent vectors to reduce the variance in the inputs to ensures that on average educe the conditional dependence on the inputs . See Equations 5 in Appendix A.4 for more d $\mathbf { x }$ . Thisails of $\left[ \mathbf { z } _ { \mathrm { d i f f } } ^ { b } \right] _ { y } < \left[ \mathbf { z } _ { \mathrm { d i f f } } ^ { b } \right] _ { \{ \backslash y \} }$ +the process. + +# 4 EXPERIMENTS + +In this section we first qualitatively demonstrate that our proposed $\beta$ -VAE framework consistently discovers more latent factors and disentangles them in a cleaner fashion that either unmodified VAE (Kingma & Welling, 2014) or state of the art unsupervised (InfoGAN: Chen et al., 2016) and semisupervised (DC-IGN: Kulkarni et al., 2015) solutions for disentangled factor learning on a variety of benchmarks. We then quantify and characterise the differences in disentangled factor learning between our $\beta$ -VAE framework and a variety of benchmarks using our proposed new disentangling metric. + +# 4.1 QUALITATIVE BENCHMARKS + +We trained $\beta$ -VAE (see Tbl. 1 for architecture details) on a variety of datasets commonly used to evaluate disentangling performance of models: celebA (Liu et al., 2015), chairs (Aubry et al., 2014) and faces (Paysan et al., 2009). Figures 1-3 provide a qualitative comparison of the disentangling performance of $\beta$ -VAE, VAE $\beta = 1$ ) (Kingma & Welling, 2014), InfoGAN (Chen et al., 2016) and DC-IGN (Kulkarni et al., 2015) as appropriate. + +It can be seen that across all datasets $\beta$ -VAE is able to automatically discover and learn to disentangle all of the factors learnt by the semi-supervised DC-IGN (Kulkarni et al., 2015): azimuth (Fig. 3a, Fig. 2a), lighting and elevation (Fig. 3b,c)). Often it acts as a more convincing inverse graphics network than DC-IGN (e.g. Fig. 3a) or InfoGAN (e.g. Fig. 2a, Fig. 1a-c or Fig. 3a). Furthermore, unlike DC-IGN, $\beta$ -VAE requires no supervision and hence can learn about extra unlabelled data generative factors that DC-IGN can not learn by design, such as chair width or leg style (Fig. 2b,c). The unsupervised InfoGAN (Chen et al., 2016) approach shares this quality with $\beta$ -VAE, and the two frameworks tend to discover overlapping, but not necessarily identical sets of data generative factors. For example, both $\beta$ -VAE and InfoGAN (but not DC-IGN) learn about the width of chairs (Fig. 2b). Only $\beta$ -VAE, however, learns about the chair leg style (Fig. 2c). It is interesting to note how $\beta$ -VAE is able to generate an armchair with a round office chair base, even though such armchairs do not exist in the dataset (or, perhaps, reality). Furthermore, only $\beta$ -VAE is able to discover all three factors of variation (chair azimuth, width and leg style) within a single model, while InfoGAN learns to allocate its continuous latent variable to either azimuth or width. InfoGAN sometimes discovers factors that $\beta$ -VAE does not precisely disentangle, such as the presence of sunglasses in celebA. $\beta$ -VAE does, however, discover numerous extra factors such as skin colour, image saturation, and age/gender that are not reported in the InfoGAN paper (Chen et al., 2016) (Fig. 4). Furthermore, $\beta$ -VAE latents tend to learn a smooth continuous transformation over a wider range of factor values than InfoGAN (e.g. rotation over a wider range of angles as shown in Figs. 1-3a). + +Overall $\beta$ -VAE tends to consistently and robustly discover more latent factors and learn cleaner disentangled representations of them than either InfoGAN or DC-IGN. This holds even on such challenging datasets as celebA. Furthermore, unlike InfoGAN and DC-IGN, $\beta$ -VAE requires no design decisions or assumptions about the data, and is very stable to train. + +When compared to the unmodified VAE baseline ( $\beta = 1$ ) $\beta$ -VAE consistently learns significantly more disentangled latent representations. For example, when learning about chairs, VAE entangles chair width with leg style (Fig. 2b). When learning about celebA, VAE entangles azimuth with emotion and gender (Fig. 1a); emotion with hair style, skin colour and identity (Fig. 1b); while the VAE fringe latent also codes for baldness and head size (Fig. 1c). Although VAE performs relatively well on the faces dataset, it still struggles to learn a clean representation of azimuth (Fig. 3a). This, however, suggests that a continuum of disentanglement quality exists, and it can be traversed by varying $\beta$ within the $\beta$ -VAE framework. While increasing $\beta$ often leads to better disentanglement, it may come at the cost of blurrier reconstructions and losing representations for some factors, particularly those that correspond to only minor changes in pixel space. + +# 4.2 QUANTITATIVE BENCHMARKS + +In order to quantitatively compare the disentangling performance of $\beta$ -VAE against various baselines, we created a synthetic dataset of 737,280 binary 2D shapes (heart, oval and square) generated from the Cartesian product of the shape and four independent generative factors $v _ { k }$ defined in vector graphics: position X (32 values), position Y (32 values), scale (6 values) and rotation (40 values over the $2 \pi$ range). To ensure smooth affine object transforms, each two subsequent values for each factor $v _ { k }$ were chosen to ensure minimal differences in pixel space given 64x64 pixel image resolution. This dataset was chosen because it contains no confounding factors apart from its five independent data generative factors (identity, position X, position Y, scale and rotation). This gives us knowledge of the ground truth for comparing the disentangling performance of different models in an objective manner. + +We used our proposed disentanglement metric (see Sec. 3) to quantitatively compare the ability of $\beta$ -VAE to automatically discover and learn a disentangled representation of the data generative factors of the synthetic dataset of 2D shapes described above with that of a number of benchmarks (see Tbl. 1 in Appendix for model architecture details). The table in Fig. 6 (left) reports the classification accuracy of the disentanglement metric for 5,000 test samples. It can be seen that $\beta$ -VAE $\beta = 4 ,$ ) significantly outperforms all baselines, such as an untrained VAE and the original VAE formulation of Kingma & Welling (2014) $( \beta = 1 )$ ) with the same architecture as $\beta$ -VAE, the top ten PCA or ICA components of the data (see Sec. A.3 for details), or when using the raw pixels directly. $\beta$ -VAE also does better than InfoGAN. Remarkably, $\beta$ -VAE performs on the same level as DC-IGN despite the latter being semi-supervised and the former wholly unsupervised. Furthermore, $\beta$ -VAE achieved similar classification accuracy as the ground truth vectors used for data generation, thus suggesting that it was able to learn a very good disentangled representation of the data generative factors. + +We also examined qualitatively the representations learnt by $\beta$ -VAE, VAE, InfoGAN and DC-IGN on the synthetic dataset of 2D shapes. Fig. 7A demonstrates that after training, $\beta$ -VAE with $\beta = 4$ learnt a good (while not perfect) disentangled representation of the data generative factors, and its decoder learnt to act as a rendering engine. Its performance was comparative to that of DCIGN (Fig. 7C), with the difference that DC-IGN required a priori knowledge about the quantity of the data generative factors, while $\beta$ -VAE was able to discover them in an unsupervised manner. The most informative latent units $z _ { m }$ of $\beta$ -VAE have the highest KL divergence from the unit Gaussian prior $( p ( z ) = \mathcal { N } ( 0 , I ) )$ , while the uninformative latents have KL divergence close to zero. Fig. 7A demonstrates the selectivity of each latent $z _ { m }$ to the independent data generating factors: $z _ { m } ^ { \mu } = f ( v _ { k } ) \forall v _ { k } \in \left\{ v _ { p o s i t i o n X } , v _ { p o s i t i o n Y } , v _ { s c a l e } , v _ { r o t a t i o n } \right\}$ (top three rows), where $z _ { m } ^ { \mu }$ is the learnt Gaussian mean of latent unit $z _ { m }$ . The effect of traversing each latent $z _ { m }$ on the resulting reconstructions is shown in the bottom five rows of Fig. 7A. The latents $z _ { 6 }$ and $z _ { 2 }$ learnt to encode $\mathrm { X }$ and $\mathrm { Y }$ coordinates of the objects respectively; unit $z _ { 1 }$ learnt to encode scale; and units $z _ { 5 }$ and $z _ { 7 }$ learnt to encode rotation. The frequency of oscillations in each rotational latent corresponds to the rotational symmetry of the corresponding object $2 \pi$ for heart, $\pi$ for oval and $\pi / 2$ for square). Furthermore, the two rotational latents seem to encode cos and sin rotational coordinates, while the positional latents align with the Cartesian axes. While such alignment with intuitive factors for humans is not guaranteed, empirically we found it to be very common. Fig. 7B demonstrates that the unmodified + +
ModelDisentanglementmetric score
Ground truthRaw pixelsPCAICA100%45.75 ± 0.8%84.9 ±0.4%42.03 ± 10.6%
DC-IGNInfoGAN99.3 ± 0.1%73.5 ± 0.9%
VAE untrainedVAEβ-VAE44.14 ± 2.5%61.58 ± 0.5%
61.58 ± 0.5%99.23±0.1%
+ +![](images/b161f1f7e0f90af0103ab47c8dc24899548b222d5a87984f5f762ba8e18c80a4.jpg) +Figure 6: Disentanglement metric classification accuracy for 2D shapes dataset. Left: Accuracy for different models and training regimes Right: Positive correlation is present between the size of $\mathbf { z }$ and the optimal normalised values of $\beta$ for disentangled factor learning for a fixed $\beta$ -VAE architecture. $\beta$ values are normalised by latent $\mathbf { z }$ size $m$ and input $\mathbf { x }$ size $n$ . Note that $\beta$ values are not uniformly sampled. Orange approximately corresponds to unnormalised $\beta = 1$ . Good reconstructions are associated with entangled representations (lower disentanglement scores). Disentangled representations (high disentanglement scores) often result in blurry reconstructions. + +VAE baseline $\begin{array} { r } { \beta = 1 } \end{array}$ ) is not able to disentangle generative factors in the data as well as $\beta$ -VAE with appropriate learning pressures. Instead each latent $\mathbf { z }$ (apart from $\mathbf { z } _ { 9 }$ , which learnt rotation) encodes at least two data generative factors. InfoGAN also achieved a degree of disentangling (see Fig. 7D), particularly for positional factors. However, despite our best efforts to train InfoGAN, we were not able to achieve the same degree of disentangling in other factors, such as rotation, scale and shape. We also found its ability to generate the different shapes in the dataset to be inaccurate and unstable during training, possibly due to reported limitations of the GAN framework, which can struggle to learn the full data distribution and instead will often learn a small subset of its modes (Salimans et al., 2016; Zhao et al., 2016). + +Understanding the effects of $\beta$ We hypothesised that constrained optimisation is important for enabling deep unsupervised models to learn disentangled representations of the independent data generative factors (Sec. 2). In the $\beta$ -VAE framework this corresponds to tuning the $\beta$ coefficient. One way to view $\beta$ is as a mixing coefficient (see Sec. A.6 for a derivation) for balancing the magnitudes of gradients from the reconstruction and the prior-matching components of the VAE lower bound formulation in Eq. 4 during training. In this context it makes sense to normalise $\beta$ by latent $\mathbf { z }$ size $m$ and input $\mathbf { x }$ size $n$ in order to compare its different values across different latent layer sizes and different datasets $\begin{array} { r } { ( \beta _ { n o r m } = \frac { \beta M } { N } . } \end{array}$ ). We found that larger latent $\mathbf { z }$ layer sizes $m$ require higher constraint pressures (higher $\beta$ values), see Fig. 6 (Right). Furthermore, the relationship of $\beta$ for a given $m$ is characterised by an inverted U curve. When $\beta$ is too low or too high the model learns an entangled latent representation due to either too much or too little capacity in the latent $\mathbf { z }$ bottleneck. We find that in general $\beta > 1$ is necessary to achieve good disentanglement. However if $\beta$ is too high and the resulting capacity of the latent channel is lower than the number of data generative factors, then the learnt representation necessarily has to be entangled (as a low-rank projection of the true data generative factors will compress them in a non-factorial way to still capture the full data distribution well). We also note that VAE reconstruction quality is a poor indicator of learnt disentanglement. Good disentangled representations often lead to blurry reconstructions due to the restricted capacity of the latent information channel $\mathbf { z }$ , while entangled representations often result in the sharpest reconstructions. We therefore suggest that one should not necessarily strive for perfect reconstructions when using $\beta$ -VAEs as unsupervised feature learners - though it is often possible to find the right $\beta$ -VAE architecture and the right value of $\beta$ to have both well disentangled latent representations and good reconstructions. + +We proposed a principled way of choosing $\beta$ for datasets with at least weak label information. If label information exists for at least a small subset of the independent data generative factors of variation, one can apply the disentanglement metric described in Sec. 3 to approximate the level of learnt disentanglement for various $\beta$ choices during a hyperparameter sweep. When such labelled information is not available, the optimal value of $\beta$ can be found through visual inspection of what + +![](images/289e67ba96015590e7c3119eba2427223eff790ce99ac76938ab55156ba3a1fd.jpg) + +Figure 7: A: Representations learnt by a $\beta$ -VAE $\beta = 4$ ). Each column represents a latent $z _ { i }$ , ordered according to the learnt Gaussian variance (last row). Row 1 (position) shows the mean activation (red represents high values) of each latent $z _ { i }$ as a function of all $3 2 \mathrm { x } 3 2$ locations averaged across objects, rotations and scales. Row 2 and 3 show the mean activation of each unit $z _ { i }$ as a function of scale (respectively rotation), averaged across rotations and positions (respectively scales and positions). Square is red, oval is green and heart is blue. Rows 4-8 (second group) show reconstructions resulting from the traversal of each latent $z _ { i }$ over three standard deviations around the unit Gaussian prior mean while keeping the remaining 9/10 latent units fixed to the values obtained by running inference on an image from the dataset. B: Similar analysis for VAE $\begin{array} { r } { \beta = 1 ) } \end{array}$ . C: Similar analysis for DCIGN, clamping a single latent each for scale, positions, orientation and 5 for shape. D: Similar analysis for InfoGAN, using 5 continuous latents regularized using the mutual information cost, and 5 additional unconstrained noise latents (not shown). + +effect the traversal of each single latent unit $z _ { m }$ has on the generated images $\mathbf { \tau } ( \mathbf { x } | \mathbf { z } )$ in pixel space (as shown in Fig. 7 rows 4-8). For the 2D shapes dataset, we have found that the optimal values of $\beta$ as determined by visual inspection match closely the optimal values as determined by the disentanglement metric. + +# 5 CONCLUSION + +In this paper we have reformulated the standard VAE framework (Kingma & Welling, 2014; Rezende et al., 2014) as a constrained optimisation problem with strong latent capacity constraint and independence prior pressures. By augmenting the lower bound formulation with the $\beta$ coefficient that regulates the strength of such pressures and, as a consequence, the qualitative nature of the representations learnt by the model, we have achieved state of the art results for learning disentangled representations of data generative factors. We have shown that our proposed $\beta$ -VAE framework significantly outperforms both qualitatively and quantitatively the original VAE (Kingma & Welling, 2014), as well as state-of-the-art unsupervised (InfoGAN: Chen et al., 2016) and semi-supervised (DC-IGN: Kulkarni et al., 2015) approaches to disentangled factor learning. Furthermore, we have shown that $\beta$ -VAE consistently and robustly discovers more factors of variation in the data, and it learns a representation that covers a wider range of factor values and is disentangled more cleanly than other benchmarks, all in a completely unsupervised manner. Unlike InfoGAN and DC-IGN, our approach does not depend on any a priori knowledge about the number or the nature of data generative factors. Our preliminary investigations suggest that the performance of the $\beta$ -VAE framework may depend on the sampling density of the data generative factors within a training dataset (see Appendix A.8 for more details). It appears that having more densely sampled data generative factors results in better disentangling performance of $\beta$ -VAE, however we leave a more principled investigation of this effect to future work. + +$\beta$ -VAE is robust with respect to different architectures, optimisation parameters and datasets, hence requiring few design decisions. Our approach relies on the optimisation of a single hyperparameter $\beta$ , which can be found directly through a hyperparameter search if weakly labelled data is available to calculate our new proposed disentangling metric. Alternatively the optimal $\beta$ can be estimated heuristically in purely unsupervised scenarios. Learning an interpretable factorised representation of the independent data generative factors in a completely unsupervised manner is an important precursor for the development of artificial intelligence that understands the world in the same way that humans do (Lake et al., 2016). We believe that using our approach as an unsupervised pretraining stage for supervised or reinforcement learning will produce significant improvements for scenarios such as transfer or fast learning. + +# 6 ACKNOWLEDGEMENTS + +We would like to thank Charles Blundell, Danilo Rezende, Tejas Kulkarni and David Pfau for helpful comments that improved the manuscript. + +# REFERENCES + +M. Aubry, D. Maturana, A. Efros, B. Russell, and J. Sivic. Seeing 3d chairs: exemplar part-based 2d-3d alignment using a large dataset of cad models. In CVPR, 2014. +Y. Bengio, A. Courville, and P. Vincent. Representation learning: A review and new perspectives. In IEEE Transactions on Pattern Analysis & Machine Intelligence, 2013. +Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. arXiv, 2016. +Brian Cheung, Jesse A. Levezey, Arjun K. Bansal, and Bruno A. Olshausen. Discovering hidden factors of variation in deep networks. In Proceedings of the International Conference on Learning Representations, Workshop Track, 2015. +T. Cohen and M. Welling. Transformation properties of learned visual representations. In ICLR, 2015. +Taco Cohen and Max Welling. Learning the irreducible representations of commutative lie groups. arXiv, 2014. +G. Desjardins, A. Courville, and Y. Bengio. Disentangling factors of variation via generative entangling. arXiv, 2012. +Carl Doersch. Tutorial on variational autoencoders. arxiv, 2016. +John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 2011. +I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. NIPS, pp. 2672–2680, 2014. +Ross Goroshin, Michael Mathieu, and Yann LeCun. Learning to linearize under uncertainty. NIPS, 2015. +G. Hinton, A. Krizhevsky, and S. D. Wang. Transforming auto-encoders. International Conference on Artificial Neural Networks, 2011. +Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. An introduction to variational methods for graphical models. Machine learning, 37(2):183–233, 1999. +Theofanis Karaletsos, Serge Belongie, and Gunnar Rätsch. Bayesian representation learning with oracle constraints. ICLR, 2016. +W. Karush. Minima of Functions of Several Variables with Inequalities as Side Constraints. Master’s thesis, Univ. of Chicago, Chicago, Illinois, 1939. +D. P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2014. +D. P. Kingma and M. Welling. Auto-encoding variational bayes. ICLR, 2014. +H. W. Kuhn and A. W. Tucker. Nonlinear programming. In Proceedings of 2nd Berkeley Symposium, pp. 481–492, 1951. +Tejas Kulkarni, William Whitney, Pushmeet Kohli, and Joshua Tenenbaum. Deep convolutional inverse graphics network. NIPS, 2015. +Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building machines that learn and think like people. arXiv, 2016. +Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. ICCV, 2015. +P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3d face model for pose and illumination invariant face recognition. AVSS, 2009. +Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, and David Cournapeau. Scikit-learn: Machine learning in python. Journal of Machine Learning Research, 2011. +Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv, 2015. +Scott Reed, Kihyuk Sohn, Yuting Zhang, and Honglak Lee. Learning to disentangle factors of variation with manifold interaction. ICML, 2014. +Danilo J. Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. arXiv, 2014. +Karl Ridgeway. A survey of inductive biases for factorial Representation-Learning. arXiv, 2016. URL http://arxiv.org/abs/1612.05299. +Oren Rippel and Ryan Prescott Adams. High-dimensional probability estimation with deep density models. arXiv, 2013. +Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. arXiv, 2016. URL http://arxiv.org/abs/1606. 03498. +Jürgen Schmidhuber. Learning factorial codes by predictability minimization. Neural Computation, 4(6):863–869, 1992. +Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-Time single image and video Super-Resolution using an efficient Sub-Pixel convolutional neural network. arXiv, 2016. +Yichuan Tang, Ruslan Salakhutdinov, and Geoffrey Hinton. Tensor analyzers. In Proceedings of the 30th International Conference on Machine Learning, 2013, Atlanta, USA, 2013. +William F. Whitney, Michael Chang, Tejas Kulkarni, and Joshua B. Tenenbaum. Understanding visual concepts with continuation learning. arXiv, 2016. URL http://arxiv.org/pdf/ 1602.06822.pdf. +Jimei Yang, Scott Reed, Ming-Hsuan Yang, and Honglak Lee. Weakly-supervised disentangling with recurrent transformations for 3d view synthesis. NIPS, 2015. +Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network. arXiv, 2016. URL http://arxiv.org/abs/1609.03126. +Z. Zhu, P. Luo, X. Wang, and X. Tang. Multi-view perceptron: a deep model for learning face identity and view representations. In Advances in Neural Information Processing Systems 27. 2014. + +# A APPENDIX + +# A.1 MODEL ARCHITECTURE DETAILS + +A summary of all model architectures used in this paper can be seen in Tbl 1. + +# A.2 INFOGAN TRAINING + +To train the InfoGAN network described in Tbl. 1 on the 2D shapes dataset (Fig. 7), we followed the training paradigm described in Chen et al. (2016) with the following modifications. For the mutual information regularised latent code, we used 5 continuous variables $c _ { i }$ sampled uniformly from $( - 1 , 1 )$ . We used 5 noise variables $z _ { i }$ , as we found that using a reduced number of noise variables improved the quality of generated samples for this dataset. To help stabilise training, we used the instance noise trick described in Shi et al. (2016), adding Gaussian noise to the discriminator inputs (0.2 standard deviation on images scaled to $[ - 1 , 1 ] )$ . We followed Radford et al. (2015) for the architecture of the convolutional layers, and used batch normalisation in all layers except the last in the generator and the first in the discriminator. + +Table 1: Details of model architectures used in the paper. The models were trained using either adagrad (Duchi et al., 2011) or adam (Kingma & Ba, 2014) optimisers. + +
DatasetOptimiserArchitecture
2D shapesAdagradInput4096 (flattened 64x64x1).
(VAE)1e-2EncoderFC 1200,1200.ReLU activation.
Latents Decoder10
FC 1200,1200,1200,4096. Tanh activation.Bernoulli.
2D shapesrmspropInput64x64x1.
(DC-IGN)(as in Kulkarni et al., 2015)EncoderConv 96x3x3,48x3x3,48x3x3 (padding 1). ReLU activation and Max pooling 2x2.
Latents10
DecoderUnpooling,Conv 48x3x3,96x3x3,1x3x3. ReLU activation, Sigmoid.
2D shapes (InfoGAN)Adam 1e-3 (gen)GeneratorFC 256,256,Deconv 128x4x4, 64x4x4 (stride 2). Tanh.
2e-4 (dis)DiscriminatorConv and FC reverse of generator.Leaky ReLU activation. FC1. Sigmoid activation.
Recognition LatentsConv and FC shared with discriminator.FC 128,5.Gaussian 10: z1...5 ~ Unif(-1,1),c1...5 ~ Unif(-1,1)
ChairsAdamInput64x64x1.
(VAE)1e-4EncoderConv 32x4x4 (stride 2),32x4x4 (stride 2), 64x4x4 (stride 2),
Latents64x4x4 (stride 2),FC 256.ReLU activation. 32
DecoderDeconv reverse of encoder. ReLU activation.Bernoulli.
CelebA
(VAE)Adam 1e-4Input Encoder64x64x3. Conv 32x4x4 (stride 2),32x4x4 (stride 2),64x4x4 (stride 2),
Latents64x4x4 (stride 2),FC 256.ReLU activation. 32
DecoderDeconv reverse of encoder.ReLU activation.Gaussian.
3DFacesAdam64x64x1.
(VAE)1e-4Input EncoderConv 32x4x4 (stride 2),32x4x4 (stride 2), 64x4x4 (stride 2),
64x4x4 (stride 2),FC 256.ReLU activation. 32
Latents DecoderDeconv reverse of encoder.ReLU activation.Bernoulli.
+ +# A.3 ICA AND PCA BASELINES + +In order to calculate the ICA benchmark, we applied fastICA (Pedregosa et al., 2011) algorithm to the whitened pixel data. Due to memory limitations we had to apply the algorithm to pairwise combinations of the subsets of the dataset corresponding to the transforms of each of the three 2D object identities. We calculated the disentangling metric for all three ICA models trained on each of the three pairwise combinations of 2D objects, before presenting the average of these scores in Fig. 6. + +We performed PCA on the raw and whitened pixel data. Both approaches resulted in similar disentangling metric scores. Fig. 6 reports the PCA results calculated using whitened pixel data for more direct comparison with the ICA score. + +# A.4 DISENTANGLEMENT METRIC DETAILS + +We used a linear classifier to learn the identity of the generative factor that produced $\mathbf { z } _ { \mathrm { d i f f } } ^ { b }$ (see Equations (5) for the process used to obtain samples of $\mathbf { z } _ { \mathrm { d i f f } } ^ { b } .$ ). We used a fully connected linear classifier to predict $p ( y | \mathbf { z } _ { \mathrm { d i f f } } ^ { b } )$ , where $y$ is one of four generative factors (position X, position Y, scale and rotation). We used softmax output nonlinearity and a negative log likelihood loss function. The classifier was trained using the Adagrad (Duchi et al., 2011) optimisation algorithm with learning rate of 1e-2 until convergence. + +$$ +\mathcal { D } = \{ V \in \mathbb { R } ^ { K } , W \in \mathbb { R } ^ { H } , X \in \mathbb { R } ^ { N } \} , y \sim U n i f [ 1 . . . K ] +$$ + +Repeat for $b = 1 \dots B$ : + +$$ +\begin{array} { l } { { \displaystyle { \bf v } _ { 1 , l } \sim p ( { \bf v } ) , ~ { \bf w } _ { 1 , l } \sim p ( { \bf w } ) , ~ { \bf w } _ { 2 , l } \sim p ( { \bf w } ) , ~ [ { \bf v } _ { 2 , l } ] _ { k } = \left\{ \begin{array} { l l } { { \displaystyle { \bf v } _ { 1 , l } ] _ { k } , } } & { { \mathrm { i f ~ } k = y } } \\ { { \displaystyle \sim p ( v _ { k } ) , } } & { { \mathrm { o t h e r w i s e } } } \end{array} \right. } \ ~ } \\ { { \displaystyle { \bf x } _ { 1 , l } \sim \mathrm { S i m } ( { \bf v } _ { 1 , 1 } , { \bf w } _ { 1 , 1 } ) , ~ { \bf x } _ { 2 , 1 } \sim \mathrm { S i m } ( { \bf v } _ { 2 , 1 } , { \bf w } _ { 2 , 1 } ) , } \ ~ } \\ { { \displaystyle q ( { \bf z } | { \bf x } ) \sim \mathcal N ( \mu ( { \bf x } ) , \sigma ( { \bf x } ) ) , ~ { \bf z } _ { 1 , l } = \mu ( { \bf x } _ { 1 , l } ) , ~ { \bf z } _ { 2 , l } = \mu ( { \bf x } _ { 2 , l } ) } \ ~ } \\ { \displaystyle { \bf z } _ { \mathrm { d i f f } } ^ { l } = | { \bf z } _ { 1 , l } - { \bf z } _ { 2 , l } | , ~ { \bf z } _ { \mathrm { d i f f } } ^ { b } = \frac { 1 } { L } \sum _ { l = 1 } ^ { L } { \bf z } _ { \mathrm { d i f f } } ^ { l } } \end{array} +$$ + +All disentanglement metric score results reported in the paper were calculated in the following manner. Ten replicas of each model with the same hyperparameters were trained using different random seeds to obtain disentangled representations. Each of the ten trained model replicas was evaluated three times using the disentanglement metric score algorithm, each time using a different random seed to initialise the linear classifier. We then discarded the bottom $50 \%$ of the thirty resulting scores and reported the remaining results. This was done to control for the outlier results from the few experiments that diverged during training. + +The results reported in table in Fig. 6 (left) were calculated using the following data. Ground truth uses independent data generating factors $\mathbf { v }$ (our dataset did not contain any correlated data generating factors w). PCA and ICA decompositions keep the first ten components (PCA components explain $6 0 . 8 \%$ of variance). $\beta$ -VAE $\beta = 4 ,$ ), VAE $\begin{array} { r } { \beta = 1 , } \end{array}$ ) and VAE untrained have the same fully connected architecture with ten latent units $\mathbf { z }$ . InfoGAN uses “inferred” values of the five continuous latents that were regularised with the mutual information objective during training. + +# A.5 CLASSIFYING THE GROUND TRUTH DATA GENERATIVE FACTORS VALUES + +In order to further verify the validity of our proposed disentanglement metric we ran an extra quantitative test: we trained a linear classifier to predict the ground truth value of each of the five data generative factors used to generate the 2D shapes dataset. While this test does not measure disentangling directly (since it does not measure independence of the latent representation), a disentangled representation should make such a classification trivial. It can be seen in Table 2 that the representation learnt by $\beta$ -VAE is on average the best representation for factor classification across all five factors. It is closely followed by DC-IGN. It is interesting to note that ICA does well only at encoding object identity, while PCA manages to learn a very good representation of object position. + +
ModelClassification accuracy
idscalerotationposition Xposition Yaverage
PCA43.3836.085.9660.6660.1541.25
ICA59.634.47.6125.9625.1230.54
DC-IGN44.8245.9215.8947.6445.8840.03
InfoGAN44.4740.916.3927.5123.7328.60
VAEuntrained39.4425.336.0916.6914.3920.39
VAE41.5524.07816.518.7221.77
β-VAE50.0843.0320.3652.2549.543.04
+ +Table 2: Linear classifier classification accuracy for predicting the ground truth values for each data generative factor from different latent representations. Each factor could take a variable number of possible values: 3 for id, 6 for scale, 40 for rotation and 32 for position X or Y. Best performing model results in each column are printed in bold. + +# A.6 INTERPRETING NORMALISED $\beta$ + +We start with the $\beta$ -VAE constrained optimisation formulation that we have derived in Sec. 2. + +$$ +\mathcal { L } ( \boldsymbol { \theta } , \boldsymbol { \phi } ; \mathbf { x } , \mathbf { z } , \beta ) = \mathbb { E } _ { q _ { \phi } ( \mathbf { z } | \mathbf { x } ) } [ \log p _ { \theta } ( \mathbf { x } | \mathbf { z } ) ] - \beta D _ { K L } ( q _ { \phi } ( \mathbf { z } | \mathbf { x } ) | | p ( \mathbf { z } ) ) +$$ + +We make the assumption that every pixel $n$ in $\mathbf { x } \in \mathbb { R } ^ { N }$ is conditionally independent given $\mathbf { z }$ (Doersch, 2016). The first term of Eq. 6 then becomes: + +$$ +\mathbb { E } _ { q _ { \phi } ( \mathbf { z } | \mathbf { x } ) } [ \log p _ { \theta } ( \mathbf { x } | \mathbf { z } ) ] = \mathbb { E } _ { q _ { \phi } ( \mathbf { z } | \mathbf { x } ) } [ \log \prod _ { n } p _ { \theta } ( x _ { n } | \mathbf { z } ) ] = \mathbb { E } _ { q _ { \phi } ( \mathbf { z } | \mathbf { x } ) } [ \sum _ { n } \log p _ { \theta } ( x _ { n } | \mathbf { z } ) ] +$$ + +Dividing both sides of Eq. 6 by $N$ produces: + +$$ +\mathcal { L } ( \boldsymbol { \theta } , \boldsymbol { \phi } ; \mathbf { x } , \mathbf { z } , \beta ) \propto \mathbb { E } _ { q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) } \mathbb { E } _ { n } [ \log p _ { \boldsymbol { \theta } } ( x _ { n } | \mathbf { z } ) ] - \frac { \beta } { N } D _ { K L } ( q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) | | p ( \mathbf { z } ) ) +$$ + +We design $\beta$ -VAE to learn conditionally independent factors of variation in the data. Hence we assume conditional independence of every latent $z _ { m }$ given $x$ (where $m \in 1 . . . M$ , and $M$ is the dimensionality of $\mathbf { z }$ ). Since our prior $p ( \mathbf { z } )$ is an isotropic unit Gaussian, we can re-write the second term of Eq. 6 as: + +$$ +D _ { K L } ( q _ { \phi } ( \mathbf { z } | \mathbf { x } ) | | p ( \mathbf { z } ) ) = \int _ { z } q _ { \phi } ( \mathbf { z } | \mathbf { x } ) l o g { \frac { q _ { \phi } ( \mathbf { z } | \mathbf { x } ) } { p ( \mathbf { z } ) } } = \sum _ { m } \int _ { z _ { m } } q _ { \phi } ( z _ { m } | \mathbf { x } ) l o g { \frac { q _ { \phi } ( z _ { m } | \mathbf { x } ) } { p ( z _ { m } ) } } +$$ + +Multiplying the second term in Eq. 8 by a factor $\textstyle { \frac { M } { M } }$ produces: + +$$ +\begin{array} { r } { \displaystyle \mathcal { L } ( \boldsymbol { \theta } , \boldsymbol { \phi } ; \mathbf { x } , \mathbf { z } , \boldsymbol { \beta } ) \propto \mathbb { E } _ { q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) } \mathbb { E } _ { n } [ \log p _ { \boldsymbol { \theta } } ( x _ { n } | \mathbf { z } ) ] - \frac { \beta M } { N } \mathbb { E } _ { m } \int _ { z _ { m } } [ q _ { \boldsymbol { \phi } } ( z _ { m } | \mathbf { x } ) l o g \frac { q _ { \boldsymbol { \phi } } ( z _ { m } | \mathbf { x } ) } { p ( z _ { m } ) } ] } \\ { \displaystyle = \mathbb { E } _ { q _ { \boldsymbol { \phi } } ( \mathbf { z } | \mathbf { x } ) } \mathbb { E } _ { n } [ \log p _ { \boldsymbol { \theta } } ( x _ { n } | \mathbf { z } ) ] - \frac { \beta M } { N } \mathbb { E } _ { m } [ D _ { K L } ( q _ { \boldsymbol { \phi } } ( z _ { m } | \mathbf { x } ) | | p ( z _ { m } ) ) ] } \end{array} +$$ + +Hence using + +$$ +\beta _ { n o r m } = { \frac { \beta M } { N } } +$$ + +in Eq. 10 is equivalent to optimising the original $\beta$ -VAE formulation from Sec. 2, but with the additional independence assumptions that let us calculate data log likelihood and KL divergence terms in expectation over the individual pixels $x _ { n }$ and individual latents $z _ { m }$ . + +# A.7 RELATIONSHIP BETWEEN $\beta$ AND $\epsilon$ + +For a given $\epsilon$ we can solve the constrained optimisation problem in Eq. 3 (find the optimal $( \theta ^ { * } , \phi ^ { * } , \beta ^ { * } )$ , such that $\Delta \mathcal { F } ( \theta ^ { * } , \phi ^ { * } , \beta ^ { * } ) = 0 )$ . We can then re-write our optimal solution to the original optimisation problem in Eq. 2 as a function of $\epsilon$ : + +$$ +\mathcal { G } ( \theta ^ { * } ( \epsilon ) , \phi ^ { * } ( \epsilon ) ) = \mathbb { E } _ { q _ { \phi ^ { * } ( \epsilon ) } ( \mathbf { z } | \mathbf { x } ) } [ \log p _ { \theta ^ { * } ( \epsilon ) } ( \mathbf { x } | \mathbf { z } ) ] +$$ + +Now $\beta$ can be interpreted as the rate of change of the optimal solution $( \theta ^ { * } , \phi ^ { * } )$ to $\mathcal { G }$ when varying the constraint $\epsilon$ : + +$$ +\frac { \delta \mathcal { G } } { \delta \epsilon } = \beta ^ { * } ( \epsilon ) +$$ + +# A.8 DATA CONTINUITY + +We hypothesise that data continuity plays a role in guiding unsupervised models towards learning the correct data manifolds. To test this idea we measure how the degree of learnt disentangling changes with reduced continuity in the 2D shapes dataset. We trained a $\beta$ -VAE with $\beta = 4$ (Figure 7A) on subsamples of the original 2D shapes dataset, where we progressively decreased the generative factor sampling density. Reduction in data continuity negatively correlates with the average pixel wise (Hamming) distance between two consecutive transforms of each object (normalised by the average number of pixels occupied by each of the two adjacent transforms of an object to account for object scale). Figure 8 demonstrates that as the continuity in the data reduces, the degree of disentanglement in the learnt representations also drops. This effect holds after additional hyperparameter tuning and can not solely be explained by the decrease in dataset size, since the same VAE can learn disentangled representations from a data subset that preserves data continuity but is approximately $55 \%$ of the original size (results not shown). + +![](images/459a24497b2b3896af58ec63ae32465f95812d23a12e44b5ccde1d283f06622a.jpg) +Figure 8: Negative correlation between data transform continuity and the degree of disentangling achieved by $\beta$ -VAE. Abscissa is the average normalized Hamming distance between each of the two consecutive transforms of each object. Ordinate is disentanglement metric score. Disentangling performance is robust to Bernoulli noise added to the data at test time, as shown by slowly degrading classification accuracy up to $10 \%$ noise level, considering that the 2D objects occupy on average between $2 - 7 \%$ of the image depending on scale. Fluctuations in classification accuracy for similar Hamming distances are due the different nature of subsampled generative factors (i.e. symmetries are present in rotation but are lacking in position). + +# A.9 $\beta$ -VAE SAMPLES + +Samples from $\beta$ -VAE that learnt disentangled $\beta = 4$ ) and entangled $\beta = 1 \mathrm { \hbar }$ ) representations can be seen in Figure 9. + +# A.10 EXTRA $\beta$ -VAE TRAVERSAL PLOTS + +We present extra latent traversal plots from $\beta$ -VAE that learnt disentangled representations of 3D chairs (Figures 10-11) and CelebA (Figures 12-14) datasets. Here we show traversals from all informative latents from a large number of seed images. + +![](images/ae67406e8913a55260a5562a058ee2699d2b4069b4ef385b295db2de9abbca03.jpg) +Figure 9: Samples from $\beta$ -VAE trained on the dataset of 2D shapes that learnt either a disentangled (left, $\beta = 4$ ) or an entangled (right, $\beta = 1$ ) representation of the data generative factors. It can be seen that sampling from an entangled representation results in some unrealistic looking samples. A disentangled representation that inverts the original data generation process does not suffer from such errors. + +![](images/9049f21d4258d2516eb480744efc9e5f4109fd0f5a53c4fd5d1945d693c566ff.jpg) +Figure 10: Latent traversal plots from $\beta$ -VAE that learnt disentangled representations on the 3D chairs dataset. + +![](images/1a2cc13dc4e5fa96eb6370dc49a989426f7b720c781798e806809d71f5d4675c.jpg) +Figure 11: Latent traversal plots from $\beta$ -VAE that learnt disentangled representations on the 3D chairs dataset. + +![](images/ef3fae64fc29c55f9c1a18226beb0eb9b8bf0d442519e028e1cb69113cd260a7.jpg) +Figure 12: Latent traversal plots from $\beta$ -VAE that learnt disentangled representations on the CelebA dataset. + +![](images/01703dbdb4f3532ca41bf9156e8f6ca60335af5d67cbb56aab3078a7e2c7b3c2.jpg) +Figure 13: Latent traversal plots from $\beta$ -VAE that learnt disentangled representations on the CelebA dataset. + +![](images/f61801f2b3e9448756df3fb7e5d52bdc561084346f95faae387be2246ff24200.jpg) +Figure 14: Latent traversal plots from $\beta$ -VAE that learnt disentangled representations on the CelebA dataset. \ No newline at end of file diff --git a/parse/train/Sy2fzU9gl/Sy2fzU9gl_middle.json b/parse/train/Sy2fzU9gl/Sy2fzU9gl_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..2295f83efb70d5c4719085e57a27ecb3901db927 --- /dev/null +++ b/parse/train/Sy2fzU9gl/Sy2fzU9gl_middle.json @@ -0,0 +1,52027 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 506, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 118, + 98 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 78, + 506, + 98 + ], + "score": 1.0, + "content": "-VAE: LEARNING BASIC VISUAL CONCEPTS WITH A", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 429, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 429, + 118 + ], + "score": 1.0, + "content": "CONSTRAINED VARIATIONAL FRAMEWORK", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 111, + 135, + 436, + 158 + ], + "lines": [ + { + "bbox": [ + 111, + 133, + 437, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 133, + 437, + 149 + ], + "score": 1.0, + "content": "Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 368, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 368, + 159 + ], + "score": 1.0, + "content": "Matthew Botvinick, Shakir Mohamed, Alexander Lerchner", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 113, + 158, + 357, + 190 + ], + "lines": [ + { + "bbox": [ + 112, + 158, + 190, + 168 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 190, + 168 + ], + "score": 1.0, + "content": "Google DeepMind", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 168, + 358, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 358, + 181 + ], + "score": 1.0, + "content": "{irinah,lmatthey,arkap,cpburgess,glorotx,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 179, + 336, + 191 + ], + "spans": [ + { + "bbox": [ + 111, + 179, + 336, + 191 + ], + "score": 1.0, + "content": "botvinick,shakir,lerchner}@google.com", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 278, + 204, + 333, + 217 + ], + "lines": [ + { + "bbox": [ + 277, + 204, + 335, + 218 + ], + "spans": [ + { + "bbox": [ + 277, + 204, + 335, + 218 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 142, + 225, + 469, + 424 + ], + "lines": [ + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "score": 1.0, + "content": "Learning an interpretable factorised representation of the independent data gen-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "erative factors of the world without supervision is an important precursor for the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 247, + 469, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 247, + 469, + 260 + ], + "score": 1.0, + "content": "development of artificial intelligence that is able to learn and reason in the same", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 259, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 285, + 271 + ], + "score": 1.0, + "content": "way that humans do. We introduce", + "type": "text" + }, + { + "bbox": [ + 285, + 259, + 293, + 270 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 259, + 470, + 271 + ], + "score": 1.0, + "content": "-VAE, a new state-of-the-art framework for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 270, + 470, + 282 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 470, + 282 + ], + "score": 1.0, + "content": "automated discovery of interpretable factorised latent representations from raw", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 280, + 470, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 294 + ], + "score": 1.0, + "content": "image data in a completely unsupervised manner. Our approach is a modification", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 290, + 470, + 305 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 305 + ], + "score": 1.0, + "content": "of the variational autoencoder (VAE) framework. We introduce an adjustable hy-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 302, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 196, + 315 + ], + "score": 1.0, + "content": "perparameter", + "type": "text" + }, + { + "bbox": [ + 196, + 303, + 204, + 314 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 302, + 470, + 315 + ], + "score": 1.0, + "content": "that balances latent channel capacity and independence constraints", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 314, + 470, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 314, + 340, + 326 + ], + "score": 1.0, + "content": "with reconstruction accuracy. We demonstrate that", + "type": "text" + }, + { + "bbox": [ + 341, + 314, + 348, + 325 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 314, + 470, + 326 + ], + "score": 1.0, + "content": "-VAE with appropriately tuned", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 325, + 471, + 338 + ], + "spans": [ + { + "bbox": [ + 142, + 325, + 170, + 336 + ], + "score": 0.91, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 325, + 303, + 338 + ], + "score": 1.0, + "content": "qualitatively outperforms VAE", + "type": "text" + }, + { + "bbox": [ + 304, + 325, + 333, + 336 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\beta = 1 , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 325, + 471, + 338 + ], + "score": 1.0, + "content": "), as well as state of the art unsu-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 335, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 140, + 335, + 470, + 348 + ], + "score": 1.0, + "content": "pervised (InfoGAN) and semi-supervised (DC-IGN) approaches to disentangled", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 347, + 469, + 359 + ], + "spans": [ + { + "bbox": [ + 141, + 347, + 469, + 359 + ], + "score": 1.0, + "content": "factor learning on a variety of datasets (celebA, faces and chairs). Furthermore, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 358, + 469, + 370 + ], + "spans": [ + { + "bbox": [ + 141, + 358, + 469, + 370 + ], + "score": 1.0, + "content": "devise a protocol to quantitatively compare the degree of disentanglement learnt", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "score": 1.0, + "content": "by different models, and show that our approach also significantly outperforms", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 380, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 380, + 325, + 392 + ], + "score": 1.0, + "content": "all baselines quantitatively. Unlike InfoGAN,", + "type": "text" + }, + { + "bbox": [ + 325, + 380, + 333, + 390 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 380, + 469, + 392 + ], + "score": 1.0, + "content": "-VAE is stable to train, makes few", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 391, + 469, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 391, + 431, + 402 + ], + "score": 1.0, + "content": "assumptions about the data and relies on tuning a single hyperparameter", + "type": "text" + }, + { + "bbox": [ + 431, + 391, + 439, + 402 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 391, + 469, + 402 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "spans": [ + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "score": 1.0, + "content": "can be directly optimised through a hyperparameter search using weakly labelled", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 412, + 431, + 425 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 431, + 425 + ], + "score": 1.0, + "content": "data or through heuristic visual inspection for purely unsupervised data.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 437, + 206, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "The difficulty of learning a task for a given machine learning approach can vary significantly", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 473, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 504, + 486 + ], + "score": 1.0, + "content": "depending on the choice of the data representation. Having a representation that is well suited to the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "particular task and data domain can significantly improve the learning success and robustness of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "chosen model (Bengio et al., 2013). It has been suggested that learning a disentangled representation", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "score": 1.0, + "content": "of the generative factors in the data can be useful for a large variety of tasks and domains (Bengio", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "et al., 2013; Ridgeway, 2016). A disentangled representation can be defined as one where single", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "latent units are sensitive to changes in single generative factors, while being relatively invariant to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "changes in other factors (Bengio et al., 2013). For example, a model trained on a dataset of 3D objects", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "might learn independent latent units sensitive to single independent data generative factors, such as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "object identity, position, scale, lighting or colour, thus acting as an inverse graphics model (Kulkarni", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "et al., 2015). In a disentangled representation, knowledge about one factor can generalise to novel", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "configurations of other factors. According to Lake et al. (2016), disentangled representations could", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "boost the performance of state-of-the-art AI approaches in situations where they still struggle but", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "where humans excel. Such scenarios include those which require knowledge transfer, where faster", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "learning is achieved by reusing learnt representations for numerous tasks; zero-shot inference, where", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 627, + 504, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 504, + 640 + ], + "score": 1.0, + "content": "reasoning about new data is enabled by recombining previously learnt factors; or novelty detection.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "Unsupervised learning of a disentangled posterior distribution over the underlying generative factors", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "of sensory data is a major challenge in AI research (Bengio et al., 2013; Lake et al., 2016). Most", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "previous attempts required a priori knowledge of the number and/or nature of the data generative", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "score": 1.0, + "content": "factors (Hinton et al., 2011; Rippel & Adams, 2013; Reed et al., 2014; Zhu et al., 2014; Yang", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "et al., 2015; Goroshin et al., 2015; Kulkarni et al., 2015; Cheung et al., 2015; Whitney et al., 2016;", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Karaletsos et al., 2016). This is not always feasible in the real world, where the newly initialised", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "learner may be exposed to complex data where no a priori knowledge of the generative factors exists,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "and little to no supervision for discovering the factors is available. Until recently purely unsupervised", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 506, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 118, + 98 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 78, + 506, + 98 + ], + "score": 1.0, + "content": "-VAE: LEARNING BASIC VISUAL CONCEPTS WITH A", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 429, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 429, + 118 + ], + "score": 1.0, + "content": "CONSTRAINED VARIATIONAL FRAMEWORK", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 111, + 135, + 436, + 158 + ], + "lines": [ + { + "bbox": [ + 111, + 133, + 437, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 133, + 437, + 149 + ], + "score": 1.0, + "content": "Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 368, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 368, + 159 + ], + "score": 1.0, + "content": "Matthew Botvinick, Shakir Mohamed, Alexander Lerchner", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 111, + 133, + 437, + 159 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 158, + 357, + 190 + ], + "lines": [ + { + "bbox": [ + 112, + 158, + 190, + 168 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 190, + 168 + ], + "score": 1.0, + "content": "Google DeepMind", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 168, + 358, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 358, + 181 + ], + "score": 1.0, + "content": "{irinah,lmatthey,arkap,cpburgess,glorotx,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 179, + 336, + 191 + ], + "spans": [ + { + "bbox": [ + 111, + 179, + 336, + 191 + ], + "score": 1.0, + "content": "botvinick,shakir,lerchner}@google.com", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 111, + 158, + 358, + 191 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 204, + 333, + 217 + ], + "lines": [ + { + "bbox": [ + 277, + 204, + 335, + 218 + ], + "spans": [ + { + "bbox": [ + 277, + 204, + 335, + 218 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 142, + 225, + 469, + 424 + ], + "lines": [ + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "score": 1.0, + "content": "Learning an interpretable factorised representation of the independent data gen-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "erative factors of the world without supervision is an important precursor for the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 247, + 469, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 247, + 469, + 260 + ], + "score": 1.0, + "content": "development of artificial intelligence that is able to learn and reason in the same", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 259, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 285, + 271 + ], + "score": 1.0, + "content": "way that humans do. We introduce", + "type": "text" + }, + { + "bbox": [ + 285, + 259, + 293, + 270 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 259, + 470, + 271 + ], + "score": 1.0, + "content": "-VAE, a new state-of-the-art framework for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 270, + 470, + 282 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 470, + 282 + ], + "score": 1.0, + "content": "automated discovery of interpretable factorised latent representations from raw", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 280, + 470, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 294 + ], + "score": 1.0, + "content": "image data in a completely unsupervised manner. Our approach is a modification", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 290, + 470, + 305 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 305 + ], + "score": 1.0, + "content": "of the variational autoencoder (VAE) framework. We introduce an adjustable hy-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 302, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 196, + 315 + ], + "score": 1.0, + "content": "perparameter", + "type": "text" + }, + { + "bbox": [ + 196, + 303, + 204, + 314 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 302, + 470, + 315 + ], + "score": 1.0, + "content": "that balances latent channel capacity and independence constraints", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 314, + 470, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 314, + 340, + 326 + ], + "score": 1.0, + "content": "with reconstruction accuracy. We demonstrate that", + "type": "text" + }, + { + "bbox": [ + 341, + 314, + 348, + 325 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 314, + 470, + 326 + ], + "score": 1.0, + "content": "-VAE with appropriately tuned", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 325, + 471, + 338 + ], + "spans": [ + { + "bbox": [ + 142, + 325, + 170, + 336 + ], + "score": 0.91, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 325, + 303, + 338 + ], + "score": 1.0, + "content": "qualitatively outperforms VAE", + "type": "text" + }, + { + "bbox": [ + 304, + 325, + 333, + 336 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\beta = 1 , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 325, + 471, + 338 + ], + "score": 1.0, + "content": "), as well as state of the art unsu-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 335, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 140, + 335, + 470, + 348 + ], + "score": 1.0, + "content": "pervised (InfoGAN) and semi-supervised (DC-IGN) approaches to disentangled", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 347, + 469, + 359 + ], + "spans": [ + { + "bbox": [ + 141, + 347, + 469, + 359 + ], + "score": 1.0, + "content": "factor learning on a variety of datasets (celebA, faces and chairs). Furthermore, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 358, + 469, + 370 + ], + "spans": [ + { + "bbox": [ + 141, + 358, + 469, + 370 + ], + "score": 1.0, + "content": "devise a protocol to quantitatively compare the degree of disentanglement learnt", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 381 + ], + "score": 1.0, + "content": "by different models, and show that our approach also significantly outperforms", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 380, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 380, + 325, + 392 + ], + "score": 1.0, + "content": "all baselines quantitatively. Unlike InfoGAN,", + "type": "text" + }, + { + "bbox": [ + 325, + 380, + 333, + 390 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 380, + 469, + 392 + ], + "score": 1.0, + "content": "-VAE is stable to train, makes few", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 391, + 469, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 391, + 431, + 402 + ], + "score": 1.0, + "content": "assumptions about the data and relies on tuning a single hyperparameter", + "type": "text" + }, + { + "bbox": [ + 431, + 391, + 439, + 402 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 391, + 469, + 402 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "spans": [ + { + "bbox": [ + 141, + 401, + 469, + 414 + ], + "score": 1.0, + "content": "can be directly optimised through a hyperparameter search using weakly labelled", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 412, + 431, + 425 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 431, + 425 + ], + "score": 1.0, + "content": "data or through heuristic visual inspection for purely unsupervised data.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 16.5, + "bbox_fs": [ + 140, + 225, + 471, + 425 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 437, + 206, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 208, + 452 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "The difficulty of learning a task for a given machine learning approach can vary significantly", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 473, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 504, + 486 + ], + "score": 1.0, + "content": "depending on the choice of the data representation. Having a representation that is well suited to the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "particular task and data domain can significantly improve the learning success and robustness of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "chosen model (Bengio et al., 2013). It has been suggested that learning a disentangled representation", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "score": 1.0, + "content": "of the generative factors in the data can be useful for a large variety of tasks and domains (Bengio", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "et al., 2013; Ridgeway, 2016). A disentangled representation can be defined as one where single", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "latent units are sensitive to changes in single generative factors, while being relatively invariant to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "changes in other factors (Bengio et al., 2013). For example, a model trained on a dataset of 3D objects", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "might learn independent latent units sensitive to single independent data generative factors, such as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "object identity, position, scale, lighting or colour, thus acting as an inverse graphics model (Kulkarni", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "et al., 2015). In a disentangled representation, knowledge about one factor can generalise to novel", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "configurations of other factors. According to Lake et al. (2016), disentangled representations could", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "boost the performance of state-of-the-art AI approaches in situations where they still struggle but", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "where humans excel. Such scenarios include those which require knowledge transfer, where faster", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "learning is achieved by reusing learnt representations for numerous tasks; zero-shot inference, where", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 627, + 504, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 504, + 640 + ], + "score": 1.0, + "content": "reasoning about new data is enabled by recombining previously learnt factors; or novelty detection.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 461, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "Unsupervised learning of a disentangled posterior distribution over the underlying generative factors", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "of sensory data is a major challenge in AI research (Bengio et al., 2013; Lake et al., 2016). Most", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "previous attempts required a priori knowledge of the number and/or nature of the data generative", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "score": 1.0, + "content": "factors (Hinton et al., 2011; Rippel & Adams, 2013; Reed et al., 2014; Zhu et al., 2014; Yang", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "et al., 2015; Goroshin et al., 2015; Kulkarni et al., 2015; Cheung et al., 2015; Whitney et al., 2016;", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Karaletsos et al., 2016). This is not always feasible in the real world, where the newly initialised", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "learner may be exposed to complex data where no a priori knowledge of the generative factors exists,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "and little to no supervision for discovering the factors is available. Until recently purely unsupervised", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "approaches to disentangled factor learning have not scaled well (Schmidhuber, 1992; Desjardins", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 510, + 357, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 510, + 357, + 524 + ], + "score": 1.0, + "content": "et al., 2012; Tang et al., 2013; Cohen & Welling, 2014; 2015).", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 644, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 79, + 498, + 363 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 79, + 498, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 79, + 498, + 363 + ], + "spans": [ + { + "bbox": [ + 108, + 79, + 498, + 363 + ], + "score": 0.969, + "type": "image", + "image_path": "6bae38335f4edd4fe6fe641c6af928cf5435a2ee5710353b8d5ef567cae439cf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 79, + 498, + 173.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 173.66666666666669, + 498, + 268.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 268.33333333333337, + 498, + 363.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 367, + 505, + 487 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Figure 1: Manipulating latent variables on celebA: Qualitative results comparing disentangling", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 171, + 390 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 172, + 378, + 179, + 389 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 378, + 206, + 390 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 207, + 378, + 244, + 389 + ], + "score": 0.86, + "content": "\\beta = 2 5 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 378, + 387, + 390 + ], + "score": 1.0, + "content": "), VAE (Kingma & Welling, 2014)", + "type": "text" + }, + { + "bbox": [ + 387, + 378, + 415, + 389 + ], + "score": 0.86, + "content": "( \\beta = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 378, + 506, + 390 + ], + "score": 1.0, + "content": ") and InfoGAN (Chen", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "et al., 2016). In all figures of latent code traversal each block corresponds to the traversal of a single", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 371, + 412 + ], + "score": 1.0, + "content": "latent variable while keeping others fixed to either their inferred (", + "type": "text" + }, + { + "bbox": [ + 371, + 400, + 378, + 411 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "-VAE, VAE and DC-IGN where", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "applicable) or sampled (InfoGAN) values. Each row represents a different seed image used to infer", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 422, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 506, + 434 + ], + "score": 1.0, + "content": "the latent values in the VAE-based models, or a random sample of the noise variables in InfoGAN.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 433, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 107, + 433, + 114, + 444 + ], + "score": 0.78, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 433, + 263, + 445 + ], + "score": 1.0, + "content": "-VAE and VAE traversal is over the", + "type": "text" + }, + { + "bbox": [ + 264, + 433, + 293, + 444 + ], + "score": 0.35, + "content": "I { \\cdot } 3 , \\ 3 I", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 433, + 506, + 445 + ], + "score": 1.0, + "content": "range. InfoGAN traversal is over ten dimensional", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 245, + 456 + ], + "score": 1.0, + "content": "categorical latent variables. Only", + "type": "text" + }, + { + "bbox": [ + 245, + 444, + 252, + 455 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "-VAE and InfoGAN learnt to disentangle factors like azimuth", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "(a), emotion (b) and hair style (c), whereas VAE learnt an entangled representation (e.g. azimuth is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "score": 1.0, + "content": "entangled with emotion, presence of glasses and gender). InfoGAN images adapted from Chen et al.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 477, + 249, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 249, + 489 + ], + "score": 1.0, + "content": "(2016). Reprinted with permission.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "approaches to disentangled factor learning have not scaled well (Schmidhuber, 1992; Desjardins", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 510, + 357, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 510, + 357, + 524 + ], + "score": 1.0, + "content": "et al., 2012; Tang et al., 2013; Cohen & Welling, 2014; 2015).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "Recently a scalable unsupervised approach for disentangled factor learning has been developed,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 540, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 506, + 551 + ], + "score": 1.0, + "content": "called InfoGAN (Chen et al., 2016). InfoGAN extends the generative adversarial network (GAN)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "(Goodfellow et al., 2014) framework to additionally maximise the mutual information between a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "subset of the generating noise variables and the output of a recognition network. It has been reported", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "to be capable of discovering at least a subset of data generative factors and of learning a disentangled", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "representation of these factors. The reliance of InfoGAN on the GAN framework, however, comes", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "at the cost of training instability and reduced sample diversity. Furthermore, InfoGAN requires", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "some a priori knowledge of the data, since its performance is sensitive to the choice of the prior", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "distribution and the number of the regularised noise variables. InfoGAN also lacks a principled", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "inference network (although the recognition network can be used as one). The ability to infer the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "posterior latent distribution from sensory input is important when using the unsupervised model in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "transfer learning or zero-shot inference scenarios. Hence, while InfoGAN is an important step in the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "right direction, we believe that further improvements are necessary to achieve a principled way of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "using unsupervised learning for developing more human-like learning and reasoning in algorithms as", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 681, + 235, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 235, + 694 + ], + "score": 1.0, + "content": "described by Lake et al. (2016).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Finally, there is currently no general method for quantifying the degree of learnt disentanglement.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Therefore there is no way to quantitatively compare the degree of disentanglement achieved by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 719, + 412, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 412, + 734 + ], + "score": 1.0, + "content": "different models or when optimising the hyperparameters of a single model.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 79, + 498, + 363 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 79, + 498, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 79, + 498, + 363 + ], + "spans": [ + { + "bbox": [ + 108, + 79, + 498, + 363 + ], + "score": 0.969, + "type": "image", + "image_path": "6bae38335f4edd4fe6fe641c6af928cf5435a2ee5710353b8d5ef567cae439cf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 79, + 498, + 173.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 173.66666666666669, + 498, + 268.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 268.33333333333337, + 498, + 363.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 367, + 505, + 487 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Figure 1: Manipulating latent variables on celebA: Qualitative results comparing disentangling", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 171, + 390 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 172, + 378, + 179, + 389 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 378, + 206, + 390 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 207, + 378, + 244, + 389 + ], + "score": 0.86, + "content": "\\beta = 2 5 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 378, + 387, + 390 + ], + "score": 1.0, + "content": "), VAE (Kingma & Welling, 2014)", + "type": "text" + }, + { + "bbox": [ + 387, + 378, + 415, + 389 + ], + "score": 0.86, + "content": "( \\beta = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 378, + 506, + 390 + ], + "score": 1.0, + "content": ") and InfoGAN (Chen", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "et al., 2016). In all figures of latent code traversal each block corresponds to the traversal of a single", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 371, + 412 + ], + "score": 1.0, + "content": "latent variable while keeping others fixed to either their inferred (", + "type": "text" + }, + { + "bbox": [ + 371, + 400, + 378, + 411 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "-VAE, VAE and DC-IGN where", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "applicable) or sampled (InfoGAN) values. Each row represents a different seed image used to infer", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 422, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 506, + 434 + ], + "score": 1.0, + "content": "the latent values in the VAE-based models, or a random sample of the noise variables in InfoGAN.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 433, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 107, + 433, + 114, + 444 + ], + "score": 0.78, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 433, + 263, + 445 + ], + "score": 1.0, + "content": "-VAE and VAE traversal is over the", + "type": "text" + }, + { + "bbox": [ + 264, + 433, + 293, + 444 + ], + "score": 0.35, + "content": "I { \\cdot } 3 , \\ 3 I", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 433, + 506, + 445 + ], + "score": 1.0, + "content": "range. InfoGAN traversal is over ten dimensional", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 245, + 456 + ], + "score": 1.0, + "content": "categorical latent variables. Only", + "type": "text" + }, + { + "bbox": [ + 245, + 444, + 252, + 455 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "-VAE and InfoGAN learnt to disentangle factors like azimuth", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "(a), emotion (b) and hair style (c), whereas VAE learnt an entangled representation (e.g. azimuth is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 507, + 478 + ], + "score": 1.0, + "content": "entangled with emotion, presence of glasses and gender). InfoGAN images adapted from Chen et al.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 477, + 249, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 249, + 489 + ], + "score": 1.0, + "content": "(2016). Reprinted with permission.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 504, + 523 + ], + "lines": [], + "index": 14.5, + "bbox_fs": [ + 104, + 500, + 505, + 524 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "Recently a scalable unsupervised approach for disentangled factor learning has been developed,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 540, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 506, + 551 + ], + "score": 1.0, + "content": "called InfoGAN (Chen et al., 2016). InfoGAN extends the generative adversarial network (GAN)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "(Goodfellow et al., 2014) framework to additionally maximise the mutual information between a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "subset of the generating noise variables and the output of a recognition network. It has been reported", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "to be capable of discovering at least a subset of data generative factors and of learning a disentangled", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "representation of these factors. The reliance of InfoGAN on the GAN framework, however, comes", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "at the cost of training instability and reduced sample diversity. Furthermore, InfoGAN requires", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "some a priori knowledge of the data, since its performance is sensitive to the choice of the prior", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "distribution and the number of the regularised noise variables. InfoGAN also lacks a principled", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "inference network (although the recognition network can be used as one). The ability to infer the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "posterior latent distribution from sensory input is important when using the unsupervised model in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "transfer learning or zero-shot inference scenarios. Hence, while InfoGAN is an important step in the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "right direction, we believe that further improvements are necessary to achieve a principled way of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "using unsupervised learning for developing more human-like learning and reasoning in algorithms as", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 681, + 235, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 235, + 694 + ], + "score": 1.0, + "content": "described by Lake et al. (2016).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 528, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Finally, there is currently no general method for quantifying the degree of learnt disentanglement.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Therefore there is no way to quantitatively compare the degree of disentanglement achieved by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 719, + 412, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 412, + 734 + ], + "score": 1.0, + "content": "different models or when optimising the hyperparameters of a single model.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 698, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 77, + 503, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 77, + 503, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 77, + 503, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 503, + 264 + ], + "score": 0.967, + "type": "image", + "image_path": "504fafaff72a9120d6f80158ae647468090601903bd9e60571b38eabc767fa49.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 77, + 503, + 139.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 139.33333333333334, + 503, + 201.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 201.66666666666669, + 503, + 264.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 272, + 506, + 371 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 271, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 271, + 506, + 286 + ], + "score": 1.0, + "content": "Figure 2: Manipulating latent variables on 3D chairs: Qualitative results comparing disentangling", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 283, + 507, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 171, + 297 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 172, + 284, + 179, + 295 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 283, + 206, + 297 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 207, + 284, + 234, + 295 + ], + "score": 0.85, + "content": "\\beta = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 283, + 378, + 297 + ], + "score": 1.0, + "content": "), VAE (Kingma & Welling, 2014)", + "type": "text" + }, + { + "bbox": [ + 378, + 284, + 406, + 295 + ], + "score": 0.87, + "content": "\\mathcal { \\beta } = 1 \\mathrm { \\check { \\beta } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 283, + 507, + 297 + ], + "score": 1.0, + "content": "), InfoGAN (Chen et al.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 293, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 104, + 293, + 506, + 308 + ], + "score": 1.0, + "content": "2016) and DC-IGN (Kulkarni et al., 2015). InfoGAN traversal is over the [-1, 1] range. VAE always", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 305, + 507, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 507, + 319 + ], + "score": 1.0, + "content": "learns an entangled representation (e.g. chair width is entangled with azimuth and leg style (b)).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 316, + 507, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 507, + 329 + ], + "score": 1.0, + "content": "All models apart from VAE learnt to disentangle the labelled data generative factor, azimuth (a).", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 163, + 340 + ], + "score": 1.0, + "content": "InfoGAN and", + "type": "text" + }, + { + "bbox": [ + 163, + 328, + 171, + 339 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "-VAE were also able to discover unlabelled factors in the dataset, such as chair width", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 147, + 351 + ], + "score": 1.0, + "content": "(b). Only", + "type": "text" + }, + { + "bbox": [ + 147, + 339, + 154, + 350 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "-VAE, however, learnt about the unlabelled factor of chair leg style (c). InfoGAN and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 348, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 506, + 363 + ], + "score": 1.0, + "content": "DC-IGN images adapted from Chen et al. (2016) and Kulkarni et al. (2015), respectively. Reprinted", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 361, + 174, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 174, + 373 + ], + "score": 1.0, + "content": "with permission.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 346, + 403 + ], + "score": 1.0, + "content": "In this paper we attempt to address these issues. We propose", + "type": "text" + }, + { + "bbox": [ + 347, + 391, + 354, + 402 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "-VAE, a deep unsupervised generative", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "approach for disentangled factor learning that can automatically discover the independent latent", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "score": 1.0, + "content": "factors of variation in unsupervised data. Our approach is based on the variational autoencoder (VAE)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 421, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 438 + ], + "score": 1.0, + "content": "framework (Kingma & Welling, 2014; Rezende et al., 2014), which brings scalability and training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "stability. While the original VAE work has been shown to achieve limited disentangling performance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "on simple datasets, such as FreyFaces or MNIST (Kingma & Welling, 2014), disentangling perfor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "mance does not scale to more complex datasets (e.g. Aubry et al., 2014; Paysan et al., 2009; Liu et al.,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "2015), prompting the development of more elaborate semi-supervised VAE-based approaches for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 478, + 428, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 428, + 491 + ], + "score": 1.0, + "content": "learning disentangled factors (e.g. Kulkarni et al., 2015; Karaletsos et al., 2016).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 435, + 509 + ], + "score": 1.0, + "content": "We propose augmenting the original VAE framework with a single hyperparameter", + "type": "text" + }, + { + "bbox": [ + 436, + 496, + 444, + 507 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "that modulates", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "the learning constraints applied to the model. These constraints impose a limit on the capacity of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 518, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 506, + 530 + ], + "score": 1.0, + "content": "the latent information channel and control the emphasis on learning statistically independent latent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 527, + 507, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 139, + 542 + ], + "score": 1.0, + "content": "factors.", + "type": "text" + }, + { + "bbox": [ + 139, + 529, + 147, + 540 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 527, + 191, + 542 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + }, + { + "bbox": [ + 191, + 528, + 217, + 540 + ], + "score": 0.91, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 527, + 507, + 542 + ], + "score": 1.0, + "content": "corresponds to the original VAE framework (Kingma & Welling, 2014;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 217, + 552 + ], + "score": 1.0, + "content": "Rezende et al., 2014). With", + "type": "text" + }, + { + "bbox": [ + 218, + 540, + 243, + 551 + ], + "score": 0.91, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "the model is pushed to learn a more efficient latent representation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "of the data, which is disentangled if the data contains at least some underlying factors of variation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 376, + 574 + ], + "score": 1.0, + "content": "that are independent. We show that this simple modification allows", + "type": "text" + }, + { + "bbox": [ + 376, + 562, + 384, + 573 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "-VAE to significantly improve", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "the degree of disentanglement in learnt latent representations compared to the unmodified VAE", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 473, + 595 + ], + "score": 1.0, + "content": "framework (Kingma & Welling, 2014; Rezende et al., 2014). Furthermore, we show that", + "type": "text" + }, + { + "bbox": [ + 474, + 583, + 481, + 594 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 594, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 507, + 607 + ], + "score": 1.0, + "content": "achieves state of the art disentangling performance against both the best unsupervised (InfoGAN:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "Chen et al., 2016) and semi-supervised (DC-IGN: Kulkarni et al., 2015) approaches for disentangled", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "factor learning on a number of benchmark datasets, such as CelebA (Liu et al., 2015), chairs (Aubry", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "et al., 2014) and faces (Paysan et al., 2009) using qualitative evaluation. Finally, to help quantify", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 421, + 651 + ], + "score": 1.0, + "content": "the differences, we develop a new measure of disentanglement and show that", + "type": "text" + }, + { + "bbox": [ + 421, + 639, + 428, + 649 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "-VAE significantly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 650, + 504, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 504, + 660 + ], + "score": 1.0, + "content": "outperforms all our baselines on this measure (ICA, PCA, VAE Kingma & Ba (2014), DC-IGN", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 659, + 335, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 335, + 672 + ], + "score": 1.0, + "content": "Kulkarni et al. (2015), and InfoGAN Chen et al. (2016)).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 336, + 689 + ], + "score": 1.0, + "content": "Our main contributions are the following: 1) we propose", + "type": "text" + }, + { + "bbox": [ + 337, + 678, + 343, + 689 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "-VAE, a new unsupervised approach for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "learning disentangled representations of independent visual data generative factors; 2) we devise a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "protocol to quantitatively compare the degree of disentanglement learnt by different models; 3) we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 337, + 723 + ], + "score": 1.0, + "content": "demonstrate both qualitatively and quantitatively that our", + "type": "text" + }, + { + "bbox": [ + 337, + 712, + 344, + 721 + ], + "score": 0.87, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "-VAE approach achieves state-of-the-art", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 483, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 483, + 734 + ], + "score": 1.0, + "content": "disentanglement performance compared to various baselines on a variety of complex datasets.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 77, + 503, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 77, + 503, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 77, + 503, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 503, + 264 + ], + "score": 0.967, + "type": "image", + "image_path": "504fafaff72a9120d6f80158ae647468090601903bd9e60571b38eabc767fa49.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 77, + 503, + 139.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 139.33333333333334, + 503, + 201.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 201.66666666666669, + 503, + 264.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 272, + 506, + 371 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 271, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 271, + 506, + 286 + ], + "score": 1.0, + "content": "Figure 2: Manipulating latent variables on 3D chairs: Qualitative results comparing disentangling", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 283, + 507, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 171, + 297 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 172, + 284, + 179, + 295 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 283, + 206, + 297 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 207, + 284, + 234, + 295 + ], + "score": 0.85, + "content": "\\beta = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 283, + 378, + 297 + ], + "score": 1.0, + "content": "), VAE (Kingma & Welling, 2014)", + "type": "text" + }, + { + "bbox": [ + 378, + 284, + 406, + 295 + ], + "score": 0.87, + "content": "\\mathcal { \\beta } = 1 \\mathrm { \\check { \\beta } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 283, + 507, + 297 + ], + "score": 1.0, + "content": "), InfoGAN (Chen et al.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 293, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 104, + 293, + 506, + 308 + ], + "score": 1.0, + "content": "2016) and DC-IGN (Kulkarni et al., 2015). InfoGAN traversal is over the [-1, 1] range. VAE always", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 305, + 507, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 507, + 319 + ], + "score": 1.0, + "content": "learns an entangled representation (e.g. chair width is entangled with azimuth and leg style (b)).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 316, + 507, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 507, + 329 + ], + "score": 1.0, + "content": "All models apart from VAE learnt to disentangle the labelled data generative factor, azimuth (a).", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 163, + 340 + ], + "score": 1.0, + "content": "InfoGAN and", + "type": "text" + }, + { + "bbox": [ + 163, + 328, + 171, + 339 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "-VAE were also able to discover unlabelled factors in the dataset, such as chair width", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 147, + 351 + ], + "score": 1.0, + "content": "(b). Only", + "type": "text" + }, + { + "bbox": [ + 147, + 339, + 154, + 350 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "-VAE, however, learnt about the unlabelled factor of chair leg style (c). InfoGAN and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 348, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 506, + 363 + ], + "score": 1.0, + "content": "DC-IGN images adapted from Chen et al. (2016) and Kulkarni et al. (2015), respectively. Reprinted", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 361, + 174, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 174, + 373 + ], + "score": 1.0, + "content": "with permission.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 346, + 403 + ], + "score": 1.0, + "content": "In this paper we attempt to address these issues. We propose", + "type": "text" + }, + { + "bbox": [ + 347, + 391, + 354, + 402 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "-VAE, a deep unsupervised generative", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "approach for disentangled factor learning that can automatically discover the independent latent", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "score": 1.0, + "content": "factors of variation in unsupervised data. Our approach is based on the variational autoencoder (VAE)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 421, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 438 + ], + "score": 1.0, + "content": "framework (Kingma & Welling, 2014; Rezende et al., 2014), which brings scalability and training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "stability. While the original VAE work has been shown to achieve limited disentangling performance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "on simple datasets, such as FreyFaces or MNIST (Kingma & Welling, 2014), disentangling perfor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "mance does not scale to more complex datasets (e.g. Aubry et al., 2014; Paysan et al., 2009; Liu et al.,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "2015), prompting the development of more elaborate semi-supervised VAE-based approaches for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 478, + 428, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 428, + 491 + ], + "score": 1.0, + "content": "learning disentangled factors (e.g. Kulkarni et al., 2015; Karaletsos et al., 2016).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 391, + 506, + 491 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 435, + 509 + ], + "score": 1.0, + "content": "We propose augmenting the original VAE framework with a single hyperparameter", + "type": "text" + }, + { + "bbox": [ + 436, + 496, + 444, + 507 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "that modulates", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "the learning constraints applied to the model. These constraints impose a limit on the capacity of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 518, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 506, + 530 + ], + "score": 1.0, + "content": "the latent information channel and control the emphasis on learning statistically independent latent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 527, + 507, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 139, + 542 + ], + "score": 1.0, + "content": "factors.", + "type": "text" + }, + { + "bbox": [ + 139, + 529, + 147, + 540 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 527, + 191, + 542 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + }, + { + "bbox": [ + 191, + 528, + 217, + 540 + ], + "score": 0.91, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 527, + 507, + 542 + ], + "score": 1.0, + "content": "corresponds to the original VAE framework (Kingma & Welling, 2014;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 217, + 552 + ], + "score": 1.0, + "content": "Rezende et al., 2014). With", + "type": "text" + }, + { + "bbox": [ + 218, + 540, + 243, + 551 + ], + "score": 0.91, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "the model is pushed to learn a more efficient latent representation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "of the data, which is disentangled if the data contains at least some underlying factors of variation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 376, + 574 + ], + "score": 1.0, + "content": "that are independent. We show that this simple modification allows", + "type": "text" + }, + { + "bbox": [ + 376, + 562, + 384, + 573 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "-VAE to significantly improve", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "the degree of disentanglement in learnt latent representations compared to the unmodified VAE", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 473, + 595 + ], + "score": 1.0, + "content": "framework (Kingma & Welling, 2014; Rezende et al., 2014). Furthermore, we show that", + "type": "text" + }, + { + "bbox": [ + 474, + 583, + 481, + 594 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 594, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 507, + 607 + ], + "score": 1.0, + "content": "achieves state of the art disentangling performance against both the best unsupervised (InfoGAN:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "Chen et al., 2016) and semi-supervised (DC-IGN: Kulkarni et al., 2015) approaches for disentangled", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "factor learning on a number of benchmark datasets, such as CelebA (Liu et al., 2015), chairs (Aubry", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "et al., 2014) and faces (Paysan et al., 2009) using qualitative evaluation. Finally, to help quantify", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 421, + 651 + ], + "score": 1.0, + "content": "the differences, we develop a new measure of disentanglement and show that", + "type": "text" + }, + { + "bbox": [ + 421, + 639, + 428, + 649 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "-VAE significantly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 650, + 504, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 504, + 660 + ], + "score": 1.0, + "content": "outperforms all our baselines on this measure (ICA, PCA, VAE Kingma & Ba (2014), DC-IGN", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 659, + 335, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 335, + 672 + ], + "score": 1.0, + "content": "Kulkarni et al. (2015), and InfoGAN Chen et al. (2016)).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 495, + 507, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 336, + 689 + ], + "score": 1.0, + "content": "Our main contributions are the following: 1) we propose", + "type": "text" + }, + { + "bbox": [ + 337, + 678, + 343, + 689 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "-VAE, a new unsupervised approach for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "learning disentangled representations of independent visual data generative factors; 2) we devise a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "protocol to quantitatively compare the degree of disentanglement learnt by different models; 3) we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 337, + 723 + ], + "score": 1.0, + "content": "demonstrate both qualitatively and quantitatively that our", + "type": "text" + }, + { + "bbox": [ + 337, + 712, + 344, + 721 + ], + "score": 0.87, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "-VAE approach achieves state-of-the-art", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 483, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 483, + 734 + ], + "score": 1.0, + "content": "disentanglement performance compared to various baselines on a variety of complex datasets.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 78, + 503, + 383 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 78, + 503, + 383 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 78, + 503, + 383 + ], + "spans": [ + { + "bbox": [ + 107, + 78, + 503, + 383 + ], + "score": 0.957, + "type": "image", + "image_path": "4afdd3550ccd39290f23e0e238e9a039d6d3016ff06f8dad0ca88f63a1ea1b20.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 78, + 503, + 179.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 179.66666666666669, + 503, + 281.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 281.33333333333337, + 503, + 383.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 391, + 506, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "score": 1.0, + "content": "Figure 3: Manipulating latent variables on 3D faces: Qualitative results comparing disentangling", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 171, + 415 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 171, + 403, + 178, + 414 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 402, + 205, + 415 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 205, + 402, + 237, + 413 + ], + "score": 0.86, + "content": "\\beta = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 402, + 379, + 415 + ], + "score": 1.0, + "content": "), VAE (Kingma & Welling, 2014)", + "type": "text" + }, + { + "bbox": [ + 380, + 402, + 407, + 414 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\beta = 1 , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "), InfoGAN (Chen et al.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 413, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 506, + 425 + ], + "score": 1.0, + "content": "2016) and DC-IGN (Kulkarni et al., 2015). InfoGAN traversal is over the [-1, 1] range. All models", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "learnt to disentangle lighting (b) and elevation (c). DC-IGN and VAE struggled to continuously", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 325, + 447 + ], + "score": 1.0, + "content": "interpolate between different azimuth angles (a), unlike", + "type": "text" + }, + { + "bbox": [ + 325, + 435, + 333, + 446 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "-VAE, which additionally learnt to encode a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "wider range of azimuth angles than other models. InfoGAN and DC-IGN images adapted from Chen", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 457, + 428, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 428, + 469 + ], + "score": 1.0, + "content": "et al. (2016) and Kulkarni et al. (2015), respectively. Reprinted with permission.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 108, + 487, + 502, + 590 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 487, + 502, + 590 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 487, + 502, + 590 + ], + "spans": [ + { + "bbox": [ + 108, + 487, + 502, + 590 + ], + "score": 0.97, + "type": "image", + "image_path": "40660321d43bd855a3ea6abb7c3408b840648a9ba4a972704a38a38e1104972e.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 487, + 502, + 521.3333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 521.3333333333334, + 502, + 555.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 555.6666666666667, + 502, + 590.0000000000001 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 597, + 504, + 632 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 598, + 504, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 249, + 610 + ], + "score": 1.0, + "content": "Figure 4: Latent factors learnt by", + "type": "text" + }, + { + "bbox": [ + 250, + 598, + 257, + 610 + ], + "score": 0.8, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 598, + 504, + 610 + ], + "score": 1.0, + "content": "-VAE on celebA: traversal of individual latents demonstrates", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 123, + 621 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 609, + 131, + 620 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "-VAE discovered in an unsupervised manner factors that encode skin colour, transition from an", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 620, + 324, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 324, + 632 + ], + "score": 1.0, + "content": "elderly male to younger female, and image saturation.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 645, + 299, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 300, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 124, + 659 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 125, + 645, + 134, + 658 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 644, + 300, + 659 + ], + "score": 1.0, + "content": "-VAE FRAMEWORK DERIVATION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 674, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 674, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 123, + 688 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 675, + 191, + 687 + ], + "score": 0.92, + "content": "\\mathcal { D } = \\{ X , V , W \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 674, + 330, + 688 + ], + "score": 1.0, + "content": "be the set that consists of images", + "type": "text" + }, + { + "bbox": [ + 330, + 675, + 365, + 685 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 674, + 506, + 688 + ], + "score": 1.0, + "content": "and two sets of ground truth data", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 682, + 507, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 317, + 701 + ], + "score": 1.0, + "content": "generative factors: conditionally independent factors", + "type": "text" + }, + { + "bbox": [ + 317, + 686, + 351, + 697 + ], + "score": 0.91, + "content": "\\mathbf { \\bar { v } } \\in \\mathbb { R } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 682, + 381, + 701 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 382, + 686, + 502, + 699 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\log p ( \\mathbf { v } | \\mathbf { x } ) = \\sum _ { k } \\log p ( v _ { k } | \\mathbf { x } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 682, + 507, + 701 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 255, + 712 + ], + "score": 1.0, + "content": "and conditionally dependent factors", + "type": "text" + }, + { + "bbox": [ + 255, + 698, + 292, + 709 + ], + "score": 0.89, + "content": "\\mathbf { w } \\in \\mathbb { R } ^ { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 697, + 410, + 712 + ], + "score": 1.0, + "content": ". We assume that the images", + "type": "text" + }, + { + "bbox": [ + 411, + 701, + 419, + 709 + ], + "score": 0.48, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 697, + 506, + 712 + ], + "score": 1.0, + "content": "are generated by the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 450, + 723 + ], + "score": 1.0, + "content": "true world simulator using the corresponding ground truth data generative factors:", + "type": "text" + }, + { + "bbox": [ + 450, + 709, + 505, + 722 + ], + "score": 0.91, + "content": "p ( \\mathbf { x } | \\mathbf { v } , \\dot { \\mathbf { w } } ) =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 719, + 158, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 153, + 732 + ], + "score": 0.91, + "content": "\\mathbf { S i m } ( \\mathbf { v } , \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 719, + 158, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 78, + 503, + 383 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 78, + 503, + 383 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 78, + 503, + 383 + ], + "spans": [ + { + "bbox": [ + 107, + 78, + 503, + 383 + ], + "score": 0.957, + "type": "image", + "image_path": "4afdd3550ccd39290f23e0e238e9a039d6d3016ff06f8dad0ca88f63a1ea1b20.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 78, + 503, + 179.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 179.66666666666669, + 503, + 281.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 281.33333333333337, + 503, + 383.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 391, + 506, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "score": 1.0, + "content": "Figure 3: Manipulating latent variables on 3D faces: Qualitative results comparing disentangling", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 171, + 415 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 171, + 403, + 178, + 414 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 402, + 205, + 415 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 205, + 402, + 237, + 413 + ], + "score": 0.86, + "content": "\\beta = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 402, + 379, + 415 + ], + "score": 1.0, + "content": "), VAE (Kingma & Welling, 2014)", + "type": "text" + }, + { + "bbox": [ + 380, + 402, + 407, + 414 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\beta = 1 , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "), InfoGAN (Chen et al.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 413, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 506, + 425 + ], + "score": 1.0, + "content": "2016) and DC-IGN (Kulkarni et al., 2015). InfoGAN traversal is over the [-1, 1] range. All models", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "learnt to disentangle lighting (b) and elevation (c). DC-IGN and VAE struggled to continuously", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 325, + 447 + ], + "score": 1.0, + "content": "interpolate between different azimuth angles (a), unlike", + "type": "text" + }, + { + "bbox": [ + 325, + 435, + 333, + 446 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "-VAE, which additionally learnt to encode a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "wider range of azimuth angles than other models. InfoGAN and DC-IGN images adapted from Chen", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 457, + 428, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 428, + 469 + ], + "score": 1.0, + "content": "et al. (2016) and Kulkarni et al. (2015), respectively. Reprinted with permission.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 108, + 487, + 502, + 590 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 487, + 502, + 590 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 487, + 502, + 590 + ], + "spans": [ + { + "bbox": [ + 108, + 487, + 502, + 590 + ], + "score": 0.97, + "type": "image", + "image_path": "40660321d43bd855a3ea6abb7c3408b840648a9ba4a972704a38a38e1104972e.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 487, + 502, + 521.3333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 521.3333333333334, + 502, + 555.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 555.6666666666667, + 502, + 590.0000000000001 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 597, + 504, + 632 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 598, + 504, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 249, + 610 + ], + "score": 1.0, + "content": "Figure 4: Latent factors learnt by", + "type": "text" + }, + { + "bbox": [ + 250, + 598, + 257, + 610 + ], + "score": 0.8, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 598, + 504, + 610 + ], + "score": 1.0, + "content": "-VAE on celebA: traversal of individual latents demonstrates", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 123, + 621 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 609, + 131, + 620 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "-VAE discovered in an unsupervised manner factors that encode skin colour, transition from an", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 620, + 324, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 324, + 632 + ], + "score": 1.0, + "content": "elderly male to younger female, and image saturation.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 645, + 299, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 300, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 124, + 659 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 125, + 645, + 134, + 658 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 644, + 300, + 659 + ], + "score": 1.0, + "content": "-VAE FRAMEWORK DERIVATION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 674, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 674, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 123, + 688 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 675, + 191, + 687 + ], + "score": 0.92, + "content": "\\mathcal { D } = \\{ X , V , W \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 674, + 330, + 688 + ], + "score": 1.0, + "content": "be the set that consists of images", + "type": "text" + }, + { + "bbox": [ + 330, + 675, + 365, + 685 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 674, + 506, + 688 + ], + "score": 1.0, + "content": "and two sets of ground truth data", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 682, + 507, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 317, + 701 + ], + "score": 1.0, + "content": "generative factors: conditionally independent factors", + "type": "text" + }, + { + "bbox": [ + 317, + 686, + 351, + 697 + ], + "score": 0.91, + "content": "\\mathbf { \\bar { v } } \\in \\mathbb { R } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 682, + 381, + 701 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 382, + 686, + 502, + 699 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\log p ( \\mathbf { v } | \\mathbf { x } ) = \\sum _ { k } \\log p ( v _ { k } | \\mathbf { x } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 682, + 507, + 701 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 255, + 712 + ], + "score": 1.0, + "content": "and conditionally dependent factors", + "type": "text" + }, + { + "bbox": [ + 255, + 698, + 292, + 709 + ], + "score": 0.89, + "content": "\\mathbf { w } \\in \\mathbb { R } ^ { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 697, + 410, + 712 + ], + "score": 1.0, + "content": ". We assume that the images", + "type": "text" + }, + { + "bbox": [ + 411, + 701, + 419, + 709 + ], + "score": 0.48, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 697, + 506, + 712 + ], + "score": 1.0, + "content": "are generated by the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 450, + 723 + ], + "score": 1.0, + "content": "true world simulator using the corresponding ground truth data generative factors:", + "type": "text" + }, + { + "bbox": [ + 450, + 709, + 505, + 722 + ], + "score": 0.91, + "content": "p ( \\mathbf { x } | \\mathbf { v } , \\dot { \\mathbf { w } } ) =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 719, + 158, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 153, + 732 + ], + "score": 0.91, + "content": "\\mathbf { S i m } ( \\mathbf { v } , \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 719, + 158, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 674, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 454, + 95 + ], + "score": 1.0, + "content": "We want to develop an unsupervised deep generative model that, using samples from", + "type": "text" + }, + { + "bbox": [ + 454, + 83, + 464, + 92 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "only, can", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 90, + 507, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 90, + 263, + 107 + ], + "score": 1.0, + "content": "learn the joint distribution of the data", + "type": "text" + }, + { + "bbox": [ + 263, + 95, + 272, + 104 + ], + "score": 0.68, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 90, + 425, + 107 + ], + "score": 1.0, + "content": "and a set of generative latent factors", + "type": "text" + }, + { + "bbox": [ + 425, + 96, + 432, + 104 + ], + "score": 0.5, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 93, + 473, + 104 + ], + "score": 0.87, + "content": "{ \\bf \\Psi } ( { \\bf z } \\in \\mathbb { R } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 90, + 507, + 107 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 142, + 115 + ], + "score": 0.88, + "content": "M \\geq K", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 104, + 185, + 117 + ], + "score": 1.0, + "content": ") such that", + "type": "text" + }, + { + "bbox": [ + 185, + 106, + 192, + 114 + ], + "score": 0.56, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 104, + 320, + 117 + ], + "score": 1.0, + "content": "can generate the observed data", + "type": "text" + }, + { + "bbox": [ + 320, + 106, + 327, + 114 + ], + "score": 0.37, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 104, + 361, + 117 + ], + "score": 1.0, + "content": "; that is,", + "type": "text" + }, + { + "bbox": [ + 361, + 105, + 503, + 116 + ], + "score": 0.92, + "content": "p ( \\mathbf { x } | \\mathbf { z } ) \\approx p ( \\mathbf { x } | \\mathbf { v } , \\mathbf { w } ) = \\mathbf { S i m } ( \\mathbf { v } , \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 485, + 128 + ], + "score": 1.0, + "content": "Thus a suitable objective is to maximise the marginal (log-)likelihood of the observed data", + "type": "text" + }, + { + "bbox": [ + 486, + 118, + 493, + 125 + ], + "score": 0.69, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 340, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 329, + 139 + ], + "score": 1.0, + "content": "expectation over the whole distribution of latent factors", + "type": "text" + }, + { + "bbox": [ + 329, + 128, + 336, + 136 + ], + "score": 0.64, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 126, + 340, + 139 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 146, + 348, + 164 + ], + "lines": [ + { + "bbox": [ + 262, + 146, + 348, + 164 + ], + "spans": [ + { + "bbox": [ + 262, + 146, + 348, + 164 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { p _ { \\theta } ( \\mathbf { z } ) } [ p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ]", + "type": "interline_equation", + "image_path": "36556623e132280866cf141f8ad7cdb4fd471175cb7f382c1f51d2253939f743.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 262, + 146, + 348, + 164 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 201, + 185 + ], + "score": 1.0, + "content": "For a given observation", + "type": "text" + }, + { + "bbox": [ + 201, + 175, + 209, + 182 + ], + "score": 0.65, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 172, + 485, + 185 + ], + "score": 1.0, + "content": ", we describe the inferred posterior configurations of the latent factors", + "type": "text" + }, + { + "bbox": [ + 485, + 174, + 492, + 182 + ], + "score": 0.69, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 205, + 196 + ], + "score": 1.0, + "content": "a probability distribution", + "type": "text" + }, + { + "bbox": [ + 205, + 183, + 238, + 195 + ], + "score": 0.93, + "content": "q _ { \\phi } ( { \\bf z } | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 183, + 440, + 196 + ], + "score": 1.0, + "content": ". Our aim is to ensure that the inferred latent factors", + "type": "text" + }, + { + "bbox": [ + 440, + 183, + 473, + 195 + ], + "score": 0.93, + "content": "q _ { \\phi } ( { \\bf z } | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "capture", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 198, + 207 + ], + "score": 1.0, + "content": "the generative factors", + "type": "text" + }, + { + "bbox": [ + 198, + 196, + 206, + 204 + ], + "score": 0.54, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "in a disentangled manner. The conditionally dependent data generative", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 324, + 217 + ], + "score": 1.0, + "content": "factors w can remain entangled in a separate subset of", + "type": "text" + }, + { + "bbox": [ + 324, + 207, + 330, + 215 + ], + "score": 0.72, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 205, + 459, + 217 + ], + "score": 1.0, + "content": "that is not used for representing", + "type": "text" + }, + { + "bbox": [ + 459, + 207, + 467, + 216 + ], + "score": 0.43, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 205, + 505, + 217 + ], + "score": 1.0, + "content": ". In order", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 326, + 229 + ], + "score": 1.0, + "content": "to encourage this disentangling property in the inferred", + "type": "text" + }, + { + "bbox": [ + 327, + 216, + 359, + 228 + ], + "score": 0.93, + "content": "q _ { \\phi } ( { \\bf z } | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 216, + 505, + 229 + ], + "score": 1.0, + "content": ", we introduce a constraint over it by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 213, + 240 + ], + "score": 1.0, + "content": "trying to match it to a prior", + "type": "text" + }, + { + "bbox": [ + 213, + 227, + 232, + 239 + ], + "score": 0.92, + "content": "p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "that can both control the capacity of the latent information bottleneck,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "score": 1.0, + "content": "and embodies the desiderata of statistical independence mentioned above. This can be achieved if", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 299, + 262 + ], + "score": 1.0, + "content": "we set the prior to be an isotropic unit Gaussian", + "type": "text" + }, + { + "bbox": [ + 300, + 249, + 369, + 261 + ], + "score": 0.91, + "content": "( p ( \\mathbf { z } ) = \\mathcal { N } ( \\mathbf { 0 } , I ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 249, + 506, + 262 + ], + "score": 1.0, + "content": ", hence arriving at the constrained", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 452, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 257, + 273 + ], + "score": 1.0, + "content": "optimisation problem in Eq. 2, where", + "type": "text" + }, + { + "bbox": [ + 258, + 262, + 263, + 270 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 260, + 452, + 273 + ], + "score": 1.0, + "content": "specifies the strength of the applied constraint.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 279, + 456, + 300 + ], + "lines": [ + { + "bbox": [ + 155, + 279, + 456, + 300 + ], + "spans": [ + { + "bbox": [ + 155, + 279, + 456, + 300 + ], + "score": 0.89, + "content": "\\operatorname* { m a x } _ { \\phi , \\theta } \\mathbb { E } _ { x \\sim \\mathbf { D } } \\left[ \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ] \\right] \\quad \\mathrm { ~ s u b j e c t ~ t o ~ } D _ { K L } ( q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) ) < \\epsilon", + "type": "interline_equation", + "image_path": "33895b8d80dbcb06fb8743b687e0217c351fe24e0185e9f10719a72c5ff5f7a9.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 155, + 279, + 456, + 300 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 507, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "Re-writing Eq. 2 as a Lagrangian under the KKT conditions (Kuhn & Tucker, 1951; Karush, 1939),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 316, + 152, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 152, + 331 + ], + "score": 1.0, + "content": "we obtain:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 160, + 333, + 451, + 348 + ], + "lines": [ + { + "bbox": [ + 160, + 333, + 451, + 348 + ], + "spans": [ + { + "bbox": [ + 160, + 333, + 451, + 348 + ], + "score": 0.88, + "content": "\\mathcal { F } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } , \\boldsymbol { \\beta } ; \\mathbf { x } , \\mathbf { z } ) = \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } | \\mathbf { z } ) ] - \\beta \\left( D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) ) - \\epsilon \\right)", + "type": "interline_equation", + "image_path": "c002cfea42f7bb618492e31160968e8801fbc12bbbf7cd2fb1aab463571a4957.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 160, + 333, + 451, + 348 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 212, + 367 + ], + "score": 1.0, + "content": "where the KKT multiplier", + "type": "text" + }, + { + "bbox": [ + 212, + 355, + 219, + 367 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "is the regularisation coefficient that constrains the capacity of the latent", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 365, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 191, + 379 + ], + "score": 1.0, + "content": "information channel", + "type": "text" + }, + { + "bbox": [ + 192, + 368, + 199, + 376 + ], + "score": 0.67, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 365, + 505, + 379 + ], + "score": 1.0, + "content": "and puts implicit independence pressure on the learnt posterior due to the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 254, + 389 + ], + "score": 1.0, + "content": "isotropic nature of the Gaussian prior", + "type": "text" + }, + { + "bbox": [ + 255, + 377, + 273, + 389 + ], + "score": 0.92, + "content": "p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 378, + 302, + 389 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 302, + 377, + 336, + 388 + ], + "score": 0.92, + "content": "\\beta , \\epsilon \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "according to the complementary slackness", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 319, + 399 + ], + "score": 1.0, + "content": "KKT condition, Eq. 3 can be re-written to arrive at the", + "type": "text" + }, + { + "bbox": [ + 319, + 389, + 326, + 399 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 388, + 505, + 399 + ], + "score": 1.0, + "content": "-VAE formulation - as the familiar variational", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 399, + 504, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 496, + 411 + ], + "score": 1.0, + "content": "free energy objective function as described by Jordan et al. (1999), but with the addition of the", + "type": "text" + }, + { + "bbox": [ + 497, + 401, + 504, + 411 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 410, + 154, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 154, + 421 + ], + "score": 1.0, + "content": "coefficient:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 135, + 429, + 475, + 444 + ], + "lines": [ + { + "bbox": [ + 135, + 429, + 475, + 444 + ], + "spans": [ + { + "bbox": [ + 135, + 429, + 475, + 444 + ], + "score": 0.9, + "content": "\\mathcal { F } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } , \\boldsymbol { \\beta } ; \\mathbf { x } , \\mathbf { z } ) \\geq \\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\boldsymbol { \\beta } ) = \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } | \\mathbf { z } ) ] - \\beta D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) )", + "type": "interline_equation", + "image_path": "29d1949935f8919270e59c8ac83451224e585288bd1d3c7ad51f0440accb0000.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 135, + 429, + 475, + 444 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 450, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 139, + 465 + ], + "score": 1.0, + "content": "Varying", + "type": "text" + }, + { + "bbox": [ + 139, + 452, + 147, + 463 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "changes the degree of applied learning pressure during training, thus encouraging different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 196, + 475 + ], + "score": 1.0, + "content": "learnt representations.", + "type": "text" + }, + { + "bbox": [ + 196, + 463, + 203, + 474 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 462, + 253, + 475 + ], + "score": 1.0, + "content": "-VAE where", + "type": "text" + }, + { + "bbox": [ + 253, + 463, + 279, + 474 + ], + "score": 0.93, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "corresponds to the original VAE formulation of (Kingma", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "& Welling, 2014). We postulate that in order to learn disentangled representations of the conditionally", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 249, + 498 + ], + "score": 1.0, + "content": "independent data generative factors", + "type": "text" + }, + { + "bbox": [ + 249, + 487, + 257, + 495 + ], + "score": 0.54, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 484, + 342, + 498 + ], + "score": 1.0, + "content": ", it is important to set", + "type": "text" + }, + { + "bbox": [ + 343, + 485, + 368, + 496 + ], + "score": 0.91, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 484, + 506, + 498 + ], + "score": 1.0, + "content": ", thus putting a stronger constraint", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "on the latent bottleneck than in the original VAE formulation of Kingma & Welling (2014). These", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 232, + 519 + ], + "score": 1.0, + "content": "constraints limit the capacity of", + "type": "text" + }, + { + "bbox": [ + 233, + 508, + 239, + 516 + ], + "score": 0.66, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 506, + 506, + 519 + ], + "score": 1.0, + "content": ", which, combined with the pressure to maximise the log likelihood", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 189, + 530 + ], + "score": 1.0, + "content": "of the training data", + "type": "text" + }, + { + "bbox": [ + 190, + 519, + 198, + 528 + ], + "score": 0.51, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "under the model, should encourage the model to learn the most efficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 267, + 541 + ], + "score": 1.0, + "content": "representation of the data. Since the data", + "type": "text" + }, + { + "bbox": [ + 267, + 530, + 275, + 538 + ], + "score": 0.64, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "is generated using at least some conditionally independent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 189, + 552 + ], + "score": 1.0, + "content": "ground truth factors", + "type": "text" + }, + { + "bbox": [ + 189, + 541, + 196, + 549 + ], + "score": 0.39, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 539, + 232, + 552 + ], + "score": 1.0, + "content": ", and the", + "type": "text" + }, + { + "bbox": [ + 232, + 540, + 254, + 550 + ], + "score": 0.91, + "content": "D _ { K L }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 539, + 302, + 552 + ], + "score": 1.0, + "content": "term of the", + "type": "text" + }, + { + "bbox": [ + 303, + 540, + 310, + 551 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "-VAE objective function encourages conditional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 549, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 177, + 564 + ], + "score": 1.0, + "content": "independence in", + "type": "text" + }, + { + "bbox": [ + 178, + 550, + 210, + 563 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 549, + 374, + 564 + ], + "score": 1.0, + "content": ", we hypothesise that higher values of", + "type": "text" + }, + { + "bbox": [ + 374, + 550, + 382, + 562 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 549, + 506, + 564 + ], + "score": 1.0, + "content": "should encourage learning a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 559, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 232, + 576 + ], + "score": 1.0, + "content": "disentangled representation of", + "type": "text" + }, + { + "bbox": [ + 233, + 563, + 240, + 571 + ], + "score": 0.4, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 559, + 405, + 576 + ], + "score": 1.0, + "content": ". The extra pressures coming from high", + "type": "text" + }, + { + "bbox": [ + 405, + 562, + 413, + 573 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 559, + 506, + 576 + ], + "score": 1.0, + "content": "values, however, may", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "create a trade-off between reconstruction fidelity and the quality of disentanglement within the learnt", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "latent representations. Disentangled representations emerge when the right balance is found between", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "information preservation (reconstruction cost as regularisation) and latent channel capacity restriction", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 109, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 109, + 605, + 137, + 616 + ], + "score": 0.85, + "content": "( \\beta > 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "). The latter can lead to poorer reconstructions due to the loss of high frequency details when", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "passing through a constrained latent bottleneck. Hence, the log likelihood of the data under the learnt", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 327, + 640 + ], + "score": 1.0, + "content": "model is a poor metric for evaluating disentangling in", + "type": "text" + }, + { + "bbox": [ + 327, + 627, + 335, + 639 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "-VAEs. Instead we propose a quantitative", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 639, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 482, + 650 + ], + "score": 1.0, + "content": "metric that directly measures the degree of learnt disentanglement in the latent representation.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 248, + 667 + ], + "score": 1.0, + "content": "Since our proposed hyperparameter", + "type": "text" + }, + { + "bbox": [ + 248, + 655, + 256, + 666 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "directly affects the degree of learnt disentanglement, we would", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 215, + 678 + ], + "score": 1.0, + "content": "like to estimate the optimal", + "type": "text" + }, + { + "bbox": [ + 216, + 666, + 223, + 677 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "for learning a disentangled latent representation directly. However, it is", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 306, + 689 + ], + "score": 1.0, + "content": "not possible to do so. This is because the optimal", + "type": "text" + }, + { + "bbox": [ + 306, + 679, + 313, + 689 + ], + "score": 0.87, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 677, + 425, + 689 + ], + "score": 1.0, + "content": "will depend on the value of", + "type": "text" + }, + { + "bbox": [ + 425, + 679, + 431, + 687 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "in Eq.2. Different", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 431, + 699 + ], + "score": 1.0, + "content": "datasets and different model architectures will require different optimal values of", + "type": "text" + }, + { + "bbox": [ + 431, + 690, + 436, + 698 + ], + "score": 0.38, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 688, + 505, + 699 + ], + "score": 1.0, + "content": ". However, when", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 151, + 711 + ], + "score": 1.0, + "content": "optimising", + "type": "text" + }, + { + "bbox": [ + 152, + 699, + 159, + 711 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 699, + 331, + 711 + ], + "score": 1.0, + "content": "in Eq. 4, we are indirectly also optimising", + "type": "text" + }, + { + "bbox": [ + 331, + 704, + 336, + 708 + ], + "score": 0.83, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "for the best disentanglement (see Sec.A.7", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 353, + 722 + ], + "score": 1.0, + "content": "for details), and while we can not learn the optimal value of", + "type": "text" + }, + { + "bbox": [ + 354, + 712, + 360, + 721 + ], + "score": 0.88, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "directly, we can instead estimate it", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "using either our proposed disentanglement metric (see Sec. 3) or through visual inspection heuristics.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 454, + 95 + ], + "score": 1.0, + "content": "We want to develop an unsupervised deep generative model that, using samples from", + "type": "text" + }, + { + "bbox": [ + 454, + 83, + 464, + 92 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "only, can", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 90, + 507, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 90, + 263, + 107 + ], + "score": 1.0, + "content": "learn the joint distribution of the data", + "type": "text" + }, + { + "bbox": [ + 263, + 95, + 272, + 104 + ], + "score": 0.68, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 90, + 425, + 107 + ], + "score": 1.0, + "content": "and a set of generative latent factors", + "type": "text" + }, + { + "bbox": [ + 425, + 96, + 432, + 104 + ], + "score": 0.5, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 93, + 473, + 104 + ], + "score": 0.87, + "content": "{ \\bf \\Psi } ( { \\bf z } \\in \\mathbb { R } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 90, + 507, + 107 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 142, + 115 + ], + "score": 0.88, + "content": "M \\geq K", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 104, + 185, + 117 + ], + "score": 1.0, + "content": ") such that", + "type": "text" + }, + { + "bbox": [ + 185, + 106, + 192, + 114 + ], + "score": 0.56, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 104, + 320, + 117 + ], + "score": 1.0, + "content": "can generate the observed data", + "type": "text" + }, + { + "bbox": [ + 320, + 106, + 327, + 114 + ], + "score": 0.37, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 104, + 361, + 117 + ], + "score": 1.0, + "content": "; that is,", + "type": "text" + }, + { + "bbox": [ + 361, + 105, + 503, + 116 + ], + "score": 0.92, + "content": "p ( \\mathbf { x } | \\mathbf { z } ) \\approx p ( \\mathbf { x } | \\mathbf { v } , \\mathbf { w } ) = \\mathbf { S i m } ( \\mathbf { v } , \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 485, + 128 + ], + "score": 1.0, + "content": "Thus a suitable objective is to maximise the marginal (log-)likelihood of the observed data", + "type": "text" + }, + { + "bbox": [ + 486, + 118, + 493, + 125 + ], + "score": 0.69, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 340, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 329, + 139 + ], + "score": 1.0, + "content": "expectation over the whole distribution of latent factors", + "type": "text" + }, + { + "bbox": [ + 329, + 128, + 336, + 136 + ], + "score": 0.64, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 126, + 340, + 139 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 104, + 81, + 507, + 139 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 146, + 348, + 164 + ], + "lines": [ + { + "bbox": [ + 262, + 146, + 348, + 164 + ], + "spans": [ + { + "bbox": [ + 262, + 146, + 348, + 164 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { \\theta } \\mathbb { E } _ { p _ { \\theta } ( \\mathbf { z } ) } [ p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ]", + "type": "interline_equation", + "image_path": "36556623e132280866cf141f8ad7cdb4fd471175cb7f382c1f51d2253939f743.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 262, + 146, + 348, + 164 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 201, + 185 + ], + "score": 1.0, + "content": "For a given observation", + "type": "text" + }, + { + "bbox": [ + 201, + 175, + 209, + 182 + ], + "score": 0.65, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 172, + 485, + 185 + ], + "score": 1.0, + "content": ", we describe the inferred posterior configurations of the latent factors", + "type": "text" + }, + { + "bbox": [ + 485, + 174, + 492, + 182 + ], + "score": 0.69, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 205, + 196 + ], + "score": 1.0, + "content": "a probability distribution", + "type": "text" + }, + { + "bbox": [ + 205, + 183, + 238, + 195 + ], + "score": 0.93, + "content": "q _ { \\phi } ( { \\bf z } | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 183, + 440, + 196 + ], + "score": 1.0, + "content": ". Our aim is to ensure that the inferred latent factors", + "type": "text" + }, + { + "bbox": [ + 440, + 183, + 473, + 195 + ], + "score": 0.93, + "content": "q _ { \\phi } ( { \\bf z } | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "capture", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 198, + 207 + ], + "score": 1.0, + "content": "the generative factors", + "type": "text" + }, + { + "bbox": [ + 198, + 196, + 206, + 204 + ], + "score": 0.54, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "in a disentangled manner. The conditionally dependent data generative", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 324, + 217 + ], + "score": 1.0, + "content": "factors w can remain entangled in a separate subset of", + "type": "text" + }, + { + "bbox": [ + 324, + 207, + 330, + 215 + ], + "score": 0.72, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 205, + 459, + 217 + ], + "score": 1.0, + "content": "that is not used for representing", + "type": "text" + }, + { + "bbox": [ + 459, + 207, + 467, + 216 + ], + "score": 0.43, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 205, + 505, + 217 + ], + "score": 1.0, + "content": ". In order", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 326, + 229 + ], + "score": 1.0, + "content": "to encourage this disentangling property in the inferred", + "type": "text" + }, + { + "bbox": [ + 327, + 216, + 359, + 228 + ], + "score": 0.93, + "content": "q _ { \\phi } ( { \\bf z } | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 216, + 505, + 229 + ], + "score": 1.0, + "content": ", we introduce a constraint over it by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 213, + 240 + ], + "score": 1.0, + "content": "trying to match it to a prior", + "type": "text" + }, + { + "bbox": [ + 213, + 227, + 232, + 239 + ], + "score": 0.92, + "content": "p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "that can both control the capacity of the latent information bottleneck,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 250 + ], + "score": 1.0, + "content": "and embodies the desiderata of statistical independence mentioned above. This can be achieved if", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 299, + 262 + ], + "score": 1.0, + "content": "we set the prior to be an isotropic unit Gaussian", + "type": "text" + }, + { + "bbox": [ + 300, + 249, + 369, + 261 + ], + "score": 0.91, + "content": "( p ( \\mathbf { z } ) = \\mathcal { N } ( \\mathbf { 0 } , I ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 249, + 506, + 262 + ], + "score": 1.0, + "content": ", hence arriving at the constrained", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 452, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 257, + 273 + ], + "score": 1.0, + "content": "optimisation problem in Eq. 2, where", + "type": "text" + }, + { + "bbox": [ + 258, + 262, + 263, + 270 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 260, + 452, + 273 + ], + "score": 1.0, + "content": "specifies the strength of the applied constraint.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 172, + 506, + 273 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 279, + 456, + 300 + ], + "lines": [ + { + "bbox": [ + 155, + 279, + 456, + 300 + ], + "spans": [ + { + "bbox": [ + 155, + 279, + 456, + 300 + ], + "score": 0.89, + "content": "\\operatorname* { m a x } _ { \\phi , \\theta } \\mathbb { E } _ { x \\sim \\mathbf { D } } \\left[ \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ] \\right] \\quad \\mathrm { ~ s u b j e c t ~ t o ~ } D _ { K L } ( q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) ) < \\epsilon", + "type": "interline_equation", + "image_path": "33895b8d80dbcb06fb8743b687e0217c351fe24e0185e9f10719a72c5ff5f7a9.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 155, + 279, + 456, + 300 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 507, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "Re-writing Eq. 2 as a Lagrangian under the KKT conditions (Kuhn & Tucker, 1951; Karush, 1939),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 316, + 152, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 152, + 331 + ], + "score": 1.0, + "content": "we obtain:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 306, + 506, + 331 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 160, + 333, + 451, + 348 + ], + "lines": [ + { + "bbox": [ + 160, + 333, + 451, + 348 + ], + "spans": [ + { + "bbox": [ + 160, + 333, + 451, + 348 + ], + "score": 0.88, + "content": "\\mathcal { F } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } , \\boldsymbol { \\beta } ; \\mathbf { x } , \\mathbf { z } ) = \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } | \\mathbf { z } ) ] - \\beta \\left( D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) ) - \\epsilon \\right)", + "type": "interline_equation", + "image_path": "c002cfea42f7bb618492e31160968e8801fbc12bbbf7cd2fb1aab463571a4957.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 160, + 333, + 451, + 348 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 212, + 367 + ], + "score": 1.0, + "content": "where the KKT multiplier", + "type": "text" + }, + { + "bbox": [ + 212, + 355, + 219, + 367 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "is the regularisation coefficient that constrains the capacity of the latent", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 365, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 191, + 379 + ], + "score": 1.0, + "content": "information channel", + "type": "text" + }, + { + "bbox": [ + 192, + 368, + 199, + 376 + ], + "score": 0.67, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 365, + 505, + 379 + ], + "score": 1.0, + "content": "and puts implicit independence pressure on the learnt posterior due to the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 254, + 389 + ], + "score": 1.0, + "content": "isotropic nature of the Gaussian prior", + "type": "text" + }, + { + "bbox": [ + 255, + 377, + 273, + 389 + ], + "score": 0.92, + "content": "p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 378, + 302, + 389 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 302, + 377, + 336, + 388 + ], + "score": 0.92, + "content": "\\beta , \\epsilon \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "according to the complementary slackness", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 319, + 399 + ], + "score": 1.0, + "content": "KKT condition, Eq. 3 can be re-written to arrive at the", + "type": "text" + }, + { + "bbox": [ + 319, + 389, + 326, + 399 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 388, + 505, + 399 + ], + "score": 1.0, + "content": "-VAE formulation - as the familiar variational", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 399, + 504, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 496, + 411 + ], + "score": 1.0, + "content": "free energy objective function as described by Jordan et al. (1999), but with the addition of the", + "type": "text" + }, + { + "bbox": [ + 497, + 401, + 504, + 411 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 410, + 154, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 154, + 421 + ], + "score": 1.0, + "content": "coefficient:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 354, + 506, + 421 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 135, + 429, + 475, + 444 + ], + "lines": [ + { + "bbox": [ + 135, + 429, + 475, + 444 + ], + "spans": [ + { + "bbox": [ + 135, + 429, + 475, + 444 + ], + "score": 0.9, + "content": "\\mathcal { F } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } , \\boldsymbol { \\beta } ; \\mathbf { x } , \\mathbf { z } ) \\geq \\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\boldsymbol { \\beta } ) = \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } | \\mathbf { z } ) ] - \\beta D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) )", + "type": "interline_equation", + "image_path": "29d1949935f8919270e59c8ac83451224e585288bd1d3c7ad51f0440accb0000.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 135, + 429, + 475, + 444 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 450, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 139, + 465 + ], + "score": 1.0, + "content": "Varying", + "type": "text" + }, + { + "bbox": [ + 139, + 452, + 147, + 463 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "changes the degree of applied learning pressure during training, thus encouraging different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 196, + 475 + ], + "score": 1.0, + "content": "learnt representations.", + "type": "text" + }, + { + "bbox": [ + 196, + 463, + 203, + 474 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 462, + 253, + 475 + ], + "score": 1.0, + "content": "-VAE where", + "type": "text" + }, + { + "bbox": [ + 253, + 463, + 279, + 474 + ], + "score": 0.93, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "corresponds to the original VAE formulation of (Kingma", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "& Welling, 2014). We postulate that in order to learn disentangled representations of the conditionally", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 249, + 498 + ], + "score": 1.0, + "content": "independent data generative factors", + "type": "text" + }, + { + "bbox": [ + 249, + 487, + 257, + 495 + ], + "score": 0.54, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 484, + 342, + 498 + ], + "score": 1.0, + "content": ", it is important to set", + "type": "text" + }, + { + "bbox": [ + 343, + 485, + 368, + 496 + ], + "score": 0.91, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 484, + 506, + 498 + ], + "score": 1.0, + "content": ", thus putting a stronger constraint", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "on the latent bottleneck than in the original VAE formulation of Kingma & Welling (2014). These", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 232, + 519 + ], + "score": 1.0, + "content": "constraints limit the capacity of", + "type": "text" + }, + { + "bbox": [ + 233, + 508, + 239, + 516 + ], + "score": 0.66, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 506, + 506, + 519 + ], + "score": 1.0, + "content": ", which, combined with the pressure to maximise the log likelihood", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 189, + 530 + ], + "score": 1.0, + "content": "of the training data", + "type": "text" + }, + { + "bbox": [ + 190, + 519, + 198, + 528 + ], + "score": 0.51, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "under the model, should encourage the model to learn the most efficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 267, + 541 + ], + "score": 1.0, + "content": "representation of the data. Since the data", + "type": "text" + }, + { + "bbox": [ + 267, + 530, + 275, + 538 + ], + "score": 0.64, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "is generated using at least some conditionally independent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 189, + 552 + ], + "score": 1.0, + "content": "ground truth factors", + "type": "text" + }, + { + "bbox": [ + 189, + 541, + 196, + 549 + ], + "score": 0.39, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 539, + 232, + 552 + ], + "score": 1.0, + "content": ", and the", + "type": "text" + }, + { + "bbox": [ + 232, + 540, + 254, + 550 + ], + "score": 0.91, + "content": "D _ { K L }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 539, + 302, + 552 + ], + "score": 1.0, + "content": "term of the", + "type": "text" + }, + { + "bbox": [ + 303, + 540, + 310, + 551 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "-VAE objective function encourages conditional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 549, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 177, + 564 + ], + "score": 1.0, + "content": "independence in", + "type": "text" + }, + { + "bbox": [ + 178, + 550, + 210, + 563 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 549, + 374, + 564 + ], + "score": 1.0, + "content": ", we hypothesise that higher values of", + "type": "text" + }, + { + "bbox": [ + 374, + 550, + 382, + 562 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 549, + 506, + 564 + ], + "score": 1.0, + "content": "should encourage learning a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 559, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 232, + 576 + ], + "score": 1.0, + "content": "disentangled representation of", + "type": "text" + }, + { + "bbox": [ + 233, + 563, + 240, + 571 + ], + "score": 0.4, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 559, + 405, + 576 + ], + "score": 1.0, + "content": ". The extra pressures coming from high", + "type": "text" + }, + { + "bbox": [ + 405, + 562, + 413, + 573 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 559, + 506, + 576 + ], + "score": 1.0, + "content": "values, however, may", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "create a trade-off between reconstruction fidelity and the quality of disentanglement within the learnt", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "latent representations. Disentangled representations emerge when the right balance is found between", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "information preservation (reconstruction cost as regularisation) and latent channel capacity restriction", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 109, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 109, + 605, + 137, + 616 + ], + "score": 0.85, + "content": "( \\beta > 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "). The latter can lead to poorer reconstructions due to the loss of high frequency details when", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "passing through a constrained latent bottleneck. Hence, the log likelihood of the data under the learnt", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 327, + 640 + ], + "score": 1.0, + "content": "model is a poor metric for evaluating disentangling in", + "type": "text" + }, + { + "bbox": [ + 327, + 627, + 335, + 639 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "-VAEs. Instead we propose a quantitative", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 639, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 482, + 650 + ], + "score": 1.0, + "content": "metric that directly measures the degree of learnt disentanglement in the latent representation.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 451, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 248, + 667 + ], + "score": 1.0, + "content": "Since our proposed hyperparameter", + "type": "text" + }, + { + "bbox": [ + 248, + 655, + 256, + 666 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "directly affects the degree of learnt disentanglement, we would", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 215, + 678 + ], + "score": 1.0, + "content": "like to estimate the optimal", + "type": "text" + }, + { + "bbox": [ + 216, + 666, + 223, + 677 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "for learning a disentangled latent representation directly. However, it is", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 306, + 689 + ], + "score": 1.0, + "content": "not possible to do so. This is because the optimal", + "type": "text" + }, + { + "bbox": [ + 306, + 679, + 313, + 689 + ], + "score": 0.87, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 677, + 425, + 689 + ], + "score": 1.0, + "content": "will depend on the value of", + "type": "text" + }, + { + "bbox": [ + 425, + 679, + 431, + 687 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "in Eq.2. Different", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 431, + 699 + ], + "score": 1.0, + "content": "datasets and different model architectures will require different optimal values of", + "type": "text" + }, + { + "bbox": [ + 431, + 690, + 436, + 698 + ], + "score": 0.38, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 688, + 505, + 699 + ], + "score": 1.0, + "content": ". However, when", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 151, + 711 + ], + "score": 1.0, + "content": "optimising", + "type": "text" + }, + { + "bbox": [ + 152, + 699, + 159, + 711 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 699, + 331, + 711 + ], + "score": 1.0, + "content": "in Eq. 4, we are indirectly also optimising", + "type": "text" + }, + { + "bbox": [ + 331, + 704, + 336, + 708 + ], + "score": 0.83, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "for the best disentanglement (see Sec.A.7", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 353, + 722 + ], + "score": 1.0, + "content": "for details), and while we can not learn the optimal value of", + "type": "text" + }, + { + "bbox": [ + 354, + 712, + 360, + 721 + ], + "score": 0.88, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "directly, we can instead estimate it", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "using either our proposed disentanglement metric (see Sec. 3) or through visual inspection heuristics.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 655, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 307, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 307, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 306, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 306, + 264 + ], + "score": 0.966, + "type": "image", + "image_path": "b2c9c3518517d6763a2eaa3691033afb73151e60d49d36e7012763a6045855bb.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 307, + 83.85714285714286 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 83.85714285714286, + 307, + 97.71428571428572 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 97.71428571428572, + 307, + 111.57142857142858 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 111.57142857142858, + 307, + 125.42857142857144 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 125.42857142857144, + 307, + 139.2857142857143 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 139.2857142857143, + 307, + 153.14285714285717 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 153.14285714285717, + 307, + 167.00000000000003 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 167.00000000000003, + 307, + 180.8571428571429 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 180.8571428571429, + 307, + 194.71428571428575 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 194.71428571428575, + 307, + 208.5714285714286 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 208.5714285714286, + 307, + 222.42857142857147 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 222.42857142857147, + 307, + 236.28571428571433 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 236.28571428571433, + 307, + 250.1428571428572 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 250.1428571428572, + 307, + 264.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 80, + 506, + 169 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 324, + 79, + 506, + 92 + ], + "spans": [ + { + "bbox": [ + 324, + 79, + 506, + 92 + ], + "score": 1.0, + "content": "Figure 5: Schematic of the proposed disen-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 91, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 461, + 103 + ], + "score": 1.0, + "content": "tanglement metric: over a batch of", + "type": "text" + }, + { + "bbox": [ + 461, + 92, + 469, + 101 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 91, + 506, + 103 + ], + "score": 1.0, + "content": "samples,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 324, + 102, + 505, + 114 + ], + "score": 1.0, + "content": "each pair of images has a fixed value for one", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 114, + 504, + 125 + ], + "spans": [ + { + "bbox": [ + 324, + 114, + 423, + 124 + ], + "score": 1.0, + "content": "target generative factor", + "type": "text" + }, + { + "bbox": [ + 423, + 115, + 430, + 125 + ], + "score": 0.77, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 114, + 455, + 124 + ], + "score": 1.0, + "content": "(here", + "type": "text" + }, + { + "bbox": [ + 456, + 114, + 504, + 125 + ], + "score": 0.87, + "content": "y = s c a l e )", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 324, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "and differs on all others. A linear classifier", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 323, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 323, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "is then trained to identify the target factor us-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 142, + 507, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 142, + 462, + 161 + ], + "score": 1.0, + "content": "ing the average pairwise difference", + "type": "text" + }, + { + "bbox": [ + 463, + 145, + 480, + 158 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 142, + 507, + 161 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 324, + 157, + 439, + 168 + ], + "spans": [ + { + "bbox": [ + 324, + 157, + 393, + 168 + ], + "score": 1.0, + "content": "latent space over", + "type": "text" + }, + { + "bbox": [ + 393, + 158, + 402, + 167 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 157, + 439, + 168 + ], + "score": 1.0, + "content": "samples.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5 + } + ], + "index": 12.0 + }, + { + "type": "title", + "bbox": [ + 108, + 286, + 272, + 298 + ], + "lines": [ + { + "bbox": [ + 104, + 284, + 274, + 301 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 274, + 301 + ], + "score": 1.0, + "content": "3 DISENTANGLEMENT METRIC", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 325 + ], + "score": 1.0, + "content": "It is important to be able to quantify the level of disentanglement achieved by different models.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 324, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 504, + 335 + ], + "score": 1.0, + "content": "Designing a metric for this, however, is not straightforward. We begin by defining the properties", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 347 + ], + "score": 1.0, + "content": "that we expect a disentangled representation to have. Then we describe our proposed solution for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 345, + 385, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 385, + 358 + ], + "score": 1.0, + "content": "quantifying the presence of such properties in a learnt representation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 362, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 504, + 374 + ], + "score": 1.0, + "content": "As stated above, we assume that the data is generated by a ground truth simulation process which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "uses a number of data generative factors, some of which are conditionally independent, and we also", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "assume that they are interpretable. For example, the simulator might sample independent factors", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "corresponding to object shape, colour and size to generate an image of a small green apple. Because", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 406, + 507, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 507, + 419 + ], + "score": 1.0, + "content": "of the independence property, the simulator can also generate small red apples or big green apples.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "A representation of the data that is disentangled with respect to these generative factors, i.e. which", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "encodes them in separate latents, would enable robust classification even using very simple linear", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 438, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 452 + ], + "score": 1.0, + "content": "classifiers (hence providing interpretability). For example, a classifier that learns a decision boundary", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "that relies on object shape would perform as well when other data generative factors, such as size or", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 461, + 180, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 180, + 473 + ], + "score": 1.0, + "content": "colour, are varied.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 492 + ], + "score": 1.0, + "content": "Note that a representation consisting of independent latents is not necessarily disentangled, according", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "to our desiderata. Independence can readily be achieved by a variety of approaches (such as PCA or", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "ICA) that learn to project the data onto independent bases. Representations learnt by such approaches", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "do not in general align with the data generative factors and hence may lack interpretability. For this", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 104, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "reason, a simple cross-correlation calculation between the inferred latents would not suffice as a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 533, + 205, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 205, + 545 + ], + "score": 1.0, + "content": "disentanglement metric.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "Our proposed disentangling metric, therefore, measures both the independence and interpretability", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "(due to the use of a simple classifier) of the inferred latents. To apply our metric, we run inference", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "on a number of images that are generated by fixing the value of one data generative factor while", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "randomly sampling all others. If the independence and interpretability properties hold for the inferred", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "representations, there will be less variance in the inferred latents that correspond to the fixed generative", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "factor. We use a low capacity linear classifier to identify this factor and report the accuracy value as", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "score": 1.0, + "content": "the final disentanglement metric score. Smaller variance in the latents corresponding to the target", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "factor will make the job of this classifier easier, resulting in a higher score under the metric. See", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "score": 1.0, + "content": "Fig. 5 for a representation of the full process.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 653, + 505, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 261, + 667 + ], + "score": 1.0, + "content": "More formally, we start from a dataset", + "type": "text" + }, + { + "bbox": [ + 261, + 654, + 328, + 666 + ], + "score": 0.92, + "content": "\\mathcal { D } = \\{ X , V , W \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "as described in Sec. 2, assumed to contain a", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 664, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 286, + 678 + ], + "score": 1.0, + "content": "balanced distribution of ground truth factors", + "type": "text" + }, + { + "bbox": [ + 287, + 666, + 314, + 677 + ], + "score": 0.88, + "content": "( \\mathbf { v } , \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 664, + 506, + 678 + ], + "score": 1.0, + "content": ", where images data points are obtained using a", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 676, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 233, + 690 + ], + "score": 1.0, + "content": "ground truth simulator process", + "type": "text" + }, + { + "bbox": [ + 233, + 676, + 299, + 688 + ], + "score": 0.91, + "content": "\\mathbf { x } \\sim \\mathbf { S i m } ( \\mathbf { v } , \\mathbf { \\dot { w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 676, + 507, + 690 + ], + "score": 1.0, + "content": ". We also assume we are given labels identifying a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 687, + 441, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 303, + 700 + ], + "score": 1.0, + "content": "subset of the independent data generative factors", + "type": "text" + }, + { + "bbox": [ + 303, + 687, + 330, + 698 + ], + "score": 0.9, + "content": "\\mathbf { v } \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 688, + 441, + 700 + ], + "score": 1.0, + "content": "for at least some instances.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 53.5 + }, + { + "type": "text", + "bbox": [ + 111, + 698, + 484, + 711 + ], + "lines": [ + { + "bbox": [ + 109, + 697, + 486, + 711 + ], + "spans": [ + { + "bbox": [ + 109, + 697, + 222, + 711 + ], + "score": 1.0, + "content": "We then construct a batch of", + "type": "text" + }, + { + "bbox": [ + 222, + 699, + 231, + 708 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 697, + 263, + 711 + ], + "score": 1.0, + "content": "vectors", + "type": "text" + }, + { + "bbox": [ + 264, + 698, + 281, + 711 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 697, + 486, + 711 + ], + "score": 1.0, + "content": ", to be fed as inputs to a linear classifier as follows:", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 56 + }, + { + "type": "text", + "bbox": [ + 130, + 720, + 385, + 733 + ], + "lines": [ + { + "bbox": [ + 130, + 719, + 385, + 734 + ], + "spans": [ + { + "bbox": [ + 130, + 719, + 207, + 734 + ], + "score": 1.0, + "content": "1. Choose a factor", + "type": "text" + }, + { + "bbox": [ + 208, + 720, + 279, + 732 + ], + "score": 0.91, + "content": "y \\sim U n i f [ 1 . . . K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 719, + 300, + 734 + ], + "score": 1.0, + "content": "(e.g.", + "type": "text" + }, + { + "bbox": [ + 300, + 721, + 342, + 732 + ], + "score": 0.79, + "content": "y = s c a l e", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 719, + 385, + 734 + ], + "score": 1.0, + "content": "in Fig. 5).", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 57 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 307, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 307, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 306, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 306, + 264 + ], + "score": 0.966, + "type": "image", + "image_path": "b2c9c3518517d6763a2eaa3691033afb73151e60d49d36e7012763a6045855bb.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 307, + 83.85714285714286 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 83.85714285714286, + 307, + 97.71428571428572 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 97.71428571428572, + 307, + 111.57142857142858 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 111.57142857142858, + 307, + 125.42857142857144 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 125.42857142857144, + 307, + 139.2857142857143 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 139.2857142857143, + 307, + 153.14285714285717 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 153.14285714285717, + 307, + 167.00000000000003 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 167.00000000000003, + 307, + 180.8571428571429 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 180.8571428571429, + 307, + 194.71428571428575 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 194.71428571428575, + 307, + 208.5714285714286 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 208.5714285714286, + 307, + 222.42857142857147 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 222.42857142857147, + 307, + 236.28571428571433 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 236.28571428571433, + 307, + 250.1428571428572 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 250.1428571428572, + 307, + 264.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 80, + 506, + 169 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 324, + 79, + 506, + 92 + ], + "spans": [ + { + "bbox": [ + 324, + 79, + 506, + 92 + ], + "score": 1.0, + "content": "Figure 5: Schematic of the proposed disen-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 91, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 461, + 103 + ], + "score": 1.0, + "content": "tanglement metric: over a batch of", + "type": "text" + }, + { + "bbox": [ + 461, + 92, + 469, + 101 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 91, + 506, + 103 + ], + "score": 1.0, + "content": "samples,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 324, + 102, + 505, + 114 + ], + "score": 1.0, + "content": "each pair of images has a fixed value for one", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 114, + 504, + 125 + ], + "spans": [ + { + "bbox": [ + 324, + 114, + 423, + 124 + ], + "score": 1.0, + "content": "target generative factor", + "type": "text" + }, + { + "bbox": [ + 423, + 115, + 430, + 125 + ], + "score": 0.77, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 114, + 455, + 124 + ], + "score": 1.0, + "content": "(here", + "type": "text" + }, + { + "bbox": [ + 456, + 114, + 504, + 125 + ], + "score": 0.87, + "content": "y = s c a l e )", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 324, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "and differs on all others. A linear classifier", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 323, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 323, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "is then trained to identify the target factor us-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 142, + 507, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 142, + 462, + 161 + ], + "score": 1.0, + "content": "ing the average pairwise difference", + "type": "text" + }, + { + "bbox": [ + 463, + 145, + 480, + 158 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 142, + 507, + 161 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 324, + 157, + 439, + 168 + ], + "spans": [ + { + "bbox": [ + 324, + 157, + 393, + 168 + ], + "score": 1.0, + "content": "latent space over", + "type": "text" + }, + { + "bbox": [ + 393, + 158, + 402, + 167 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 157, + 439, + 168 + ], + "score": 1.0, + "content": "samples.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5 + } + ], + "index": 12.0 + }, + { + "type": "title", + "bbox": [ + 108, + 286, + 272, + 298 + ], + "lines": [ + { + "bbox": [ + 104, + 284, + 274, + 301 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 274, + 301 + ], + "score": 1.0, + "content": "3 DISENTANGLEMENT METRIC", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 325 + ], + "score": 1.0, + "content": "It is important to be able to quantify the level of disentanglement achieved by different models.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 324, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 504, + 335 + ], + "score": 1.0, + "content": "Designing a metric for this, however, is not straightforward. We begin by defining the properties", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 347 + ], + "score": 1.0, + "content": "that we expect a disentangled representation to have. Then we describe our proposed solution for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 345, + 385, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 385, + 358 + ], + "score": 1.0, + "content": "quantifying the presence of such properties in a learnt representation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 313, + 506, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 362, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 504, + 374 + ], + "score": 1.0, + "content": "As stated above, we assume that the data is generated by a ground truth simulation process which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "uses a number of data generative factors, some of which are conditionally independent, and we also", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "assume that they are interpretable. For example, the simulator might sample independent factors", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "corresponding to object shape, colour and size to generate an image of a small green apple. Because", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 406, + 507, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 507, + 419 + ], + "score": 1.0, + "content": "of the independence property, the simulator can also generate small red apples or big green apples.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "A representation of the data that is disentangled with respect to these generative factors, i.e. which", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "encodes them in separate latents, would enable robust classification even using very simple linear", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 438, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 452 + ], + "score": 1.0, + "content": "classifiers (hence providing interpretability). For example, a classifier that learns a decision boundary", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "that relies on object shape would perform as well when other data generative factors, such as size or", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 461, + 180, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 180, + 473 + ], + "score": 1.0, + "content": "colour, are varied.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 363, + 507, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 492 + ], + "score": 1.0, + "content": "Note that a representation consisting of independent latents is not necessarily disentangled, according", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "to our desiderata. Independence can readily be achieved by a variety of approaches (such as PCA or", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "ICA) that learn to project the data onto independent bases. Representations learnt by such approaches", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "do not in general align with the data generative factors and hence may lack interpretability. For this", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 104, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "reason, a simple cross-correlation calculation between the inferred latents would not suffice as a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 533, + 205, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 205, + 545 + ], + "score": 1.0, + "content": "disentanglement metric.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 476, + 506, + 545 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "Our proposed disentangling metric, therefore, measures both the independence and interpretability", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "(due to the use of a simple classifier) of the inferred latents. To apply our metric, we run inference", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "on a number of images that are generated by fixing the value of one data generative factor while", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "randomly sampling all others. If the independence and interpretability properties hold for the inferred", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "representations, there will be less variance in the inferred latents that correspond to the fixed generative", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "factor. We use a low capacity linear classifier to identify this factor and report the accuracy value as", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "score": 1.0, + "content": "the final disentanglement metric score. Smaller variance in the latents corresponding to the target", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "factor will make the job of this classifier easier, resulting in a higher score under the metric. See", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 289, + 650 + ], + "score": 1.0, + "content": "Fig. 5 for a representation of the full process.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 549, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 653, + 505, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 261, + 667 + ], + "score": 1.0, + "content": "More formally, we start from a dataset", + "type": "text" + }, + { + "bbox": [ + 261, + 654, + 328, + 666 + ], + "score": 0.92, + "content": "\\mathcal { D } = \\{ X , V , W \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "as described in Sec. 2, assumed to contain a", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 664, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 286, + 678 + ], + "score": 1.0, + "content": "balanced distribution of ground truth factors", + "type": "text" + }, + { + "bbox": [ + 287, + 666, + 314, + 677 + ], + "score": 0.88, + "content": "( \\mathbf { v } , \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 664, + 506, + 678 + ], + "score": 1.0, + "content": ", where images data points are obtained using a", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 676, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 233, + 690 + ], + "score": 1.0, + "content": "ground truth simulator process", + "type": "text" + }, + { + "bbox": [ + 233, + 676, + 299, + 688 + ], + "score": 0.91, + "content": "\\mathbf { x } \\sim \\mathbf { S i m } ( \\mathbf { v } , \\mathbf { \\dot { w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 676, + 507, + 690 + ], + "score": 1.0, + "content": ". We also assume we are given labels identifying a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 687, + 441, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 303, + 700 + ], + "score": 1.0, + "content": "subset of the independent data generative factors", + "type": "text" + }, + { + "bbox": [ + 303, + 687, + 330, + 698 + ], + "score": 0.9, + "content": "\\mathbf { v } \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 688, + 441, + 700 + ], + "score": 1.0, + "content": "for at least some instances.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 53.5, + "bbox_fs": [ + 105, + 653, + 507, + 700 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 698, + 484, + 711 + ], + "lines": [ + { + "bbox": [ + 109, + 697, + 486, + 711 + ], + "spans": [ + { + "bbox": [ + 109, + 697, + 222, + 711 + ], + "score": 1.0, + "content": "We then construct a batch of", + "type": "text" + }, + { + "bbox": [ + 222, + 699, + 231, + 708 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 697, + 263, + 711 + ], + "score": 1.0, + "content": "vectors", + "type": "text" + }, + { + "bbox": [ + 264, + 698, + 281, + 711 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 697, + 486, + 711 + ], + "score": 1.0, + "content": ", to be fed as inputs to a linear classifier as follows:", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 56, + "bbox_fs": [ + 109, + 697, + 486, + 711 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 720, + 385, + 733 + ], + "lines": [ + { + "bbox": [ + 130, + 719, + 385, + 734 + ], + "spans": [ + { + "bbox": [ + 130, + 719, + 207, + 734 + ], + "score": 1.0, + "content": "1. Choose a factor", + "type": "text" + }, + { + "bbox": [ + 208, + 720, + 279, + 732 + ], + "score": 0.91, + "content": "y \\sim U n i f [ 1 . . . K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 719, + 300, + 734 + ], + "score": 1.0, + "content": "(e.g.", + "type": "text" + }, + { + "bbox": [ + 300, + 721, + 342, + 732 + ], + "score": 0.79, + "content": "y = s c a l e", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 719, + 385, + 734 + ], + "score": 1.0, + "content": "in Fig. 5).", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 57, + "bbox_fs": [ + 130, + 719, + 385, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 130, + 82, + 246, + 94 + ], + "lines": [ + { + "bbox": [ + 128, + 81, + 247, + 96 + ], + "spans": [ + { + "bbox": [ + 128, + 81, + 200, + 96 + ], + "score": 1.0, + "content": "2. For a batch of", + "type": "text" + }, + { + "bbox": [ + 200, + 83, + 209, + 92 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 81, + 247, + 96 + ], + "score": 1.0, + "content": "samples:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 145, + 99, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 145, + 98, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 145, + 98, + 351, + 113 + ], + "score": 1.0, + "content": "(a) Sample two sets of latent representations,", + "type": "text" + }, + { + "bbox": [ + 351, + 101, + 369, + 111 + ], + "score": 0.86, + "content": "\\mathbf { v } _ { 1 , l }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 98, + 392, + 113 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 392, + 101, + 409, + 111 + ], + "score": 0.87, + "content": "\\mathbf { v } _ { 2 , l }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 98, + 461, + 113 + ], + "score": 1.0, + "content": ", enforcing", + "type": "text" + }, + { + "bbox": [ + 461, + 99, + 505, + 113 + ], + "score": 0.89, + "content": "\\begin{array} { r l } { [ \\mathbf { v } _ { 1 , l } ] _ { k } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 163, + 109, + 417, + 123 + ], + "spans": [ + { + "bbox": [ + 163, + 110, + 190, + 123 + ], + "score": 0.92, + "content": "\\left[ \\mathbf { v } _ { 2 , l } \\right] _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 109, + 201, + 123 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 201, + 110, + 227, + 122 + ], + "score": 0.9, + "content": "k = y", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 109, + 334, + 123 + ], + "score": 1.0, + "content": "(so that the value of factor", + "type": "text" + }, + { + "bbox": [ + 335, + 111, + 360, + 122 + ], + "score": 0.92, + "content": "k = y", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 109, + 417, + 123 + ], + "score": 1.0, + "content": "is kept fixed).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 123, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 145, + 123, + 231, + 137 + ], + "score": 1.0, + "content": "(b) Simulate image", + "type": "text" + }, + { + "bbox": [ + 231, + 123, + 309, + 136 + ], + "score": 0.87, + "content": "\\mathbf { x } _ { 1 , l } \\sim \\mathbf { S i m } ( \\mathbf { v } _ { 1 , 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 123, + 359, + 137 + ], + "score": 1.0, + "content": ", then infer", + "type": "text" + }, + { + "bbox": [ + 359, + 124, + 423, + 137 + ], + "score": 0.92, + "content": "{ \\bf z } _ { 1 , l } = \\mu ( { \\bf x } _ { 1 , l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 123, + 506, + 137 + ], + "score": 1.0, + "content": ", using the encoder", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 163, + 135, + 271, + 147 + ], + "spans": [ + { + "bbox": [ + 163, + 135, + 265, + 147 + ], + "score": 0.89, + "content": "q ( \\mathbf { z } | \\mathbf { x } ) \\sim N \\left( \\mu ( \\mathbf { x } ) , \\sigma ( \\mathbf { x } ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 135, + 271, + 147 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 145, + 275, + 160 + ], + "spans": [ + { + "bbox": [ + 161, + 145, + 253, + 160 + ], + "score": 1.0, + "content": "Repeat the process for", + "type": "text" + }, + { + "bbox": [ + 254, + 149, + 270, + 158 + ], + "score": 0.84, + "content": "\\mathbf { v } _ { 2 , l }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 145, + 275, + 160 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 144, + 159, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 144, + 159, + 257, + 175 + ], + "score": 1.0, + "content": "(c) Compute the difference", + "type": "text" + }, + { + "bbox": [ + 258, + 160, + 334, + 174 + ], + "score": 0.92, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { l } = | \\mathbf { z } _ { 1 , l } - \\mathbf { z } _ { 2 , l } |", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 159, + 506, + 175 + ], + "score": 1.0, + "content": ", the absolute linear difference between the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 162, + 172, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 162, + 172, + 286, + 185 + ], + "score": 1.0, + "content": "inferred latent representations.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 130, + 190, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 123, + 182, + 510, + 209 + ], + "spans": [ + { + "bbox": [ + 123, + 182, + 209, + 209 + ], + "score": 1.0, + "content": "3. Use the average", + "type": "text" + }, + { + "bbox": [ + 210, + 189, + 291, + 205 + ], + "score": 0.95, + "content": "\\begin{array} { r } { \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } = \\frac { 1 } { L } \\sum _ { l = 1 } ^ { L } \\mathbf { z } _ { \\mathrm { d i f f } } ^ { l } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 182, + 334, + 209 + ], + "score": 1.0, + "content": "to predict", + "type": "text" + }, + { + "bbox": [ + 334, + 190, + 371, + 203 + ], + "score": 0.92, + "content": "p ( \\boldsymbol { y } | \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 182, + 403, + 209 + ], + "score": 1.0, + "content": "(again,", + "type": "text" + }, + { + "bbox": [ + 403, + 192, + 446, + 203 + ], + "score": 0.85, + "content": "y = s c a l e", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 182, + 510, + 209 + ], + "score": 1.0, + "content": "in Fig. 5) and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 202, + 424, + 215 + ], + "spans": [ + { + "bbox": [ + 141, + 202, + 424, + 215 + ], + "score": 1.0, + "content": "report the accuracy of this predictor as disentangement metric score.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 224, + 506, + 304 + ], + "lines": [ + { + "bbox": [ + 104, + 222, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 278, + 237 + ], + "score": 1.0, + "content": "The classifier’s goal is to predict the index", + "type": "text" + }, + { + "bbox": [ + 286, + 222, + 506, + 237 + ], + "score": 1.0, + "content": "of the generative factor that was kept fixed for a given", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 229, + 510, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 123, + 247 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 229, + 510, + 253 + ], + "score": 1.0, + "content": ". The accuracy of this classifier over multiple batches is used as our disentanglement metric score.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "We choose a linear classifier with low VC-dimension in order to ensure it has no capacity to perform", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 268 + ], + "score": 1.0, + "content": "nonlinear disentangling by itself. We take differences of two inferred latent vectors to reduce the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 101, + 268, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 101, + 268, + 204, + 299 + ], + "score": 1.0, + "content": "variance in the inputs to ensures that on average", + "type": "text" + }, + { + "bbox": [ + 287, + 268, + 473, + 299 + ], + "score": 1.0, + "content": "educe the conditional dependence on the inputs . See Equations 5 in Appendix A.4 for more d", + "type": "text" + }, + { + "bbox": [ + 473, + 270, + 481, + 278 + ], + "score": 0.54, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 268, + 506, + 299 + ], + "score": 1.0, + "content": ". Thisails of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 204, + 278, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 204, + 278, + 287, + 294 + ], + "score": 0.93, + "content": "\\left[ \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } \\right] _ { y } < \\left[ \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } \\right] _ { \\{ \\backslash y \\} }", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 156, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 156, + 306 + ], + "score": 1.0, + "content": "the process.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 322, + 200, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 201, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 201, + 336 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 376, + 361 + ], + "score": 1.0, + "content": "In this section we first qualitatively demonstrate that our proposed", + "type": "text" + }, + { + "bbox": [ + 376, + 349, + 384, + 360 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "-VAE framework consistently", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "discovers more latent factors and disentangles them in a cleaner fashion that either unmodified VAE", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "(Kingma & Welling, 2014) or state of the art unsupervised (InfoGAN: Chen et al., 2016) and semi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "supervised (DC-IGN: Kulkarni et al., 2015) solutions for disentangled factor learning on a variety", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "of benchmarks. We then quantify and characterise the differences in disentangled factor learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 158, + 416 + ], + "score": 1.0, + "content": "between our", + "type": "text" + }, + { + "bbox": [ + 158, + 403, + 165, + 414 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "-VAE framework and a variety of benchmarks using our proposed new disentangling", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 414, + 137, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 137, + 426 + ], + "score": 1.0, + "content": "metric.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 441, + 253, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 254, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 254, + 455 + ], + "score": 1.0, + "content": "4.1 QUALITATIVE BENCHMARKS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 153, + 475 + ], + "score": 1.0, + "content": "We trained", + "type": "text" + }, + { + "bbox": [ + 153, + 463, + 160, + 474 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "-VAE (see Tbl. 1 for architecture details) on a variety of datasets commonly used to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "evaluate disentangling performance of models: celebA (Liu et al., 2015), chairs (Aubry et al., 2014)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "and faces (Paysan et al., 2009). Figures 1-3 provide a qualitative comparison of the disentangling", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 170, + 508 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 170, + 496, + 177, + 507 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 495, + 228, + 508 + ], + "score": 1.0, + "content": "-VAE, VAE", + "type": "text" + }, + { + "bbox": [ + 229, + 496, + 254, + 507 + ], + "score": 0.87, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 495, + 505, + 508 + ], + "score": 1.0, + "content": ") (Kingma & Welling, 2014), InfoGAN (Chen et al., 2016) and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 505, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 296, + 519 + ], + "score": 1.0, + "content": "DC-IGN (Kulkarni et al., 2015) as appropriate.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 251, + 536 + ], + "score": 1.0, + "content": "It can be seen that across all datasets", + "type": "text" + }, + { + "bbox": [ + 251, + 523, + 259, + 534 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "-VAE is able to automatically discover and learn to disentangle", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 534, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 547 + ], + "score": 1.0, + "content": "all of the factors learnt by the semi-supervised DC-IGN (Kulkarni et al., 2015): azimuth (Fig. 3a,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Fig. 2a), lighting and elevation (Fig. 3b,c)). Often it acts as a more convincing inverse graphics", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 556, + 507, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 507, + 569 + ], + "score": 1.0, + "content": "network than DC-IGN (e.g. Fig. 3a) or InfoGAN (e.g. Fig. 2a, Fig. 1a-c or Fig. 3a). Furthermore,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 176, + 580 + ], + "score": 1.0, + "content": "unlike DC-IGN,", + "type": "text" + }, + { + "bbox": [ + 176, + 568, + 183, + 578 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "-VAE requires no supervision and hence can learn about extra unlabelled data", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 578, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 578, + 507, + 591 + ], + "score": 1.0, + "content": "generative factors that DC-IGN can not learn by design, such as chair width or leg style (Fig. 2b,c).", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 424, + 602 + ], + "score": 1.0, + "content": "The unsupervised InfoGAN (Chen et al., 2016) approach shares this quality with", + "type": "text" + }, + { + "bbox": [ + 424, + 590, + 431, + 600 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "-VAE, and the two", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 599, + 507, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 507, + 614 + ], + "score": 1.0, + "content": "frameworks tend to discover overlapping, but not necessarily identical sets of data generative factors.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 181, + 623 + ], + "score": 1.0, + "content": "For example, both", + "type": "text" + }, + { + "bbox": [ + 181, + 612, + 189, + 622 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "-VAE and InfoGAN (but not DC-IGN) learn about the width of chairs (Fig. 2b).", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 129, + 635 + ], + "score": 1.0, + "content": "Only", + "type": "text" + }, + { + "bbox": [ + 129, + 622, + 136, + 633 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 621, + 474, + 635 + ], + "score": 1.0, + "content": "-VAE, however, learns about the chair leg style (Fig. 2c). It is interesting to note how", + "type": "text" + }, + { + "bbox": [ + 474, + 622, + 481, + 633 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "is able to generate an armchair with a round office chair base, even though such armchairs do not exist", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 324, + 656 + ], + "score": 1.0, + "content": "in the dataset (or, perhaps, reality). Furthermore, only", + "type": "text" + }, + { + "bbox": [ + 324, + 644, + 331, + 655 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "-VAE is able to discover all three factors of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "variation (chair azimuth, width and leg style) within a single model, while InfoGAN learns to allocate", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "its continuous latent variable to either azimuth or width. InfoGAN sometimes discovers factors that", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 114, + 688 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 676, + 451, + 690 + ], + "score": 1.0, + "content": "-VAE does not precisely disentangle, such as the presence of sunglasses in celebA.", + "type": "text" + }, + { + "bbox": [ + 451, + 677, + 458, + 688 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "-VAE does,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "however, discover numerous extra factors such as skin colour, image saturation, and age/gender that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 426, + 712 + ], + "score": 1.0, + "content": "are not reported in the InfoGAN paper (Chen et al., 2016) (Fig. 4). Furthermore,", + "type": "text" + }, + { + "bbox": [ + 426, + 699, + 434, + 710 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "-VAE latents tend", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 708, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 507, + 723 + ], + "score": 1.0, + "content": "to learn a smooth continuous transformation over a wider range of factor values than InfoGAN (e.g.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 721, + 353, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 353, + 734 + ], + "score": 1.0, + "content": "rotation over a wider range of angles as shown in Figs. 1-3a).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 40 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 130, + 82, + 246, + 94 + ], + "lines": [ + { + "bbox": [ + 128, + 81, + 247, + 96 + ], + "spans": [ + { + "bbox": [ + 128, + 81, + 200, + 96 + ], + "score": 1.0, + "content": "2. For a batch of", + "type": "text" + }, + { + "bbox": [ + 200, + 83, + 209, + 92 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 81, + 247, + 96 + ], + "score": 1.0, + "content": "samples:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 128, + 81, + 247, + 96 + ] + }, + { + "type": "list", + "bbox": [ + 145, + 99, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 145, + 98, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 145, + 98, + 351, + 113 + ], + "score": 1.0, + "content": "(a) Sample two sets of latent representations,", + "type": "text" + }, + { + "bbox": [ + 351, + 101, + 369, + 111 + ], + "score": 0.86, + "content": "\\mathbf { v } _ { 1 , l }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 98, + 392, + 113 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 392, + 101, + 409, + 111 + ], + "score": 0.87, + "content": "\\mathbf { v } _ { 2 , l }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 98, + 461, + 113 + ], + "score": 1.0, + "content": ", enforcing", + "type": "text" + }, + { + "bbox": [ + 461, + 99, + 505, + 113 + ], + "score": 0.89, + "content": "\\begin{array} { r l } { [ \\mathbf { v } _ { 1 , l } ] _ { k } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 163, + 109, + 417, + 123 + ], + "spans": [ + { + "bbox": [ + 163, + 110, + 190, + 123 + ], + "score": 0.92, + "content": "\\left[ \\mathbf { v } _ { 2 , l } \\right] _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 109, + 201, + 123 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 201, + 110, + 227, + 122 + ], + "score": 0.9, + "content": "k = y", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 109, + 334, + 123 + ], + "score": 1.0, + "content": "(so that the value of factor", + "type": "text" + }, + { + "bbox": [ + 335, + 111, + 360, + 122 + ], + "score": 0.92, + "content": "k = y", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 109, + 417, + 123 + ], + "score": 1.0, + "content": "is kept fixed).", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 123, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 145, + 123, + 231, + 137 + ], + "score": 1.0, + "content": "(b) Simulate image", + "type": "text" + }, + { + "bbox": [ + 231, + 123, + 309, + 136 + ], + "score": 0.87, + "content": "\\mathbf { x } _ { 1 , l } \\sim \\mathbf { S i m } ( \\mathbf { v } _ { 1 , 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 123, + 359, + 137 + ], + "score": 1.0, + "content": ", then infer", + "type": "text" + }, + { + "bbox": [ + 359, + 124, + 423, + 137 + ], + "score": 0.92, + "content": "{ \\bf z } _ { 1 , l } = \\mu ( { \\bf x } _ { 1 , l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 123, + 506, + 137 + ], + "score": 1.0, + "content": ", using the encoder", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 163, + 135, + 271, + 147 + ], + "spans": [ + { + "bbox": [ + 163, + 135, + 265, + 147 + ], + "score": 0.89, + "content": "q ( \\mathbf { z } | \\mathbf { x } ) \\sim N \\left( \\mu ( \\mathbf { x } ) , \\sigma ( \\mathbf { x } ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 135, + 271, + 147 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 161, + 145, + 275, + 160 + ], + "spans": [ + { + "bbox": [ + 161, + 145, + 253, + 160 + ], + "score": 1.0, + "content": "Repeat the process for", + "type": "text" + }, + { + "bbox": [ + 254, + 149, + 270, + 158 + ], + "score": 0.84, + "content": "\\mathbf { v } _ { 2 , l }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 145, + 275, + 160 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 144, + 159, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 144, + 159, + 257, + 175 + ], + "score": 1.0, + "content": "(c) Compute the difference", + "type": "text" + }, + { + "bbox": [ + 258, + 160, + 334, + 174 + ], + "score": 0.92, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { l } = | \\mathbf { z } _ { 1 , l } - \\mathbf { z } _ { 2 , l } |", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 159, + 506, + 175 + ], + "score": 1.0, + "content": ", the absolute linear difference between the", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 172, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 162, + 172, + 286, + 185 + ], + "score": 1.0, + "content": "inferred latent representations.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + } + ], + "index": 4, + "bbox_fs": [ + 144, + 98, + 506, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 190, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 123, + 182, + 510, + 209 + ], + "spans": [ + { + "bbox": [ + 123, + 182, + 209, + 209 + ], + "score": 1.0, + "content": "3. Use the average", + "type": "text" + }, + { + "bbox": [ + 210, + 189, + 291, + 205 + ], + "score": 0.95, + "content": "\\begin{array} { r } { \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } = \\frac { 1 } { L } \\sum _ { l = 1 } ^ { L } \\mathbf { z } _ { \\mathrm { d i f f } } ^ { l } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 182, + 334, + 209 + ], + "score": 1.0, + "content": "to predict", + "type": "text" + }, + { + "bbox": [ + 334, + 190, + 371, + 203 + ], + "score": 0.92, + "content": "p ( \\boldsymbol { y } | \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 182, + 403, + 209 + ], + "score": 1.0, + "content": "(again,", + "type": "text" + }, + { + "bbox": [ + 403, + 192, + 446, + 203 + ], + "score": 0.85, + "content": "y = s c a l e", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 182, + 510, + 209 + ], + "score": 1.0, + "content": "in Fig. 5) and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 202, + 424, + 215 + ], + "spans": [ + { + "bbox": [ + 141, + 202, + 424, + 215 + ], + "score": 1.0, + "content": "report the accuracy of this predictor as disentangement metric score.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 123, + 182, + 510, + 215 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 224, + 506, + 304 + ], + "lines": [ + { + "bbox": [ + 104, + 222, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 278, + 237 + ], + "score": 1.0, + "content": "The classifier’s goal is to predict the index", + "type": "text" + }, + { + "bbox": [ + 286, + 222, + 506, + 237 + ], + "score": 1.0, + "content": "of the generative factor that was kept fixed for a given", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 229, + 510, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 123, + 247 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 229, + 510, + 253 + ], + "score": 1.0, + "content": ". The accuracy of this classifier over multiple batches is used as our disentanglement metric score.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "We choose a linear classifier with low VC-dimension in order to ensure it has no capacity to perform", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 268 + ], + "score": 1.0, + "content": "nonlinear disentangling by itself. We take differences of two inferred latent vectors to reduce the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 101, + 268, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 101, + 268, + 204, + 299 + ], + "score": 1.0, + "content": "variance in the inputs to ensures that on average", + "type": "text" + }, + { + "bbox": [ + 287, + 268, + 473, + 299 + ], + "score": 1.0, + "content": "educe the conditional dependence on the inputs . See Equations 5 in Appendix A.4 for more d", + "type": "text" + }, + { + "bbox": [ + 473, + 270, + 481, + 278 + ], + "score": 0.54, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 268, + 506, + 299 + ], + "score": 1.0, + "content": ". Thisails of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 204, + 278, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 204, + 278, + 287, + 294 + ], + "score": 0.93, + "content": "\\left[ \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } \\right] _ { y } < \\left[ \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } \\right] _ { \\{ \\backslash y \\} }", + "type": "inline_equation" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 291, + 156, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 156, + 306 + ], + "score": 1.0, + "content": "the process.", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 13, + "bbox_fs": [ + 101, + 222, + 510, + 306 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 322, + 200, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 201, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 201, + 336 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 376, + 361 + ], + "score": 1.0, + "content": "In this section we first qualitatively demonstrate that our proposed", + "type": "text" + }, + { + "bbox": [ + 376, + 349, + 384, + 360 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "-VAE framework consistently", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "discovers more latent factors and disentangles them in a cleaner fashion that either unmodified VAE", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "(Kingma & Welling, 2014) or state of the art unsupervised (InfoGAN: Chen et al., 2016) and semi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "supervised (DC-IGN: Kulkarni et al., 2015) solutions for disentangled factor learning on a variety", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "of benchmarks. We then quantify and characterise the differences in disentangled factor learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 158, + 416 + ], + "score": 1.0, + "content": "between our", + "type": "text" + }, + { + "bbox": [ + 158, + 403, + 165, + 414 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "-VAE framework and a variety of benchmarks using our proposed new disentangling", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 414, + 137, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 137, + 426 + ], + "score": 1.0, + "content": "metric.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 348, + 506, + 426 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 441, + 253, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 254, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 254, + 455 + ], + "score": 1.0, + "content": "4.1 QUALITATIVE BENCHMARKS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 153, + 475 + ], + "score": 1.0, + "content": "We trained", + "type": "text" + }, + { + "bbox": [ + 153, + 463, + 160, + 474 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "-VAE (see Tbl. 1 for architecture details) on a variety of datasets commonly used to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "evaluate disentangling performance of models: celebA (Liu et al., 2015), chairs (Aubry et al., 2014)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "and faces (Paysan et al., 2009). Figures 1-3 provide a qualitative comparison of the disentangling", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 170, + 508 + ], + "score": 1.0, + "content": "performance of", + "type": "text" + }, + { + "bbox": [ + 170, + 496, + 177, + 507 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 495, + 228, + 508 + ], + "score": 1.0, + "content": "-VAE, VAE", + "type": "text" + }, + { + "bbox": [ + 229, + 496, + 254, + 507 + ], + "score": 0.87, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 495, + 505, + 508 + ], + "score": 1.0, + "content": ") (Kingma & Welling, 2014), InfoGAN (Chen et al., 2016) and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 505, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 296, + 519 + ], + "score": 1.0, + "content": "DC-IGN (Kulkarni et al., 2015) as appropriate.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 463, + 505, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 251, + 536 + ], + "score": 1.0, + "content": "It can be seen that across all datasets", + "type": "text" + }, + { + "bbox": [ + 251, + 523, + 259, + 534 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "-VAE is able to automatically discover and learn to disentangle", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 534, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 547 + ], + "score": 1.0, + "content": "all of the factors learnt by the semi-supervised DC-IGN (Kulkarni et al., 2015): azimuth (Fig. 3a,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Fig. 2a), lighting and elevation (Fig. 3b,c)). Often it acts as a more convincing inverse graphics", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 556, + 507, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 507, + 569 + ], + "score": 1.0, + "content": "network than DC-IGN (e.g. Fig. 3a) or InfoGAN (e.g. Fig. 2a, Fig. 1a-c or Fig. 3a). Furthermore,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 176, + 580 + ], + "score": 1.0, + "content": "unlike DC-IGN,", + "type": "text" + }, + { + "bbox": [ + 176, + 568, + 183, + 578 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "-VAE requires no supervision and hence can learn about extra unlabelled data", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 578, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 578, + 507, + 591 + ], + "score": 1.0, + "content": "generative factors that DC-IGN can not learn by design, such as chair width or leg style (Fig. 2b,c).", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 424, + 602 + ], + "score": 1.0, + "content": "The unsupervised InfoGAN (Chen et al., 2016) approach shares this quality with", + "type": "text" + }, + { + "bbox": [ + 424, + 590, + 431, + 600 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "-VAE, and the two", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 599, + 507, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 507, + 614 + ], + "score": 1.0, + "content": "frameworks tend to discover overlapping, but not necessarily identical sets of data generative factors.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 181, + 623 + ], + "score": 1.0, + "content": "For example, both", + "type": "text" + }, + { + "bbox": [ + 181, + 612, + 189, + 622 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "-VAE and InfoGAN (but not DC-IGN) learn about the width of chairs (Fig. 2b).", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 129, + 635 + ], + "score": 1.0, + "content": "Only", + "type": "text" + }, + { + "bbox": [ + 129, + 622, + 136, + 633 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 621, + 474, + 635 + ], + "score": 1.0, + "content": "-VAE, however, learns about the chair leg style (Fig. 2c). It is interesting to note how", + "type": "text" + }, + { + "bbox": [ + 474, + 622, + 481, + 633 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "is able to generate an armchair with a round office chair base, even though such armchairs do not exist", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 324, + 656 + ], + "score": 1.0, + "content": "in the dataset (or, perhaps, reality). Furthermore, only", + "type": "text" + }, + { + "bbox": [ + 324, + 644, + 331, + 655 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "-VAE is able to discover all three factors of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "variation (chair azimuth, width and leg style) within a single model, while InfoGAN learns to allocate", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "its continuous latent variable to either azimuth or width. InfoGAN sometimes discovers factors that", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 114, + 688 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 676, + 451, + 690 + ], + "score": 1.0, + "content": "-VAE does not precisely disentangle, such as the presence of sunglasses in celebA.", + "type": "text" + }, + { + "bbox": [ + 451, + 677, + 458, + 688 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "-VAE does,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "however, discover numerous extra factors such as skin colour, image saturation, and age/gender that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 426, + 712 + ], + "score": 1.0, + "content": "are not reported in the InfoGAN paper (Chen et al., 2016) (Fig. 4). Furthermore,", + "type": "text" + }, + { + "bbox": [ + 426, + 699, + 434, + 710 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "-VAE latents tend", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 708, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 507, + 723 + ], + "score": 1.0, + "content": "to learn a smooth continuous transformation over a wider range of factor values than InfoGAN (e.g.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 721, + 353, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 353, + 734 + ], + "score": 1.0, + "content": "rotation over a wider range of angles as shown in Figs. 1-3a).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 522, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 140, + 95 + ], + "score": 1.0, + "content": "Overall", + "type": "text" + }, + { + "bbox": [ + 140, + 83, + 147, + 94 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "-VAE tends to consistently and robustly discover more latent factors and learn cleaner", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "disentangled representations of them than either InfoGAN or DC-IGN. This holds even on such", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 424, + 118 + ], + "score": 1.0, + "content": "challenging datasets as celebA. Furthermore, unlike InfoGAN and DC-IGN,", + "type": "text" + }, + { + "bbox": [ + 424, + 105, + 432, + 116 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "-VAE requires no", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 403, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 403, + 128 + ], + "score": 1.0, + "content": "design decisions or assumptions about the data, and is very stable to train.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 312, + 145 + ], + "score": 1.0, + "content": "When compared to the unmodified VAE baseline (", + "type": "text" + }, + { + "bbox": [ + 312, + 132, + 338, + 144 + ], + "score": 0.86, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 132, + 342, + 145 + ], + "score": 1.0, + "content": ")", + "type": "text" + }, + { + "bbox": [ + 343, + 132, + 350, + 144 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "-VAE consistently learns significantly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "more disentangled latent representations. For example, when learning about chairs, VAE entangles", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "chair width with leg style (Fig. 2b). When learning about celebA, VAE entangles azimuth with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 166, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 177 + ], + "score": 1.0, + "content": "emotion and gender (Fig. 1a); emotion with hair style, skin colour and identity (Fig. 1b); while the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "VAE fringe latent also codes for baldness and head size (Fig. 1c). Although VAE performs relatively", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "well on the faces dataset, it still struggles to learn a clean representation of azimuth (Fig. 3a). This,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "however, suggests that a continuum of disentanglement quality exists, and it can be traversed by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 139, + 222 + ], + "score": 1.0, + "content": "varying", + "type": "text" + }, + { + "bbox": [ + 139, + 210, + 147, + 221 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 209, + 191, + 222 + ], + "score": 1.0, + "content": "within the", + "type": "text" + }, + { + "bbox": [ + 191, + 210, + 199, + 221 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 209, + 343, + 222 + ], + "score": 1.0, + "content": "-VAE framework. While increasing", + "type": "text" + }, + { + "bbox": [ + 344, + 210, + 351, + 221 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "often leads to better disentanglement,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 507, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 507, + 233 + ], + "score": 1.0, + "content": "it may come at the cost of blurrier reconstructions and losing representations for some factors,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 393, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 393, + 244 + ], + "score": 1.0, + "content": "particularly those that correspond to only minor changes in pixel space.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 259, + 258, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 260, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 260, + 272 + ], + "score": 1.0, + "content": "4.2 QUANTITATIVE BENCHMARKS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 507, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 374, + 294 + ], + "score": 1.0, + "content": "In order to quantitatively compare the disentangling performance of", + "type": "text" + }, + { + "bbox": [ + 374, + 282, + 381, + 293 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 280, + 507, + 294 + ], + "score": 1.0, + "content": "-VAE against various baselines,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "score": 1.0, + "content": "we created a synthetic dataset of 737,280 binary 2D shapes (heart, oval and square) generated from", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 419, + 316 + ], + "score": 1.0, + "content": "the Cartesian product of the shape and four independent generative factors", + "type": "text" + }, + { + "bbox": [ + 420, + 304, + 431, + 314 + ], + "score": 0.86, + "content": "v _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "defined in vector", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "graphics: position X (32 values), position Y (32 values), scale (6 values) and rotation (40 values over", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 120, + 338 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 326, + 133, + 335 + ], + "score": 0.84, + "content": "2 \\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "range). To ensure smooth affine object transforms, each two subsequent values for each factor", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 107, + 338, + 117, + 347 + ], + "score": 0.83, + "content": "v _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "were chosen to ensure minimal differences in pixel space given 64x64 pixel image resolution.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "This dataset was chosen because it contains no confounding factors apart from its five independent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "data generative factors (identity, position X, position Y, scale and rotation). This gives us knowledge", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "score": 1.0, + "content": "of the ground truth for comparing the disentangling performance of different models in an objective", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 381, + 141, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 141, + 392 + ], + "score": 1.0, + "content": "manner.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 506, + 409 + ], + "score": 1.0, + "content": "We used our proposed disentanglement metric (see Sec. 3) to quantitatively compare the ability of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 107, + 408, + 114, + 419 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "-VAE to automatically discover and learn a disentangled representation of the data generative factors", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "of the synthetic dataset of 2D shapes described above with that of a number of benchmarks (see", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "Tbl. 1 in Appendix for model architecture details). The table in Fig. 6 (left) reports the classification", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 440, + 453 + ], + "score": 1.0, + "content": "accuracy of the disentanglement metric for 5,000 test samples. It can be seen that", + "type": "text" + }, + { + "bbox": [ + 440, + 441, + 447, + 452 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 441, + 474, + 453 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 475, + 441, + 502, + 452 + ], + "score": 0.84, + "content": "\\beta = 4 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 441, + 505, + 453 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "score": 1.0, + "content": "significantly outperforms all baselines, such as an untrained VAE and the original VAE formulation", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 226, + 475 + ], + "score": 1.0, + "content": "of Kingma & Welling (2014)", + "type": "text" + }, + { + "bbox": [ + 226, + 463, + 254, + 474 + ], + "score": 0.84, + "content": "( \\beta = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 463, + 375, + 475 + ], + "score": 1.0, + "content": ") with the same architecture as", + "type": "text" + }, + { + "bbox": [ + 375, + 463, + 383, + 474 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "-VAE, the top ten PCA or ICA", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 455, + 486 + ], + "score": 1.0, + "content": "components of the data (see Sec. A.3 for details), or when using the raw pixels directly.", + "type": "text" + }, + { + "bbox": [ + 456, + 474, + 463, + 485 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "-VAE also", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 269, + 497 + ], + "score": 1.0, + "content": "does better than InfoGAN. Remarkably,", + "type": "text" + }, + { + "bbox": [ + 269, + 485, + 276, + 496 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "-VAE performs on the same level as DC-IGN despite the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 434, + 508 + ], + "score": 1.0, + "content": "latter being semi-supervised and the former wholly unsupervised. Furthermore,", + "type": "text" + }, + { + "bbox": [ + 435, + 496, + 442, + 507 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "-VAE achieved", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "similar classification accuracy as the ground truth vectors used for data generation, thus suggesting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 516, + 483, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 483, + 531 + ], + "score": 1.0, + "content": "that it was able to learn a very good disentangled representation of the data generative factors.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 534, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 351, + 547 + ], + "score": 1.0, + "content": "We also examined qualitatively the representations learnt by", + "type": "text" + }, + { + "bbox": [ + 351, + 535, + 358, + 546 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "-VAE, VAE, InfoGAN and DC-IGN", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 545, + 504, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 426, + 558 + ], + "score": 1.0, + "content": "on the synthetic dataset of 2D shapes. Fig. 7A demonstrates that after training,", + "type": "text" + }, + { + "bbox": [ + 426, + 546, + 434, + 557 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 545, + 478, + 558 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + }, + { + "bbox": [ + 479, + 546, + 504, + 557 + ], + "score": 0.9, + "content": "\\beta = 4", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "score": 1.0, + "content": "learnt a good (while not perfect) disentangled representation of the data generative factors, and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "score": 1.0, + "content": "its decoder learnt to act as a rendering engine. Its performance was comparative to that of DC-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "IGN (Fig. 7C), with the difference that DC-IGN required a priori knowledge about the quantity", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 588, + 507, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 254, + 602 + ], + "score": 1.0, + "content": "of the data generative factors, while", + "type": "text" + }, + { + "bbox": [ + 254, + 590, + 262, + 600 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 588, + 507, + 602 + ], + "score": 1.0, + "content": "-VAE was able to discover them in an unsupervised manner.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 249, + 613 + ], + "score": 1.0, + "content": "The most informative latent units", + "type": "text" + }, + { + "bbox": [ + 250, + 602, + 263, + 611 + ], + "score": 0.87, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 600, + 277, + 613 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 278, + 600, + 285, + 612 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "-VAE have the highest KL divergence from the unit", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 171, + 624 + ], + "score": 1.0, + "content": "Gaussian prior", + "type": "text" + }, + { + "bbox": [ + 171, + 611, + 245, + 623 + ], + "score": 0.91, + "content": "( p ( z ) = \\mathcal { N } ( 0 , I ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 610, + 506, + 624 + ], + "score": 1.0, + "content": ", while the uninformative latents have KL divergence close to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 620, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 620, + 343, + 636 + ], + "score": 1.0, + "content": "zero. Fig. 7A demonstrates the selectivity of each latent", + "type": "text" + }, + { + "bbox": [ + 344, + 623, + 357, + 633 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 620, + 506, + 636 + ], + "score": 1.0, + "content": "to the independent data generating", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 632, + 507, + 649 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 139, + 649 + ], + "score": 1.0, + "content": "factors:", + "type": "text" + }, + { + "bbox": [ + 140, + 633, + 384, + 645 + ], + "score": 0.84, + "content": "z _ { m } ^ { \\mu } = f ( v _ { k } ) \\forall v _ { k } \\in \\left\\{ v _ { p o s i t i o n X } , v _ { p o s i t i o n Y } , v _ { s c a l e } , v _ { r o t a t i o n } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 632, + 481, + 649 + ], + "score": 1.0, + "content": "(top three rows), where", + "type": "text" + }, + { + "bbox": [ + 481, + 633, + 494, + 645 + ], + "score": 0.9, + "content": "z _ { m } ^ { \\mu }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 632, + 507, + 649 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 264, + 658 + ], + "score": 1.0, + "content": "the learnt Gaussian mean of latent unit", + "type": "text" + }, + { + "bbox": [ + 265, + 646, + 278, + 655 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 642, + 425, + 658 + ], + "score": 1.0, + "content": ". The effect of traversing each latent", + "type": "text" + }, + { + "bbox": [ + 425, + 645, + 438, + 655 + ], + "score": 0.87, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "on the resulting", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 390, + 667 + ], + "score": 1.0, + "content": "reconstructions is shown in the bottom five rows of Fig. 7A. The latents", + "type": "text" + }, + { + "bbox": [ + 390, + 657, + 401, + 666 + ], + "score": 0.86, + "content": "z _ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 655, + 419, + 667 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 419, + 657, + 429, + 666 + ], + "score": 0.85, + "content": "z _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 655, + 495, + 667 + ], + "score": 1.0, + "content": "learnt to encode", + "type": "text" + }, + { + "bbox": [ + 496, + 655, + 505, + 665 + ], + "score": 0.31, + "content": "\\mathrm { X }", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 123, + 678 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 666, + 132, + 676 + ], + "score": 0.34, + "content": "\\mathrm { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 666, + 303, + 678 + ], + "score": 1.0, + "content": "coordinates of the objects respectively; unit", + "type": "text" + }, + { + "bbox": [ + 303, + 667, + 313, + 677 + ], + "score": 0.86, + "content": "z _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 666, + 441, + 678 + ], + "score": 1.0, + "content": "learnt to encode scale; and units", + "type": "text" + }, + { + "bbox": [ + 441, + 667, + 451, + 677 + ], + "score": 0.87, + "content": "z _ { 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 666, + 468, + 678 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 469, + 667, + 479, + 677 + ], + "score": 0.86, + "content": "z _ { 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "learnt", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "to encode rotation. The frequency of oscillations in each rotational latent corresponds to the rotational", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 267, + 701 + ], + "score": 1.0, + "content": "symmetry of the corresponding object", + "type": "text" + }, + { + "bbox": [ + 268, + 689, + 280, + 698 + ], + "score": 0.82, + "content": "2 \\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 687, + 321, + 701 + ], + "score": 1.0, + "content": "for heart,", + "type": "text" + }, + { + "bbox": [ + 321, + 690, + 328, + 698 + ], + "score": 0.71, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 687, + 381, + 701 + ], + "score": 1.0, + "content": "for oval and", + "type": "text" + }, + { + "bbox": [ + 381, + 687, + 399, + 700 + ], + "score": 0.89, + "content": "\\pi / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "for square). Furthermore,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "the two rotational latents seem to encode cos and sin rotational coordinates, while the positional", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "latents align with the Cartesian axes. While such alignment with intuitive factors for humans is not", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "guaranteed, empirically we found it to be very common. Fig. 7B demonstrates that the unmodified", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 45.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 140, + 95 + ], + "score": 1.0, + "content": "Overall", + "type": "text" + }, + { + "bbox": [ + 140, + 83, + 147, + 94 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "-VAE tends to consistently and robustly discover more latent factors and learn cleaner", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "disentangled representations of them than either InfoGAN or DC-IGN. This holds even on such", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 424, + 118 + ], + "score": 1.0, + "content": "challenging datasets as celebA. Furthermore, unlike InfoGAN and DC-IGN,", + "type": "text" + }, + { + "bbox": [ + 424, + 105, + 432, + 116 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "-VAE requires no", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 403, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 403, + 128 + ], + "score": 1.0, + "content": "design decisions or assumptions about the data, and is very stable to train.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 312, + 145 + ], + "score": 1.0, + "content": "When compared to the unmodified VAE baseline (", + "type": "text" + }, + { + "bbox": [ + 312, + 132, + 338, + 144 + ], + "score": 0.86, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 132, + 342, + 145 + ], + "score": 1.0, + "content": ")", + "type": "text" + }, + { + "bbox": [ + 343, + 132, + 350, + 144 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "-VAE consistently learns significantly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "more disentangled latent representations. For example, when learning about chairs, VAE entangles", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "chair width with leg style (Fig. 2b). When learning about celebA, VAE entangles azimuth with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 166, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 177 + ], + "score": 1.0, + "content": "emotion and gender (Fig. 1a); emotion with hair style, skin colour and identity (Fig. 1b); while the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "VAE fringe latent also codes for baldness and head size (Fig. 1c). Although VAE performs relatively", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "well on the faces dataset, it still struggles to learn a clean representation of azimuth (Fig. 3a). This,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "however, suggests that a continuum of disentanglement quality exists, and it can be traversed by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 139, + 222 + ], + "score": 1.0, + "content": "varying", + "type": "text" + }, + { + "bbox": [ + 139, + 210, + 147, + 221 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 209, + 191, + 222 + ], + "score": 1.0, + "content": "within the", + "type": "text" + }, + { + "bbox": [ + 191, + 210, + 199, + 221 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 209, + 343, + 222 + ], + "score": 1.0, + "content": "-VAE framework. While increasing", + "type": "text" + }, + { + "bbox": [ + 344, + 210, + 351, + 221 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "often leads to better disentanglement,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 507, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 507, + 233 + ], + "score": 1.0, + "content": "it may come at the cost of blurrier reconstructions and losing representations for some factors,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 393, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 393, + 244 + ], + "score": 1.0, + "content": "particularly those that correspond to only minor changes in pixel space.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 132, + 507, + 244 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 259, + 258, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 260, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 260, + 272 + ], + "score": 1.0, + "content": "4.2 QUANTITATIVE BENCHMARKS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 507, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 374, + 294 + ], + "score": 1.0, + "content": "In order to quantitatively compare the disentangling performance of", + "type": "text" + }, + { + "bbox": [ + 374, + 282, + 381, + 293 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 280, + 507, + 294 + ], + "score": 1.0, + "content": "-VAE against various baselines,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "score": 1.0, + "content": "we created a synthetic dataset of 737,280 binary 2D shapes (heart, oval and square) generated from", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 419, + 316 + ], + "score": 1.0, + "content": "the Cartesian product of the shape and four independent generative factors", + "type": "text" + }, + { + "bbox": [ + 420, + 304, + 431, + 314 + ], + "score": 0.86, + "content": "v _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "defined in vector", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "graphics: position X (32 values), position Y (32 values), scale (6 values) and rotation (40 values over", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 120, + 338 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 326, + 133, + 335 + ], + "score": 0.84, + "content": "2 \\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "range). To ensure smooth affine object transforms, each two subsequent values for each factor", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 336, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 107, + 338, + 117, + 347 + ], + "score": 0.83, + "content": "v _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 336, + 506, + 349 + ], + "score": 1.0, + "content": "were chosen to ensure minimal differences in pixel space given 64x64 pixel image resolution.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "This dataset was chosen because it contains no confounding factors apart from its five independent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "data generative factors (identity, position X, position Y, scale and rotation). This gives us knowledge", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "score": 1.0, + "content": "of the ground truth for comparing the disentangling performance of different models in an objective", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 381, + 141, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 141, + 392 + ], + "score": 1.0, + "content": "manner.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 280, + 507, + 392 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 506, + 409 + ], + "score": 1.0, + "content": "We used our proposed disentanglement metric (see Sec. 3) to quantitatively compare the ability of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 107, + 408, + 114, + 419 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "-VAE to automatically discover and learn a disentangled representation of the data generative factors", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "of the synthetic dataset of 2D shapes described above with that of a number of benchmarks (see", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "Tbl. 1 in Appendix for model architecture details). The table in Fig. 6 (left) reports the classification", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 440, + 453 + ], + "score": 1.0, + "content": "accuracy of the disentanglement metric for 5,000 test samples. It can be seen that", + "type": "text" + }, + { + "bbox": [ + 440, + 441, + 447, + 452 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 441, + 474, + 453 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 475, + 441, + 502, + 452 + ], + "score": 0.84, + "content": "\\beta = 4 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 441, + 505, + 453 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 464 + ], + "score": 1.0, + "content": "significantly outperforms all baselines, such as an untrained VAE and the original VAE formulation", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 226, + 475 + ], + "score": 1.0, + "content": "of Kingma & Welling (2014)", + "type": "text" + }, + { + "bbox": [ + 226, + 463, + 254, + 474 + ], + "score": 0.84, + "content": "( \\beta = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 463, + 375, + 475 + ], + "score": 1.0, + "content": ") with the same architecture as", + "type": "text" + }, + { + "bbox": [ + 375, + 463, + 383, + 474 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "-VAE, the top ten PCA or ICA", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 455, + 486 + ], + "score": 1.0, + "content": "components of the data (see Sec. A.3 for details), or when using the raw pixels directly.", + "type": "text" + }, + { + "bbox": [ + 456, + 474, + 463, + 485 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "-VAE also", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 269, + 497 + ], + "score": 1.0, + "content": "does better than InfoGAN. Remarkably,", + "type": "text" + }, + { + "bbox": [ + 269, + 485, + 276, + 496 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "-VAE performs on the same level as DC-IGN despite the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 434, + 508 + ], + "score": 1.0, + "content": "latter being semi-supervised and the former wholly unsupervised. Furthermore,", + "type": "text" + }, + { + "bbox": [ + 435, + 496, + 442, + 507 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "-VAE achieved", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "similar classification accuracy as the ground truth vectors used for data generation, thus suggesting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 516, + 483, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 483, + 531 + ], + "score": 1.0, + "content": "that it was able to learn a very good disentangled representation of the data generative factors.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 397, + 506, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 534, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 351, + 547 + ], + "score": 1.0, + "content": "We also examined qualitatively the representations learnt by", + "type": "text" + }, + { + "bbox": [ + 351, + 535, + 358, + 546 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "-VAE, VAE, InfoGAN and DC-IGN", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 545, + 504, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 426, + 558 + ], + "score": 1.0, + "content": "on the synthetic dataset of 2D shapes. Fig. 7A demonstrates that after training,", + "type": "text" + }, + { + "bbox": [ + 426, + 546, + 434, + 557 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 545, + 478, + 558 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + }, + { + "bbox": [ + 479, + 546, + 504, + 557 + ], + "score": 0.9, + "content": "\\beta = 4", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "score": 1.0, + "content": "learnt a good (while not perfect) disentangled representation of the data generative factors, and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "score": 1.0, + "content": "its decoder learnt to act as a rendering engine. Its performance was comparative to that of DC-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "IGN (Fig. 7C), with the difference that DC-IGN required a priori knowledge about the quantity", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 588, + 507, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 254, + 602 + ], + "score": 1.0, + "content": "of the data generative factors, while", + "type": "text" + }, + { + "bbox": [ + 254, + 590, + 262, + 600 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 588, + 507, + 602 + ], + "score": 1.0, + "content": "-VAE was able to discover them in an unsupervised manner.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 249, + 613 + ], + "score": 1.0, + "content": "The most informative latent units", + "type": "text" + }, + { + "bbox": [ + 250, + 602, + 263, + 611 + ], + "score": 0.87, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 600, + 277, + 613 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 278, + 600, + 285, + 612 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "-VAE have the highest KL divergence from the unit", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 171, + 624 + ], + "score": 1.0, + "content": "Gaussian prior", + "type": "text" + }, + { + "bbox": [ + 171, + 611, + 245, + 623 + ], + "score": 0.91, + "content": "( p ( z ) = \\mathcal { N } ( 0 , I ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 610, + 506, + 624 + ], + "score": 1.0, + "content": ", while the uninformative latents have KL divergence close to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 620, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 620, + 343, + 636 + ], + "score": 1.0, + "content": "zero. Fig. 7A demonstrates the selectivity of each latent", + "type": "text" + }, + { + "bbox": [ + 344, + 623, + 357, + 633 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 620, + 506, + 636 + ], + "score": 1.0, + "content": "to the independent data generating", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 632, + 507, + 649 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 139, + 649 + ], + "score": 1.0, + "content": "factors:", + "type": "text" + }, + { + "bbox": [ + 140, + 633, + 384, + 645 + ], + "score": 0.84, + "content": "z _ { m } ^ { \\mu } = f ( v _ { k } ) \\forall v _ { k } \\in \\left\\{ v _ { p o s i t i o n X } , v _ { p o s i t i o n Y } , v _ { s c a l e } , v _ { r o t a t i o n } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 632, + 481, + 649 + ], + "score": 1.0, + "content": "(top three rows), where", + "type": "text" + }, + { + "bbox": [ + 481, + 633, + 494, + 645 + ], + "score": 0.9, + "content": "z _ { m } ^ { \\mu }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 632, + 507, + 649 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 264, + 658 + ], + "score": 1.0, + "content": "the learnt Gaussian mean of latent unit", + "type": "text" + }, + { + "bbox": [ + 265, + 646, + 278, + 655 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 642, + 425, + 658 + ], + "score": 1.0, + "content": ". The effect of traversing each latent", + "type": "text" + }, + { + "bbox": [ + 425, + 645, + 438, + 655 + ], + "score": 0.87, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "on the resulting", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 390, + 667 + ], + "score": 1.0, + "content": "reconstructions is shown in the bottom five rows of Fig. 7A. The latents", + "type": "text" + }, + { + "bbox": [ + 390, + 657, + 401, + 666 + ], + "score": 0.86, + "content": "z _ { 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 655, + 419, + 667 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 419, + 657, + 429, + 666 + ], + "score": 0.85, + "content": "z _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 655, + 495, + 667 + ], + "score": 1.0, + "content": "learnt to encode", + "type": "text" + }, + { + "bbox": [ + 496, + 655, + 505, + 665 + ], + "score": 0.31, + "content": "\\mathrm { X }", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 123, + 678 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 666, + 132, + 676 + ], + "score": 0.34, + "content": "\\mathrm { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 666, + 303, + 678 + ], + "score": 1.0, + "content": "coordinates of the objects respectively; unit", + "type": "text" + }, + { + "bbox": [ + 303, + 667, + 313, + 677 + ], + "score": 0.86, + "content": "z _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 666, + 441, + 678 + ], + "score": 1.0, + "content": "learnt to encode scale; and units", + "type": "text" + }, + { + "bbox": [ + 441, + 667, + 451, + 677 + ], + "score": 0.87, + "content": "z _ { 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 666, + 468, + 678 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 469, + 667, + 479, + 677 + ], + "score": 0.86, + "content": "z _ { 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "learnt", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "to encode rotation. The frequency of oscillations in each rotational latent corresponds to the rotational", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 267, + 701 + ], + "score": 1.0, + "content": "symmetry of the corresponding object", + "type": "text" + }, + { + "bbox": [ + 268, + 689, + 280, + 698 + ], + "score": 0.82, + "content": "2 \\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 687, + 321, + 701 + ], + "score": 1.0, + "content": "for heart,", + "type": "text" + }, + { + "bbox": [ + 321, + 690, + 328, + 698 + ], + "score": 0.71, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 687, + 381, + 701 + ], + "score": 1.0, + "content": "for oval and", + "type": "text" + }, + { + "bbox": [ + 381, + 687, + 399, + 700 + ], + "score": 0.89, + "content": "\\pi / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "for square). Furthermore,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "the two rotational latents seem to encode cos and sin rotational coordinates, while the positional", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "latents align with the Cartesian axes. While such alignment with intuitive factors for humans is not", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "guaranteed, empirically we found it to be very common. Fig. 7B demonstrates that the unmodified", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 45.5, + "bbox_fs": [ + 104, + 534, + 507, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 84, + 289, + 220 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 84, + 289, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 84, + 289, + 220 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 289, + 220 + ], + "score": 0.932, + "html": "
ModelDisentanglementmetric score
Ground truthRaw pixelsPCAICA100%45.75 ± 0.8%84.9 ±0.4%42.03 ± 10.6%
DC-IGNInfoGAN99.3 ± 0.1%73.5 ± 0.9%
VAE untrainedVAEβ-VAE44.14 ± 2.5%61.58 ± 0.5%
61.58 ± 0.5%99.23±0.1%
", + "type": "table", + "image_path": "247142d1273194a9ddfdf0438fd01012c53e69b471dd65d7a1fb3607bf205ec1.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 107, + 84, + 289, + 97.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 97.6, + 289, + 111.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 111.19999999999999, + 289, + 124.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 107, + 124.79999999999998, + 289, + 138.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 107, + 138.39999999999998, + 289, + 151.99999999999997 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 151.99999999999997, + 289, + 165.59999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 165.59999999999997, + 289, + 179.19999999999996 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 179.19999999999996, + 289, + 192.79999999999995 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 107, + 192.79999999999995, + 289, + 206.39999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 107, + 206.39999999999995, + 289, + 219.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "image", + "bbox": [ + 296, + 81, + 500, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 296, + 81, + 500, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 296, + 81, + 500, + 225 + ], + "spans": [ + { + "bbox": [ + 296, + 81, + 500, + 225 + ], + "score": 0.953, + "type": "image", + "image_path": "b161f1f7e0f90af0103ab47c8dc24899548b222d5a87984f5f762ba8e18c80a4.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 296, + 81, + 500, + 94.0909090909091 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 296, + 94.0909090909091, + 500, + 107.18181818181819 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 296, + 107.18181818181819, + 500, + 120.27272727272728 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 296, + 120.27272727272728, + 500, + 133.36363636363637 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 296, + 133.36363636363637, + 500, + 146.45454545454547 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 296, + 146.45454545454547, + 500, + 159.54545454545456 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 296, + 159.54545454545456, + 500, + 172.63636363636365 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 296, + 172.63636363636365, + 500, + 185.72727272727275 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 296, + 185.72727272727275, + 500, + 198.81818181818184 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 296, + 198.81818181818184, + 500, + 211.90909090909093 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 296, + 211.90909090909093, + 500, + 225.00000000000003 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 227, + 506, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "Figure 6: Disentanglement metric classification accuracy for 2D shapes dataset. Left: Accuracy for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 480, + 251 + ], + "score": 1.0, + "content": "different models and training regimes Right: Positive correlation is present between the size of", + "type": "text" + }, + { + "bbox": [ + 481, + 240, + 488, + 248 + ], + "score": 0.61, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 249, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 237, + 262 + ], + "score": 1.0, + "content": "the optimal normalised values of", + "type": "text" + }, + { + "bbox": [ + 238, + 250, + 245, + 261 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 249, + 414, + 262 + ], + "score": 1.0, + "content": "for disentangled factor learning for a fixed", + "type": "text" + }, + { + "bbox": [ + 414, + 250, + 421, + 261 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 249, + 496, + 262 + ], + "score": 1.0, + "content": "-VAE architecture.", + "type": "text" + }, + { + "bbox": [ + 496, + 250, + 504, + 261 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 235, + 273 + ], + "score": 1.0, + "content": "values are normalised by latent", + "type": "text" + }, + { + "bbox": [ + 236, + 262, + 243, + 271 + ], + "score": 0.43, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 260, + 262, + 273 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 262, + 262, + 272, + 271 + ], + "score": 0.65, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 260, + 314, + 273 + ], + "score": 1.0, + "content": "and input", + "type": "text" + }, + { + "bbox": [ + 314, + 262, + 322, + 271 + ], + "score": 0.28, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 260, + 341, + 273 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 342, + 262, + 349, + 270 + ], + "score": 0.61, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 260, + 394, + 273 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + }, + { + "bbox": [ + 394, + 261, + 402, + 272 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "values are not uniformly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 271, + 507, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 352, + 284 + ], + "score": 1.0, + "content": "sampled. Orange approximately corresponds to unnormalised", + "type": "text" + }, + { + "bbox": [ + 353, + 271, + 378, + 282 + ], + "score": 0.91, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 271, + 507, + 284 + ], + "score": 1.0, + "content": ". Good reconstructions are asso-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "ciated with entangled representations (lower disentanglement scores). Disentangled representations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 293, + 379, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 379, + 306 + ], + "score": 1.0, + "content": "(high disentanglement scores) often result in blurry reconstructions.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 165, + 328 + ], + "score": 1.0, + "content": "VAE baseline", + "type": "text" + }, + { + "bbox": [ + 165, + 316, + 192, + 327 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\beta = 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 315, + 454, + 328 + ], + "score": 1.0, + "content": ") is not able to disentangle generative factors in the data as well as", + "type": "text" + }, + { + "bbox": [ + 454, + 316, + 461, + 327 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 305, + 339 + ], + "score": 1.0, + "content": "appropriate learning pressures. Instead each latent", + "type": "text" + }, + { + "bbox": [ + 305, + 328, + 312, + 336 + ], + "score": 0.59, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 326, + 359, + 339 + ], + "score": 1.0, + "content": "(apart from", + "type": "text" + }, + { + "bbox": [ + 360, + 328, + 370, + 338 + ], + "score": 0.84, + "content": "\\mathbf { z } _ { 9 }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 326, + 506, + 339 + ], + "score": 1.0, + "content": ", which learnt rotation) encodes at", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 338, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 506, + 349 + ], + "score": 1.0, + "content": "least two data generative factors. InfoGAN also achieved a degree of disentangling (see Fig. 7D),", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "particularly for positional factors. However, despite our best efforts to train InfoGAN, we were not", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "score": 1.0, + "content": "able to achieve the same degree of disentangling in other factors, such as rotation, scale and shape.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "We also found its ability to generate the different shapes in the dataset to be inaccurate and unstable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "during training, possibly due to reported limitations of the GAN framework, which can struggle to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 392, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 507, + 405 + ], + "score": 1.0, + "content": "learn the full data distribution and instead will often learn a small subset of its modes (Salimans et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 403, + 207, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 207, + 415 + ], + "score": 1.0, + "content": "2016; Zhao et al., 2016).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 230, + 441 + ], + "score": 1.0, + "content": "Understanding the effects of", + "type": "text" + }, + { + "bbox": [ + 230, + 428, + 238, + 439 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "We hypothesised that constrained optimisation is important for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "enabling deep unsupervised models to learn disentangled representations of the independent data", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 238, + 463 + ], + "score": 1.0, + "content": "generative factors (Sec. 2). In the", + "type": "text" + }, + { + "bbox": [ + 239, + 451, + 246, + 461 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 450, + 431, + 463 + ], + "score": 1.0, + "content": "-VAE framework this corresponds to tuning the", + "type": "text" + }, + { + "bbox": [ + 431, + 451, + 439, + 461 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "coefficient. One", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 156, + 474 + ], + "score": 1.0, + "content": "way to view", + "type": "text" + }, + { + "bbox": [ + 156, + 462, + 164, + 473 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "is as a mixing coefficient (see Sec. A.6 for a derivation) for balancing the magnitudes", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "of gradients from the reconstruction and the prior-matching components of the VAE lower bound", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 432, + 496 + ], + "score": 1.0, + "content": "formulation in Eq. 4 during training. In this context it makes sense to normalise", + "type": "text" + }, + { + "bbox": [ + 433, + 483, + 440, + 494 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 483, + 479, + 496 + ], + "score": 1.0, + "content": "by latent", + "type": "text" + }, + { + "bbox": [ + 479, + 485, + 486, + 493 + ], + "score": 0.63, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "size", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 107, + 496, + 117, + 504 + ], + "score": 0.64, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 494, + 160, + 507 + ], + "score": 1.0, + "content": "and input", + "type": "text" + }, + { + "bbox": [ + 161, + 496, + 169, + 504 + ], + "score": 0.59, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 494, + 190, + 507 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 190, + 496, + 198, + 504 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "in order to compare its different values across different latent layer sizes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 103, + 502, + 507, + 522 + ], + "spans": [ + { + "bbox": [ + 103, + 502, + 198, + 522 + ], + "score": 1.0, + "content": "and different datasets", + "type": "text" + }, + { + "bbox": [ + 198, + 505, + 258, + 519 + ], + "score": 0.91, + "content": "\\begin{array} { r } { ( \\beta _ { n o r m } = \\frac { \\beta M } { N } . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 502, + 379, + 522 + ], + "score": 1.0, + "content": "). We found that larger latent", + "type": "text" + }, + { + "bbox": [ + 379, + 508, + 386, + 516 + ], + "score": 0.52, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 502, + 433, + 522 + ], + "score": 1.0, + "content": "layer sizes", + "type": "text" + }, + { + "bbox": [ + 433, + 508, + 443, + 516 + ], + "score": 0.7, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 502, + 507, + 522 + ], + "score": 1.0, + "content": "require higher", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 221, + 531 + ], + "score": 1.0, + "content": "constraint pressures (higher", + "type": "text" + }, + { + "bbox": [ + 222, + 519, + 229, + 529 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 516, + 475, + 531 + ], + "score": 1.0, + "content": "values), see Fig. 6 (Right). Furthermore, the relationship of", + "type": "text" + }, + { + "bbox": [ + 475, + 518, + 482, + 529 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "for a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 130, + 541 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 131, + 530, + 141, + 538 + ], + "score": 0.68, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 528, + 328, + 541 + ], + "score": 1.0, + "content": "is characterised by an inverted U curve. When", + "type": "text" + }, + { + "bbox": [ + 329, + 529, + 336, + 540 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "is too low or too high the model learns an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 540, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 452, + 551 + ], + "score": 1.0, + "content": "entangled latent representation due to either too much or too little capacity in the latent", + "type": "text" + }, + { + "bbox": [ + 453, + 541, + 460, + 549 + ], + "score": 0.28, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 540, + 506, + 551 + ], + "score": 1.0, + "content": "bottleneck.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 202, + 563 + ], + "score": 1.0, + "content": "We find that in general", + "type": "text" + }, + { + "bbox": [ + 202, + 550, + 229, + 561 + ], + "score": 0.92, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 549, + 471, + 563 + ], + "score": 1.0, + "content": "is necessary to achieve good disentanglement. However if", + "type": "text" + }, + { + "bbox": [ + 471, + 551, + 479, + 562 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "is too", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "high and the resulting capacity of the latent channel is lower than the number of data generative", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "score": 1.0, + "content": "factors, then the learnt representation necessarily has to be entangled (as a low-rank projection of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "the true data generative factors will compress them in a non-factorial way to still capture the full", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "data distribution well). We also note that VAE reconstruction quality is a poor indicator of learnt", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "disentanglement. Good disentangled representations often lead to blurry reconstructions due to the", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 311, + 628 + ], + "score": 1.0, + "content": "restricted capacity of the latent information channel", + "type": "text" + }, + { + "bbox": [ + 311, + 618, + 318, + 626 + ], + "score": 0.49, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 616, + 505, + 628 + ], + "score": 1.0, + "content": ", while entangled representations often result in", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "the sharpest reconstructions. We therefore suggest that one should not necessarily strive for perfect", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 221, + 650 + ], + "score": 1.0, + "content": "reconstructions when using", + "type": "text" + }, + { + "bbox": [ + 221, + 639, + 228, + 649 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "-VAEs as unsupervised feature learners - though it is often possible", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 172, + 661 + ], + "score": 1.0, + "content": "to find the right", + "type": "text" + }, + { + "bbox": [ + 172, + 649, + 180, + 660 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 649, + 344, + 661 + ], + "score": 1.0, + "content": "-VAE architecture and the right value of", + "type": "text" + }, + { + "bbox": [ + 344, + 649, + 352, + 660 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "to have both well disentangled latent", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 660, + 274, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 274, + 673 + ], + "score": 1.0, + "content": "representations and good reconstructions.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 47.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 284, + 689 + ], + "score": 1.0, + "content": "We proposed a principled way of choosing", + "type": "text" + }, + { + "bbox": [ + 284, + 679, + 291, + 688 + ], + "score": 0.89, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "for datasets with at least weak label information. If", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "label information exists for at least a small subset of the independent data generative factors of", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "variation, one can apply the disentanglement metric described in Sec. 3 to approximate the level of", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 248, + 723 + ], + "score": 1.0, + "content": "learnt disentanglement for various", + "type": "text" + }, + { + "bbox": [ + 248, + 712, + 255, + 721 + ], + "score": 0.88, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "choices during a hyperparameter sweep. When such labelled", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 305, + 733 + ], + "score": 1.0, + "content": "information is not available, the optimal value of", + "type": "text" + }, + { + "bbox": [ + 305, + 722, + 312, + 732 + ], + "score": 0.88, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "can be found through visual inspection of what", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 61 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 84, + 289, + 220 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 84, + 289, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 84, + 289, + 220 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 289, + 220 + ], + "score": 0.932, + "html": "
ModelDisentanglementmetric score
Ground truthRaw pixelsPCAICA100%45.75 ± 0.8%84.9 ±0.4%42.03 ± 10.6%
DC-IGNInfoGAN99.3 ± 0.1%73.5 ± 0.9%
VAE untrainedVAEβ-VAE44.14 ± 2.5%61.58 ± 0.5%
61.58 ± 0.5%99.23±0.1%
", + "type": "table", + "image_path": "247142d1273194a9ddfdf0438fd01012c53e69b471dd65d7a1fb3607bf205ec1.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 107, + 84, + 289, + 97.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 97.6, + 289, + 111.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 111.19999999999999, + 289, + 124.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 107, + 124.79999999999998, + 289, + 138.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 107, + 138.39999999999998, + 289, + 151.99999999999997 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 151.99999999999997, + 289, + 165.59999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 165.59999999999997, + 289, + 179.19999999999996 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 179.19999999999996, + 289, + 192.79999999999995 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 107, + 192.79999999999995, + 289, + 206.39999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 107, + 206.39999999999995, + 289, + 219.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "image", + "bbox": [ + 296, + 81, + 500, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 296, + 81, + 500, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 296, + 81, + 500, + 225 + ], + "spans": [ + { + "bbox": [ + 296, + 81, + 500, + 225 + ], + "score": 0.953, + "type": "image", + "image_path": "b161f1f7e0f90af0103ab47c8dc24899548b222d5a87984f5f762ba8e18c80a4.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 296, + 81, + 500, + 94.0909090909091 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 296, + 94.0909090909091, + 500, + 107.18181818181819 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 296, + 107.18181818181819, + 500, + 120.27272727272728 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 296, + 120.27272727272728, + 500, + 133.36363636363637 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 296, + 133.36363636363637, + 500, + 146.45454545454547 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 296, + 146.45454545454547, + 500, + 159.54545454545456 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 296, + 159.54545454545456, + 500, + 172.63636363636365 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 296, + 172.63636363636365, + 500, + 185.72727272727275 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 296, + 185.72727272727275, + 500, + 198.81818181818184 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 296, + 198.81818181818184, + 500, + 211.90909090909093 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 296, + 211.90909090909093, + 500, + 225.00000000000003 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 227, + 506, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "Figure 6: Disentanglement metric classification accuracy for 2D shapes dataset. Left: Accuracy for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 480, + 251 + ], + "score": 1.0, + "content": "different models and training regimes Right: Positive correlation is present between the size of", + "type": "text" + }, + { + "bbox": [ + 481, + 240, + 488, + 248 + ], + "score": 0.61, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 249, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 237, + 262 + ], + "score": 1.0, + "content": "the optimal normalised values of", + "type": "text" + }, + { + "bbox": [ + 238, + 250, + 245, + 261 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 249, + 414, + 262 + ], + "score": 1.0, + "content": "for disentangled factor learning for a fixed", + "type": "text" + }, + { + "bbox": [ + 414, + 250, + 421, + 261 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 249, + 496, + 262 + ], + "score": 1.0, + "content": "-VAE architecture.", + "type": "text" + }, + { + "bbox": [ + 496, + 250, + 504, + 261 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 235, + 273 + ], + "score": 1.0, + "content": "values are normalised by latent", + "type": "text" + }, + { + "bbox": [ + 236, + 262, + 243, + 271 + ], + "score": 0.43, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 260, + 262, + 273 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 262, + 262, + 272, + 271 + ], + "score": 0.65, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 260, + 314, + 273 + ], + "score": 1.0, + "content": "and input", + "type": "text" + }, + { + "bbox": [ + 314, + 262, + 322, + 271 + ], + "score": 0.28, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 260, + 341, + 273 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 342, + 262, + 349, + 270 + ], + "score": 0.61, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 260, + 394, + 273 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + }, + { + "bbox": [ + 394, + 261, + 402, + 272 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "values are not uniformly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 271, + 507, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 352, + 284 + ], + "score": 1.0, + "content": "sampled. Orange approximately corresponds to unnormalised", + "type": "text" + }, + { + "bbox": [ + 353, + 271, + 378, + 282 + ], + "score": 0.91, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 271, + 507, + 284 + ], + "score": 1.0, + "content": ". Good reconstructions are asso-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "ciated with entangled representations (lower disentanglement scores). Disentangled representations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 293, + 379, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 379, + 306 + ], + "score": 1.0, + "content": "(high disentanglement scores) often result in blurry reconstructions.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 165, + 328 + ], + "score": 1.0, + "content": "VAE baseline", + "type": "text" + }, + { + "bbox": [ + 165, + 316, + 192, + 327 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\beta = 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 315, + 454, + 328 + ], + "score": 1.0, + "content": ") is not able to disentangle generative factors in the data as well as", + "type": "text" + }, + { + "bbox": [ + 454, + 316, + 461, + 327 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 305, + 339 + ], + "score": 1.0, + "content": "appropriate learning pressures. Instead each latent", + "type": "text" + }, + { + "bbox": [ + 305, + 328, + 312, + 336 + ], + "score": 0.59, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 326, + 359, + 339 + ], + "score": 1.0, + "content": "(apart from", + "type": "text" + }, + { + "bbox": [ + 360, + 328, + 370, + 338 + ], + "score": 0.84, + "content": "\\mathbf { z } _ { 9 }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 326, + 506, + 339 + ], + "score": 1.0, + "content": ", which learnt rotation) encodes at", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 338, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 506, + 349 + ], + "score": 1.0, + "content": "least two data generative factors. InfoGAN also achieved a degree of disentangling (see Fig. 7D),", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "particularly for positional factors. However, despite our best efforts to train InfoGAN, we were not", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "score": 1.0, + "content": "able to achieve the same degree of disentangling in other factors, such as rotation, scale and shape.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "We also found its ability to generate the different shapes in the dataset to be inaccurate and unstable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "during training, possibly due to reported limitations of the GAN framework, which can struggle to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 392, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 507, + 405 + ], + "score": 1.0, + "content": "learn the full data distribution and instead will often learn a small subset of its modes (Salimans et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 403, + 207, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 207, + 415 + ], + "score": 1.0, + "content": "2016; Zhao et al., 2016).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 315, + 507, + 415 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 230, + 441 + ], + "score": 1.0, + "content": "Understanding the effects of", + "type": "text" + }, + { + "bbox": [ + 230, + 428, + 238, + 439 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "We hypothesised that constrained optimisation is important for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "enabling deep unsupervised models to learn disentangled representations of the independent data", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 238, + 463 + ], + "score": 1.0, + "content": "generative factors (Sec. 2). In the", + "type": "text" + }, + { + "bbox": [ + 239, + 451, + 246, + 461 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 450, + 431, + 463 + ], + "score": 1.0, + "content": "-VAE framework this corresponds to tuning the", + "type": "text" + }, + { + "bbox": [ + 431, + 451, + 439, + 461 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "coefficient. One", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 156, + 474 + ], + "score": 1.0, + "content": "way to view", + "type": "text" + }, + { + "bbox": [ + 156, + 462, + 164, + 473 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "is as a mixing coefficient (see Sec. A.6 for a derivation) for balancing the magnitudes", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 485 + ], + "score": 1.0, + "content": "of gradients from the reconstruction and the prior-matching components of the VAE lower bound", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 432, + 496 + ], + "score": 1.0, + "content": "formulation in Eq. 4 during training. In this context it makes sense to normalise", + "type": "text" + }, + { + "bbox": [ + 433, + 483, + 440, + 494 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 483, + 479, + 496 + ], + "score": 1.0, + "content": "by latent", + "type": "text" + }, + { + "bbox": [ + 479, + 485, + 486, + 493 + ], + "score": 0.63, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "size", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 107, + 496, + 117, + 504 + ], + "score": 0.64, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 494, + 160, + 507 + ], + "score": 1.0, + "content": "and input", + "type": "text" + }, + { + "bbox": [ + 161, + 496, + 169, + 504 + ], + "score": 0.59, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 494, + 190, + 507 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 190, + 496, + 198, + 504 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "in order to compare its different values across different latent layer sizes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 103, + 502, + 507, + 522 + ], + "spans": [ + { + "bbox": [ + 103, + 502, + 198, + 522 + ], + "score": 1.0, + "content": "and different datasets", + "type": "text" + }, + { + "bbox": [ + 198, + 505, + 258, + 519 + ], + "score": 0.91, + "content": "\\begin{array} { r } { ( \\beta _ { n o r m } = \\frac { \\beta M } { N } . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 502, + 379, + 522 + ], + "score": 1.0, + "content": "). We found that larger latent", + "type": "text" + }, + { + "bbox": [ + 379, + 508, + 386, + 516 + ], + "score": 0.52, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 502, + 433, + 522 + ], + "score": 1.0, + "content": "layer sizes", + "type": "text" + }, + { + "bbox": [ + 433, + 508, + 443, + 516 + ], + "score": 0.7, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 502, + 507, + 522 + ], + "score": 1.0, + "content": "require higher", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 221, + 531 + ], + "score": 1.0, + "content": "constraint pressures (higher", + "type": "text" + }, + { + "bbox": [ + 222, + 519, + 229, + 529 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 516, + 475, + 531 + ], + "score": 1.0, + "content": "values), see Fig. 6 (Right). Furthermore, the relationship of", + "type": "text" + }, + { + "bbox": [ + 475, + 518, + 482, + 529 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "for a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 130, + 541 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 131, + 530, + 141, + 538 + ], + "score": 0.68, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 528, + 328, + 541 + ], + "score": 1.0, + "content": "is characterised by an inverted U curve. When", + "type": "text" + }, + { + "bbox": [ + 329, + 529, + 336, + 540 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "is too low or too high the model learns an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 540, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 452, + 551 + ], + "score": 1.0, + "content": "entangled latent representation due to either too much or too little capacity in the latent", + "type": "text" + }, + { + "bbox": [ + 453, + 541, + 460, + 549 + ], + "score": 0.28, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 540, + 506, + 551 + ], + "score": 1.0, + "content": "bottleneck.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 202, + 563 + ], + "score": 1.0, + "content": "We find that in general", + "type": "text" + }, + { + "bbox": [ + 202, + 550, + 229, + 561 + ], + "score": 0.92, + "content": "\\beta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 549, + 471, + 563 + ], + "score": 1.0, + "content": "is necessary to achieve good disentanglement. However if", + "type": "text" + }, + { + "bbox": [ + 471, + 551, + 479, + 562 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "is too", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "high and the resulting capacity of the latent channel is lower than the number of data generative", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "score": 1.0, + "content": "factors, then the learnt representation necessarily has to be entangled (as a low-rank projection of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "the true data generative factors will compress them in a non-factorial way to still capture the full", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "data distribution well). We also note that VAE reconstruction quality is a poor indicator of learnt", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "disentanglement. Good disentangled representations often lead to blurry reconstructions due to the", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 311, + 628 + ], + "score": 1.0, + "content": "restricted capacity of the latent information channel", + "type": "text" + }, + { + "bbox": [ + 311, + 618, + 318, + 626 + ], + "score": 0.49, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 616, + 505, + 628 + ], + "score": 1.0, + "content": ", while entangled representations often result in", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "the sharpest reconstructions. We therefore suggest that one should not necessarily strive for perfect", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 221, + 650 + ], + "score": 1.0, + "content": "reconstructions when using", + "type": "text" + }, + { + "bbox": [ + 221, + 639, + 228, + 649 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "-VAEs as unsupervised feature learners - though it is often possible", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 172, + 661 + ], + "score": 1.0, + "content": "to find the right", + "type": "text" + }, + { + "bbox": [ + 172, + 649, + 180, + 660 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 649, + 344, + 661 + ], + "score": 1.0, + "content": "-VAE architecture and the right value of", + "type": "text" + }, + { + "bbox": [ + 344, + 649, + 352, + 660 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "to have both well disentangled latent", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 660, + 274, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 274, + 673 + ], + "score": 1.0, + "content": "representations and good reconstructions.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 47.5, + "bbox_fs": [ + 103, + 428, + 507, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 284, + 689 + ], + "score": 1.0, + "content": "We proposed a principled way of choosing", + "type": "text" + }, + { + "bbox": [ + 284, + 679, + 291, + 688 + ], + "score": 0.89, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "for datasets with at least weak label information. If", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "label information exists for at least a small subset of the independent data generative factors of", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "variation, one can apply the disentanglement metric described in Sec. 3 to approximate the level of", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 248, + 723 + ], + "score": 1.0, + "content": "learnt disentanglement for various", + "type": "text" + }, + { + "bbox": [ + 248, + 712, + 255, + 721 + ], + "score": 0.88, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "choices during a hyperparameter sweep. When such labelled", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 305, + 733 + ], + "score": 1.0, + "content": "information is not available, the optimal value of", + "type": "text" + }, + { + "bbox": [ + 305, + 722, + 312, + 732 + ], + "score": 0.88, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "can be found through visual inspection of what", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 61, + "bbox_fs": [ + 105, + 676, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 69, + 338, + 414 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 69, + 338, + 414 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 69, + 338, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 69, + 338, + 414 + ], + "score": 0.971, + "type": "image", + "image_path": "289e67ba96015590e7c3119eba2427223eff790ce99ac76938ab55156ba3a1fd.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 69, + 338, + 81.32142857142857 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 81.32142857142857, + 338, + 93.64285714285714 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 93.64285714285714, + 338, + 105.96428571428571 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 105.96428571428571, + 338, + 118.28571428571428 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 118.28571428571428, + 338, + 130.60714285714286 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 130.60714285714286, + 338, + 142.92857142857144 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 142.92857142857144, + 338, + 155.25000000000003 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 155.25000000000003, + 338, + 167.5714285714286 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 167.5714285714286, + 338, + 179.8928571428572 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 179.8928571428572, + 338, + 192.21428571428578 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 192.21428571428578, + 338, + 204.53571428571436 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 204.53571428571436, + 338, + 216.85714285714295 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 216.85714285714295, + 338, + 229.17857142857153 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 229.17857142857153, + 338, + 241.5000000000001 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 241.5000000000001, + 338, + 253.8214285714287 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 106, + 253.8214285714287, + 338, + 266.1428571428573 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 266.1428571428573, + 338, + 278.46428571428584 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 278.46428571428584, + 338, + 290.7857142857144 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 106, + 290.7857142857144, + 338, + 303.10714285714295 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 106, + 303.10714285714295, + 338, + 315.4285714285715 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 106, + 315.4285714285715, + 338, + 327.75000000000006 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 106, + 327.75000000000006, + 338, + 340.0714285714286 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 106, + 340.0714285714286, + 338, + 352.39285714285717 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 106, + 352.39285714285717, + 338, + 364.7142857142857 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 106, + 364.7142857142857, + 338, + 377.0357142857143 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 106, + 377.0357142857143, + 338, + 389.35714285714283 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 106, + 389.35714285714283, + 338, + 401.6785714285714 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 106, + 401.6785714285714, + 338, + 413.99999999999994 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 345, + 78, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 343, + 79, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 343, + 79, + 505, + 93 + ], + "score": 1.0, + "content": "Figure 7: A: Representations learnt by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 343, + 91, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 351, + 103 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 352, + 91, + 359, + 103 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 91, + 386, + 103 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 387, + 91, + 415, + 102 + ], + "score": 0.87, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 91, + 506, + 103 + ], + "score": 1.0, + "content": "). Each column repre-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 343, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 343, + 102, + 402, + 114 + ], + "score": 1.0, + "content": "sents a latent", + "type": "text" + }, + { + "bbox": [ + 402, + 105, + 411, + 113 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 102, + 505, + 114 + ], + "score": 1.0, + "content": ", ordered according to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 344, + 113, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 344, + 113, + 506, + 125 + ], + "score": 1.0, + "content": "the learnt Gaussian variance (last row).", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 343, + 124, + 507, + 136 + ], + "spans": [ + { + "bbox": [ + 343, + 124, + 507, + 136 + ], + "score": 1.0, + "content": "Row 1 (position) shows the mean acti-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 343, + 135, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 343, + 135, + 506, + 146 + ], + "score": 1.0, + "content": "vation (red represents high values) of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 343, + 146, + 504, + 157 + ], + "spans": [ + { + "bbox": [ + 343, + 146, + 389, + 157 + ], + "score": 1.0, + "content": "each latent", + "type": "text" + }, + { + "bbox": [ + 390, + 147, + 399, + 157 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 146, + 477, + 157 + ], + "score": 1.0, + "content": "as a function of all", + "type": "text" + }, + { + "bbox": [ + 477, + 146, + 504, + 156 + ], + "score": 0.63, + "content": "3 2 \\mathrm { x } 3 2", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 343, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 343, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "locations averaged across objects, rota-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 344, + 168, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 344, + 168, + 505, + 179 + ], + "score": 1.0, + "content": "tions and scales. Row 2 and 3 show the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 344, + 179, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 344, + 179, + 456, + 191 + ], + "score": 1.0, + "content": "mean activation of each unit", + "type": "text" + }, + { + "bbox": [ + 456, + 180, + 465, + 190 + ], + "score": 0.85, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 179, + 506, + 191 + ], + "score": 1.0, + "content": "as a func-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 343, + 189, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 343, + 189, + 506, + 202 + ], + "score": 1.0, + "content": "tion of scale (respectively rotation), av-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 344, + 201, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 344, + 201, + 506, + 213 + ], + "score": 1.0, + "content": "eraged across rotations and positions (re-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 343, + 212, + 506, + 224 + ], + "spans": [ + { + "bbox": [ + 343, + 212, + 506, + 224 + ], + "score": 1.0, + "content": "spectively scales and positions). Square", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 343, + 223, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 343, + 223, + 506, + 235 + ], + "score": 1.0, + "content": "is red, oval is green and heart is blue.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 343, + 233, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 343, + 233, + 506, + 246 + ], + "score": 1.0, + "content": "Rows 4-8 (second group) show recon-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 344, + 245, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 344, + 245, + 506, + 255 + ], + "score": 1.0, + "content": "structions resulting from the traversal", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 343, + 255, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 343, + 255, + 401, + 267 + ], + "score": 1.0, + "content": "of each latent", + "type": "text" + }, + { + "bbox": [ + 401, + 257, + 410, + 267 + ], + "score": 0.85, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 255, + 506, + 267 + ], + "score": 1.0, + "content": "over three standard de-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 344, + 267, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 344, + 267, + 505, + 278 + ], + "score": 1.0, + "content": "viations around the unit Gaussian prior", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 344, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 344, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "mean while keeping the remaining 9/10", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 344, + 289, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 344, + 289, + 505, + 299 + ], + "score": 1.0, + "content": "latent units fixed to the values obtained", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 343, + 300, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 343, + 300, + 505, + 312 + ], + "score": 1.0, + "content": "by running inference on an image from", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 343, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 343, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "the dataset. B: Similar analysis for VAE", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 347, + 321, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 347, + 321, + 378, + 332 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\beta = 1 ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 321, + 506, + 333 + ], + "score": 1.0, + "content": ". C: Similar analysis for DC-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 343, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 343, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "IGN, clamping a single latent each for", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 343, + 343, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 343, + 343, + 506, + 355 + ], + "score": 1.0, + "content": "scale, positions, orientation and 5 for", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 343, + 354, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 343, + 354, + 506, + 366 + ], + "score": 1.0, + "content": "shape. D: Similar analysis for InfoGAN,", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 343, + 365, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 343, + 365, + 506, + 377 + ], + "score": 1.0, + "content": "using 5 continuous latents regularized", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 344, + 377, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 344, + 377, + 505, + 387 + ], + "score": 1.0, + "content": "using the mutual information cost, and", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 343, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 343, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "5 additional unconstrained noise latents", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 343, + 398, + 396, + 410 + ], + "spans": [ + { + "bbox": [ + 343, + 398, + 396, + 410 + ], + "score": 1.0, + "content": "(not shown).", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 289, + 457 + ], + "score": 1.0, + "content": "effect the traversal of each single latent unit", + "type": "text" + }, + { + "bbox": [ + 289, + 446, + 302, + 455 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 443, + 423, + 457 + ], + "score": 1.0, + "content": "has on the generated images", + "type": "text" + }, + { + "bbox": [ + 423, + 444, + 444, + 456 + ], + "score": 0.9, + "content": "\\mathbf { \\tau } ( \\mathbf { x } | \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "in pixel space", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "score": 1.0, + "content": "(as shown in Fig. 7 rows 4-8). For the 2D shapes dataset, we have found that the optimal values", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 118, + 479 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 466, + 126, + 477 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "as determined by visual inspection match closely the optimal values as determined by the", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 476, + 205, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 205, + 489 + ], + "score": 1.0, + "content": "disentanglement metric.", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 59.5 + }, + { + "type": "title", + "bbox": [ + 108, + 507, + 195, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 506, + 197, + 523 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 197, + 523 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 62 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "In this paper we have reformulated the standard VAE framework (Kingma & Welling, 2014; Rezende", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 545, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 558 + ], + "score": 1.0, + "content": "et al., 2014) as a constrained optimisation problem with strong latent capacity constraint and in-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 450, + 569 + ], + "score": 1.0, + "content": "dependence prior pressures. By augmenting the lower bound formulation with the", + "type": "text" + }, + { + "bbox": [ + 450, + 556, + 458, + 568 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "coefficient", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "that regulates the strength of such pressures and, as a consequence, the qualitative nature of the", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "representations learnt by the model, we have achieved state of the art results for learning disentangled", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 589, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 426, + 601 + ], + "score": 1.0, + "content": "representations of data generative factors. We have shown that our proposed", + "type": "text" + }, + { + "bbox": [ + 426, + 590, + 434, + 600 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 589, + 506, + 601 + ], + "score": 1.0, + "content": "-VAE framework", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 104, + 598, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 598, + 506, + 615 + ], + "score": 1.0, + "content": "significantly outperforms both qualitatively and quantitatively the original VAE (Kingma & Welling,", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "2014), as well as state-of-the-art unsupervised (InfoGAN: Chen et al., 2016) and semi-supervised", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "(DC-IGN: Kulkarni et al., 2015) approaches to disentangled factor learning. Furthermore, we have", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 153, + 645 + ], + "score": 1.0, + "content": "shown that", + "type": "text" + }, + { + "bbox": [ + 153, + 633, + 160, + 644 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "-VAE consistently and robustly discovers more factors of variation in the data, and it", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "learns a representation that covers a wider range of factor values and is disentangled more cleanly", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "than other benchmarks, all in a completely unsupervised manner. Unlike InfoGAN and DC-IGN,", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "our approach does not depend on any a priori knowledge about the number or the nature of data", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 447, + 690 + ], + "score": 1.0, + "content": "generative factors. Our preliminary investigations suggest that the performance of the", + "type": "text" + }, + { + "bbox": [ + 447, + 677, + 454, + 688 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 677, + 507, + 690 + ], + "score": 1.0, + "content": "-VAE frame-", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "work may depend on the sampling density of the data generative factors within a training dataset", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "score": 1.0, + "content": "(see Appendix A.8 for more details). It appears that having more densely sampled data generative", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 323, + 722 + ], + "score": 1.0, + "content": "factors results in better disentangling performance of", + "type": "text" + }, + { + "bbox": [ + 323, + 711, + 330, + 721 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "-VAE, however we leave a more principled", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 105, + 721, + 274, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 274, + 732 + ], + "score": 1.0, + "content": "investigation of this effect to future work.", + "type": "text" + } + ], + "index": 80 + } + ], + "index": 71.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 69, + 338, + 414 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 69, + 338, + 414 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 69, + 338, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 69, + 338, + 414 + ], + "score": 0.971, + "type": "image", + "image_path": "289e67ba96015590e7c3119eba2427223eff790ce99ac76938ab55156ba3a1fd.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 69, + 338, + 81.32142857142857 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 81.32142857142857, + 338, + 93.64285714285714 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 93.64285714285714, + 338, + 105.96428571428571 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 105.96428571428571, + 338, + 118.28571428571428 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 118.28571428571428, + 338, + 130.60714285714286 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 130.60714285714286, + 338, + 142.92857142857144 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 142.92857142857144, + 338, + 155.25000000000003 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 155.25000000000003, + 338, + 167.5714285714286 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 167.5714285714286, + 338, + 179.8928571428572 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 179.8928571428572, + 338, + 192.21428571428578 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 192.21428571428578, + 338, + 204.53571428571436 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 204.53571428571436, + 338, + 216.85714285714295 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 216.85714285714295, + 338, + 229.17857142857153 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 229.17857142857153, + 338, + 241.5000000000001 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 241.5000000000001, + 338, + 253.8214285714287 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 106, + 253.8214285714287, + 338, + 266.1428571428573 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 266.1428571428573, + 338, + 278.46428571428584 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 278.46428571428584, + 338, + 290.7857142857144 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 106, + 290.7857142857144, + 338, + 303.10714285714295 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 106, + 303.10714285714295, + 338, + 315.4285714285715 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 106, + 315.4285714285715, + 338, + 327.75000000000006 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 106, + 327.75000000000006, + 338, + 340.0714285714286 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 106, + 340.0714285714286, + 338, + 352.39285714285717 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 106, + 352.39285714285717, + 338, + 364.7142857142857 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 106, + 364.7142857142857, + 338, + 377.0357142857143 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 106, + 377.0357142857143, + 338, + 389.35714285714283 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 106, + 389.35714285714283, + 338, + 401.6785714285714 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 106, + 401.6785714285714, + 338, + 413.99999999999994 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 345, + 78, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 343, + 79, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 343, + 79, + 505, + 93 + ], + "score": 1.0, + "content": "Figure 7: A: Representations learnt by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 343, + 91, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 351, + 103 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 352, + 91, + 359, + 103 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 91, + 386, + 103 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 387, + 91, + 415, + 102 + ], + "score": 0.87, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 91, + 506, + 103 + ], + "score": 1.0, + "content": "). Each column repre-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 343, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 343, + 102, + 402, + 114 + ], + "score": 1.0, + "content": "sents a latent", + "type": "text" + }, + { + "bbox": [ + 402, + 105, + 411, + 113 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 102, + 505, + 114 + ], + "score": 1.0, + "content": ", ordered according to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 344, + 113, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 344, + 113, + 506, + 125 + ], + "score": 1.0, + "content": "the learnt Gaussian variance (last row).", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 343, + 124, + 507, + 136 + ], + "spans": [ + { + "bbox": [ + 343, + 124, + 507, + 136 + ], + "score": 1.0, + "content": "Row 1 (position) shows the mean acti-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 343, + 135, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 343, + 135, + 506, + 146 + ], + "score": 1.0, + "content": "vation (red represents high values) of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 343, + 146, + 504, + 157 + ], + "spans": [ + { + "bbox": [ + 343, + 146, + 389, + 157 + ], + "score": 1.0, + "content": "each latent", + "type": "text" + }, + { + "bbox": [ + 390, + 147, + 399, + 157 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 146, + 477, + 157 + ], + "score": 1.0, + "content": "as a function of all", + "type": "text" + }, + { + "bbox": [ + 477, + 146, + 504, + 156 + ], + "score": 0.63, + "content": "3 2 \\mathrm { x } 3 2", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 343, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 343, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "locations averaged across objects, rota-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 344, + 168, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 344, + 168, + 505, + 179 + ], + "score": 1.0, + "content": "tions and scales. Row 2 and 3 show the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 344, + 179, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 344, + 179, + 456, + 191 + ], + "score": 1.0, + "content": "mean activation of each unit", + "type": "text" + }, + { + "bbox": [ + 456, + 180, + 465, + 190 + ], + "score": 0.85, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 179, + 506, + 191 + ], + "score": 1.0, + "content": "as a func-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 343, + 189, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 343, + 189, + 506, + 202 + ], + "score": 1.0, + "content": "tion of scale (respectively rotation), av-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 344, + 201, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 344, + 201, + 506, + 213 + ], + "score": 1.0, + "content": "eraged across rotations and positions (re-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 343, + 212, + 506, + 224 + ], + "spans": [ + { + "bbox": [ + 343, + 212, + 506, + 224 + ], + "score": 1.0, + "content": "spectively scales and positions). Square", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 343, + 223, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 343, + 223, + 506, + 235 + ], + "score": 1.0, + "content": "is red, oval is green and heart is blue.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 343, + 233, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 343, + 233, + 506, + 246 + ], + "score": 1.0, + "content": "Rows 4-8 (second group) show recon-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 344, + 245, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 344, + 245, + 506, + 255 + ], + "score": 1.0, + "content": "structions resulting from the traversal", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 343, + 255, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 343, + 255, + 401, + 267 + ], + "score": 1.0, + "content": "of each latent", + "type": "text" + }, + { + "bbox": [ + 401, + 257, + 410, + 267 + ], + "score": 0.85, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 255, + 506, + 267 + ], + "score": 1.0, + "content": "over three standard de-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 344, + 267, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 344, + 267, + 505, + 278 + ], + "score": 1.0, + "content": "viations around the unit Gaussian prior", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 344, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 344, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "mean while keeping the remaining 9/10", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 344, + 289, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 344, + 289, + 505, + 299 + ], + "score": 1.0, + "content": "latent units fixed to the values obtained", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 343, + 300, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 343, + 300, + 505, + 312 + ], + "score": 1.0, + "content": "by running inference on an image from", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 343, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 343, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "the dataset. B: Similar analysis for VAE", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 347, + 321, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 347, + 321, + 378, + 332 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\beta = 1 ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 321, + 506, + 333 + ], + "score": 1.0, + "content": ". C: Similar analysis for DC-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 343, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 343, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "IGN, clamping a single latent each for", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 343, + 343, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 343, + 343, + 506, + 355 + ], + "score": 1.0, + "content": "scale, positions, orientation and 5 for", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 343, + 354, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 343, + 354, + 506, + 366 + ], + "score": 1.0, + "content": "shape. D: Similar analysis for InfoGAN,", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 343, + 365, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 343, + 365, + 506, + 377 + ], + "score": 1.0, + "content": "using 5 continuous latents regularized", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 344, + 377, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 344, + 377, + 505, + 387 + ], + "score": 1.0, + "content": "using the mutual information cost, and", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 343, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 343, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "5 additional unconstrained noise latents", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 343, + 398, + 396, + 410 + ], + "spans": [ + { + "bbox": [ + 343, + 398, + 396, + 410 + ], + "score": 1.0, + "content": "(not shown).", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 42.5, + "bbox_fs": [ + 343, + 79, + 507, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 289, + 457 + ], + "score": 1.0, + "content": "effect the traversal of each single latent unit", + "type": "text" + }, + { + "bbox": [ + 289, + 446, + 302, + 455 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 443, + 423, + 457 + ], + "score": 1.0, + "content": "has on the generated images", + "type": "text" + }, + { + "bbox": [ + 423, + 444, + 444, + 456 + ], + "score": 0.9, + "content": "\\mathbf { \\tau } ( \\mathbf { x } | \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "in pixel space", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 467 + ], + "score": 1.0, + "content": "(as shown in Fig. 7 rows 4-8). For the 2D shapes dataset, we have found that the optimal values", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 118, + 479 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 466, + 126, + 477 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "as determined by visual inspection match closely the optimal values as determined by the", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 476, + 205, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 205, + 489 + ], + "score": 1.0, + "content": "disentanglement metric.", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 59.5, + "bbox_fs": [ + 105, + 443, + 506, + 489 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 507, + 195, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 506, + 197, + 523 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 197, + 523 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 62 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "In this paper we have reformulated the standard VAE framework (Kingma & Welling, 2014; Rezende", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 545, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 558 + ], + "score": 1.0, + "content": "et al., 2014) as a constrained optimisation problem with strong latent capacity constraint and in-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 106, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 450, + 569 + ], + "score": 1.0, + "content": "dependence prior pressures. By augmenting the lower bound formulation with the", + "type": "text" + }, + { + "bbox": [ + 450, + 556, + 458, + 568 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "coefficient", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "that regulates the strength of such pressures and, as a consequence, the qualitative nature of the", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "representations learnt by the model, we have achieved state of the art results for learning disentangled", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 589, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 426, + 601 + ], + "score": 1.0, + "content": "representations of data generative factors. We have shown that our proposed", + "type": "text" + }, + { + "bbox": [ + 426, + 590, + 434, + 600 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 589, + 506, + 601 + ], + "score": 1.0, + "content": "-VAE framework", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 104, + 598, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 598, + 506, + 615 + ], + "score": 1.0, + "content": "significantly outperforms both qualitatively and quantitatively the original VAE (Kingma & Welling,", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "2014), as well as state-of-the-art unsupervised (InfoGAN: Chen et al., 2016) and semi-supervised", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "(DC-IGN: Kulkarni et al., 2015) approaches to disentangled factor learning. Furthermore, we have", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 153, + 645 + ], + "score": 1.0, + "content": "shown that", + "type": "text" + }, + { + "bbox": [ + 153, + 633, + 160, + 644 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "-VAE consistently and robustly discovers more factors of variation in the data, and it", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "learns a representation that covers a wider range of factor values and is disentangled more cleanly", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "than other benchmarks, all in a completely unsupervised manner. Unlike InfoGAN and DC-IGN,", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "our approach does not depend on any a priori knowledge about the number or the nature of data", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 447, + 690 + ], + "score": 1.0, + "content": "generative factors. Our preliminary investigations suggest that the performance of the", + "type": "text" + }, + { + "bbox": [ + 447, + 677, + 454, + 688 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 677, + 507, + 690 + ], + "score": 1.0, + "content": "-VAE frame-", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "work may depend on the sampling density of the data generative factors within a training dataset", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "score": 1.0, + "content": "(see Appendix A.8 for more details). It appears that having more densely sampled data generative", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 323, + 722 + ], + "score": 1.0, + "content": "factors results in better disentangling performance of", + "type": "text" + }, + { + "bbox": [ + 323, + 711, + 330, + 721 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "-VAE, however we leave a more principled", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 105, + 721, + 274, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 274, + 732 + ], + "score": 1.0, + "content": "investigation of this effect to future work.", + "type": "text" + } + ], + "index": 80 + } + ], + "index": 71.5, + "bbox_fs": [ + 104, + 534, + 507, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 114, + 94 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "-VAE is robust with respect to different architectures, optimisation parameters and datasets, hence", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "requiring few design decisions. Our approach relies on the optimisation of a single hyperparameter", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 114, + 116 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 103, + 505, + 118 + ], + "score": 1.0, + "content": ", which can be found directly through a hyperparameter search if weakly labelled data is available", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 425, + 128 + ], + "score": 1.0, + "content": "to calculate our new proposed disentangling metric. Alternatively the optimal", + "type": "text" + }, + { + "bbox": [ + 426, + 116, + 433, + 127 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "can be estimated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "score": 1.0, + "content": "heuristically in purely unsupervised scenarios. Learning an interpretable factorised representation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "of the independent data generative factors in a completely unsupervised manner is an important", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 147, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 147, + 506, + 163 + ], + "score": 1.0, + "content": "precursor for the development of artificial intelligence that understands the world in the same way", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 158, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 505, + 173 + ], + "score": 1.0, + "content": "that humans do (Lake et al., 2016). We believe that using our approach as an unsupervised pretraining", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "stage for supervised or reinforcement learning will produce significant improvements for scenarios", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 236, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 236, + 195 + ], + "score": 1.0, + "content": "such as transfer or fast learning.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 108, + 210, + 243, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 245, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 245, + 224 + ], + "score": 1.0, + "content": "6 ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 504, + 258 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "We would like to thank Charles Blundell, Danilo Rezende, Tejas Kulkarni and David Pfau for helpful", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 246, + 271, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 271, + 259 + ], + "score": 1.0, + "content": "comments that improved the manuscript.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 107, + 275, + 175, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 176, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 176, + 289 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 105, + 289, + 506, + 735 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "M. Aubry, D. Maturana, A. Efros, B. Russell, and J. Sivic. Seeing 3d chairs: exemplar part-based", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 306, + 395, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 395, + 318 + ], + "score": 1.0, + "content": "2d-3d alignment using a large dataset of cad models. In CVPR, 2014.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "Y. Bengio, A. Courville, and P. Vincent. Representation learning: A review and new perspectives. In", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 327, + 399, + 341 + ], + "spans": [ + { + "bbox": [ + 114, + 327, + 399, + 341 + ], + "score": 1.0, + "content": "IEEE Transactions on Pattern Analysis & Machine Intelligence, 2013.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 507, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 507, + 353 + ], + "score": 1.0, + "content": "Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 349, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 115, + 349, + 507, + 365 + ], + "score": 1.0, + "content": "Interpretable representation learning by information maximizing generative adversarial nets. arXiv,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 362, + 142, + 374 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 142, + 374 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "Brian Cheung, Jesse A. Levezey, Arjun K. Bansal, and Bruno A. Olshausen. Discovering hidden", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 383, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 506, + 398 + ], + "score": 1.0, + "content": "factors of variation in deep networks. In Proceedings of the International Conference on Learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 395, + 279, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 395, + 279, + 407 + ], + "score": 1.0, + "content": "Representations, Workshop Track, 2015.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 405, + 507, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 405, + 507, + 420 + ], + "score": 1.0, + "content": "T. Cohen and M. Welling. Transformation properties of learned visual representations. In ICLR,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 417, + 142, + 429 + ], + "spans": [ + { + "bbox": [ + 115, + 417, + 142, + 429 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 427, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 507, + 443 + ], + "score": 1.0, + "content": "Taco Cohen and Max Welling. Learning the irreducible representations of commutative lie groups.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 440, + 168, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 440, + 168, + 451 + ], + "score": 1.0, + "content": "arXiv, 2014.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "score": 1.0, + "content": "G. Desjardins, A. Courville, and Y. Bengio. Disentangling factors of variation via generative", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 462, + 217, + 475 + ], + "spans": [ + { + "bbox": [ + 115, + 462, + 217, + 475 + ], + "score": 1.0, + "content": "entangling. arXiv, 2012.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "score": 1.0, + "content": "Carl Doersch. Tutorial on variational autoencoders. arxiv, 2016.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 495, + 401, + 509 + ], + "spans": [ + { + "bbox": [ + 114, + 495, + 401, + 509 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 2011.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 518, + 392, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 518, + 392, + 531 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial nets. NIPS, pp. 2672–2680, 2014.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "score": 1.0, + "content": "Ross Goroshin, Michael Mathieu, and Yann LeCun. Learning to linearize under uncertainty. NIPS,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 541, + 142, + 552 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 142, + 552 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 551, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 104, + 551, + 506, + 566 + ], + "score": 1.0, + "content": "G. Hinton, A. Krizhevsky, and S. D. Wang. Transforming auto-encoders. International Conference", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 564, + 264, + 575 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 264, + 575 + ], + "score": 1.0, + "content": "on Artificial Neural Networks, 2011.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "score": 1.0, + "content": "Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. An introduction to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 586, + 451, + 598 + ], + "spans": [ + { + "bbox": [ + 115, + 586, + 451, + 598 + ], + "score": 1.0, + "content": "variational methods for graphical models. Machine learning, 37(2):183–233, 1999.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "score": 1.0, + "content": "Theofanis Karaletsos, Serge Belongie, and Gunnar Rätsch. Bayesian representation learning with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 608, + 244, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 244, + 621 + ], + "score": 1.0, + "content": "oracle constraints. ICLR, 2016.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 618, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 633 + ], + "score": 1.0, + "content": "W. Karush. Minima of Functions of Several Variables with Inequalities as Side Constraints. Master’s", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 630, + 313, + 643 + ], + "spans": [ + { + "bbox": [ + 115, + 630, + 313, + 643 + ], + "score": 1.0, + "content": "thesis, Univ. of Chicago, Chicago, Illinois, 1939.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 642, + 462, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 462, + 654 + ], + "score": 1.0, + "content": "D. P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2014.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 654, + 418, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 418, + 666 + ], + "score": 1.0, + "content": "D. P. Kingma and M. Welling. Auto-encoding variational bayes. ICLR, 2014.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 664, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 506, + 678 + ], + "score": 1.0, + "content": "H. W. Kuhn and A. W. Tucker. Nonlinear programming. In Proceedings of 2nd Berkeley Symposium,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 676, + 196, + 687 + ], + "spans": [ + { + "bbox": [ + 115, + 676, + 196, + 687 + ], + "score": 1.0, + "content": "pp. 481–492, 1951.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Tejas Kulkarni, William Whitney, Pushmeet Kohli, and Joshua Tenenbaum. Deep convolutional", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 115, + 699, + 273, + 710 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 273, + 710 + ], + "score": 1.0, + "content": "inverse graphics network. NIPS, 2015.", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 116, + 720, + 339, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 339, + 733 + ], + "score": 1.0, + "content": "machines that learn and think like people. arXiv, 2016.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 33 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 114, + 94 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "-VAE is robust with respect to different architectures, optimisation parameters and datasets, hence", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "requiring few design decisions. Our approach relies on the optimisation of a single hyperparameter", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 114, + 116 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 103, + 505, + 118 + ], + "score": 1.0, + "content": ", which can be found directly through a hyperparameter search if weakly labelled data is available", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 425, + 128 + ], + "score": 1.0, + "content": "to calculate our new proposed disentangling metric. Alternatively the optimal", + "type": "text" + }, + { + "bbox": [ + 426, + 116, + 433, + 127 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "can be estimated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "score": 1.0, + "content": "heuristically in purely unsupervised scenarios. Learning an interpretable factorised representation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "of the independent data generative factors in a completely unsupervised manner is an important", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 147, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 147, + 506, + 163 + ], + "score": 1.0, + "content": "precursor for the development of artificial intelligence that understands the world in the same way", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 158, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 505, + 173 + ], + "score": 1.0, + "content": "that humans do (Lake et al., 2016). We believe that using our approach as an unsupervised pretraining", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "stage for supervised or reinforcement learning will produce significant improvements for scenarios", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 236, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 236, + 195 + ], + "score": 1.0, + "content": "such as transfer or fast learning.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 104, + 82, + 506, + 195 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 210, + 243, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 245, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 245, + 224 + ], + "score": 1.0, + "content": "6 ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 504, + 258 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "We would like to thank Charles Blundell, Danilo Rezende, Tejas Kulkarni and David Pfau for helpful", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 246, + 271, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 271, + 259 + ], + "score": 1.0, + "content": "comments that improved the manuscript.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 235, + 505, + 259 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 275, + 175, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 176, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 176, + 289 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "list", + "bbox": [ + 105, + 289, + 506, + 735 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "M. Aubry, D. Maturana, A. Efros, B. Russell, and J. Sivic. Seeing 3d chairs: exemplar part-based", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 306, + 395, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 395, + 318 + ], + "score": 1.0, + "content": "2d-3d alignment using a large dataset of cad models. In CVPR, 2014.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "Y. Bengio, A. Courville, and P. Vincent. Representation learning: A review and new perspectives. In", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 327, + 399, + 341 + ], + "spans": [ + { + "bbox": [ + 114, + 327, + 399, + 341 + ], + "score": 1.0, + "content": "IEEE Transactions on Pattern Analysis & Machine Intelligence, 2013.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 339, + 507, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 507, + 353 + ], + "score": 1.0, + "content": "Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 349, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 115, + 349, + 507, + 365 + ], + "score": 1.0, + "content": "Interpretable representation learning by information maximizing generative adversarial nets. arXiv,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 362, + 142, + 374 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 142, + 374 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "Brian Cheung, Jesse A. Levezey, Arjun K. Bansal, and Bruno A. Olshausen. Discovering hidden", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 383, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 506, + 398 + ], + "score": 1.0, + "content": "factors of variation in deep networks. In Proceedings of the International Conference on Learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 395, + 279, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 395, + 279, + 407 + ], + "score": 1.0, + "content": "Representations, Workshop Track, 2015.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 405, + 507, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 405, + 507, + 420 + ], + "score": 1.0, + "content": "T. Cohen and M. Welling. Transformation properties of learned visual representations. In ICLR,", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 417, + 142, + 429 + ], + "spans": [ + { + "bbox": [ + 115, + 417, + 142, + 429 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 427, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 507, + 443 + ], + "score": 1.0, + "content": "Taco Cohen and Max Welling. Learning the irreducible representations of commutative lie groups.", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 440, + 168, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 440, + 168, + 451 + ], + "score": 1.0, + "content": "arXiv, 2014.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "score": 1.0, + "content": "G. Desjardins, A. Courville, and Y. Bengio. Disentangling factors of variation via generative", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 462, + 217, + 475 + ], + "spans": [ + { + "bbox": [ + 115, + 462, + 217, + 475 + ], + "score": 1.0, + "content": "entangling. arXiv, 2012.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 365, + 486 + ], + "score": 1.0, + "content": "Carl Doersch. Tutorial on variational autoencoders. arxiv, 2016.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 495, + 401, + 509 + ], + "spans": [ + { + "bbox": [ + 114, + 495, + 401, + 509 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 2011.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 518, + 392, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 518, + 392, + 531 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial nets. NIPS, pp. 2672–2680, 2014.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "score": 1.0, + "content": "Ross Goroshin, Michael Mathieu, and Yann LeCun. Learning to linearize under uncertainty. NIPS,", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 541, + 142, + 552 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 142, + 552 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 551, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 104, + 551, + 506, + 566 + ], + "score": 1.0, + "content": "G. Hinton, A. Krizhevsky, and S. D. Wang. Transforming auto-encoders. International Conference", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 564, + 264, + 575 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 264, + 575 + ], + "score": 1.0, + "content": "on Artificial Neural Networks, 2011.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 588 + ], + "score": 1.0, + "content": "Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. An introduction to", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 586, + 451, + 598 + ], + "spans": [ + { + "bbox": [ + 115, + 586, + 451, + 598 + ], + "score": 1.0, + "content": "variational methods for graphical models. Machine learning, 37(2):183–233, 1999.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "score": 1.0, + "content": "Theofanis Karaletsos, Serge Belongie, and Gunnar Rätsch. Bayesian representation learning with", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 608, + 244, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 244, + 621 + ], + "score": 1.0, + "content": "oracle constraints. ICLR, 2016.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 618, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 633 + ], + "score": 1.0, + "content": "W. Karush. Minima of Functions of Several Variables with Inequalities as Side Constraints. Master’s", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 630, + 313, + 643 + ], + "spans": [ + { + "bbox": [ + 115, + 630, + 313, + 643 + ], + "score": 1.0, + "content": "thesis, Univ. of Chicago, Chicago, Illinois, 1939.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 642, + 462, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 462, + 654 + ], + "score": 1.0, + "content": "D. P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2014.", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 654, + 418, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 418, + 666 + ], + "score": 1.0, + "content": "D. P. Kingma and M. Welling. Auto-encoding variational bayes. ICLR, 2014.", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 664, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 506, + 678 + ], + "score": 1.0, + "content": "H. W. Kuhn and A. W. Tucker. Nonlinear programming. In Proceedings of 2nd Berkeley Symposium,", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 676, + 196, + 687 + ], + "spans": [ + { + "bbox": [ + 115, + 676, + 196, + 687 + ], + "score": 1.0, + "content": "pp. 481–492, 1951.", + "type": "text" + } + ], + "index": 48, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Tejas Kulkarni, William Whitney, Pushmeet Kohli, and Joshua Tenenbaum. Deep convolutional", + "type": "text" + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 699, + 273, + 710 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 273, + 710 + ], + "score": 1.0, + "content": "inverse graphics network. NIPS, 2015.", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building", + "type": "text" + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 720, + 339, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 339, + 733 + ], + "score": 1.0, + "content": "machines that learn and think like people. arXiv, 2016.", + "type": "text" + } + ], + "index": 52, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 82, + 481, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 481, + 95 + ], + "score": 1.0, + "content": "Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. ICCV, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3d face model for pose and", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 105, + 325, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 325, + 117 + ], + "score": 1.0, + "content": "illumination invariant face recognition. AVSS, 2009.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 126, + 507, + 141 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 507, + 141 + ], + "score": 1.0, + "content": "Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas,", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "Alexandre Passos, and David Cournapeau. Scikit-learn: Machine learning in python. Journal of", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 150, + 258, + 161 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 258, + 161 + ], + "score": 1.0, + "content": "Machine Learning Research, 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "score": 1.0, + "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 172, + 358, + 183 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 358, + 183 + ], + "score": 1.0, + "content": "convolutional generative adversarial networks. arXiv, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "Scott Reed, Kihyuk Sohn, Yuting Zhang, and Honglak Lee. Learning to disentangle factors of", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 195, + 315, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 315, + 207 + ], + "score": 1.0, + "content": "variation with manifold interaction. ICML, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 205, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 506, + 218 + ], + "score": 1.0, + "content": "Danilo J. Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approxi-", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 217, + 340, + 229 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 340, + 229 + ], + "score": 1.0, + "content": "mate inference in deep generative models. arXiv, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 226, + 507, + 242 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 507, + 242 + ], + "score": 1.0, + "content": "Karl Ridgeway. A survey of inductive biases for factorial Representation-Learning. arXiv, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 239, + 330, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 239, + 330, + 252 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1612.05299.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "Oren Rippel and Ryan Prescott Adams. High-dimensional probability estimation with deep density", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 262, + 204, + 272 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 204, + 272 + ], + "score": 1.0, + "content": "models. arXiv, 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 272, + 507, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 507, + 285 + ], + "score": 1.0, + "content": "Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 285, + 507, + 296 + ], + "spans": [ + { + "bbox": [ + 116, + 285, + 507, + 296 + ], + "score": 1.0, + "content": "Improved techniques for training GANs. arXiv, 2016. URL http://arxiv.org/abs/1606.", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 296, + 151, + 306 + ], + "spans": [ + { + "bbox": [ + 116, + 296, + 151, + 306 + ], + "score": 1.0, + "content": "03498.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 306, + 507, + 320 + ], + "spans": [ + { + "bbox": [ + 104, + 306, + 507, + 320 + ], + "score": 1.0, + "content": "Jürgen Schmidhuber. Learning factorial codes by predictability minimization. Neural Computation,", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 317, + 200, + 329 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 200, + 329 + ], + "score": 1.0, + "content": "4(6):863–869, 1992.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "Rueckert, and Zehan Wang. Real-Time single image and video Super-Resolution using an efficient", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 350, + 333, + 363 + ], + "spans": [ + { + "bbox": [ + 116, + 350, + 333, + 363 + ], + "score": 1.0, + "content": "Sub-Pixel convolutional neural network. arXiv, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "Yichuan Tang, Ruslan Salakhutdinov, and Geoffrey Hinton. Tensor analyzers. In Proceedings of the", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 373, + 438, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 373, + 438, + 386 + ], + "score": 1.0, + "content": "30th International Conference on Machine Learning, 2013, Atlanta, USA, 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "score": 1.0, + "content": "William F. Whitney, Michael Chang, Tejas Kulkarni, and Joshua B. Tenenbaum. Understanding", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "visual concepts with continuation learning. arXiv, 2016. URL http://arxiv.org/pdf/", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 407, + 205, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 205, + 419 + ], + "score": 1.0, + "content": "1602.06822.pdf.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Jimei Yang, Scott Reed, Ming-Hsuan Yang, and Honglak Lee. Weakly-supervised disentangling with", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 429, + 362, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 362, + 441 + ], + "score": 1.0, + "content": "recurrent transformations for 3d view synthesis. NIPS, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network. arXiv,", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 451, + 357, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 357, + 464 + ], + "score": 1.0, + "content": "2016. URL http://arxiv.org/abs/1609.03126.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Z. Zhu, P. Luo, X. Wang, and X. Tang. Multi-view perceptron: a deep model for learning face identity", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 473, + 483, + 486 + ], + "spans": [ + { + "bbox": [ + 115, + 473, + 483, + 486 + ], + "score": 1.0, + "content": "and view representations. In Advances in Neural Information Processing Systems 27. 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + } + ], + "index": 33, + "bbox_fs": [ + 104, + 294, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 78, + 507, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 481, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 481, + 95 + ], + "score": 1.0, + "content": "Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. ICCV, 2015.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3d face model for pose and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 325, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 325, + 117 + ], + "score": 1.0, + "content": "illumination invariant face recognition. AVSS, 2009.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 126, + 507, + 141 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 507, + 141 + ], + "score": 1.0, + "content": "Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "Alexandre Passos, and David Cournapeau. Scikit-learn: Machine learning in python. Journal of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 150, + 258, + 161 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 258, + 161 + ], + "score": 1.0, + "content": "Machine Learning Research, 2011.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "score": 1.0, + "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 172, + 358, + 183 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 358, + 183 + ], + "score": 1.0, + "content": "convolutional generative adversarial networks. arXiv, 2015.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "Scott Reed, Kihyuk Sohn, Yuting Zhang, and Honglak Lee. Learning to disentangle factors of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 195, + 315, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 315, + 207 + ], + "score": 1.0, + "content": "variation with manifold interaction. ICML, 2014.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 205, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 506, + 218 + ], + "score": 1.0, + "content": "Danilo J. Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approxi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 217, + 340, + 229 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 340, + 229 + ], + "score": 1.0, + "content": "mate inference in deep generative models. arXiv, 2014.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 226, + 507, + 242 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 507, + 242 + ], + "score": 1.0, + "content": "Karl Ridgeway. A survey of inductive biases for factorial Representation-Learning. arXiv, 2016.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 239, + 330, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 239, + 330, + 252 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1612.05299.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "Oren Rippel and Ryan Prescott Adams. High-dimensional probability estimation with deep density", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 262, + 204, + 272 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 204, + 272 + ], + "score": 1.0, + "content": "models. arXiv, 2013.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 272, + 507, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 507, + 285 + ], + "score": 1.0, + "content": "Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 285, + 507, + 296 + ], + "spans": [ + { + "bbox": [ + 116, + 285, + 507, + 296 + ], + "score": 1.0, + "content": "Improved techniques for training GANs. arXiv, 2016. URL http://arxiv.org/abs/1606.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 296, + 151, + 306 + ], + "spans": [ + { + "bbox": [ + 116, + 296, + 151, + 306 + ], + "score": 1.0, + "content": "03498.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 306, + 507, + 320 + ], + "spans": [ + { + "bbox": [ + 104, + 306, + 507, + 320 + ], + "score": 1.0, + "content": "Jürgen Schmidhuber. Learning factorial codes by predictability minimization. Neural Computation,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 317, + 200, + 329 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 200, + 329 + ], + "score": 1.0, + "content": "4(6):863–869, 1992.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "Rueckert, and Zehan Wang. Real-Time single image and video Super-Resolution using an efficient", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 350, + 333, + 363 + ], + "spans": [ + { + "bbox": [ + 116, + 350, + 333, + 363 + ], + "score": 1.0, + "content": "Sub-Pixel convolutional neural network. arXiv, 2016.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "Yichuan Tang, Ruslan Salakhutdinov, and Geoffrey Hinton. Tensor analyzers. In Proceedings of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 373, + 438, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 373, + 438, + 386 + ], + "score": 1.0, + "content": "30th International Conference on Machine Learning, 2013, Atlanta, USA, 2013.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "score": 1.0, + "content": "William F. Whitney, Michael Chang, Tejas Kulkarni, and Joshua B. Tenenbaum. Understanding", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "visual concepts with continuation learning. arXiv, 2016. URL http://arxiv.org/pdf/", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 407, + 205, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 205, + 419 + ], + "score": 1.0, + "content": "1602.06822.pdf.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Jimei Yang, Scott Reed, Ming-Hsuan Yang, and Honglak Lee. Weakly-supervised disentangling with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 429, + 362, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 362, + 441 + ], + "score": 1.0, + "content": "recurrent transformations for 3d view synthesis. NIPS, 2015.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network. arXiv,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 451, + 357, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 357, + 464 + ], + "score": 1.0, + "content": "2016. URL http://arxiv.org/abs/1609.03126.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Z. Zhu, P. Luo, X. Wang, and X. Tang. Multi-view perceptron: a deep model for learning face identity", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 473, + 483, + 486 + ], + "spans": [ + { + "bbox": [ + 115, + 473, + 483, + 486 + ], + "score": 1.0, + "content": "and view representations. In Advances in Neural Information Processing Systems 27. 2014.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 521, + 183, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 184, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 184, + 536 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 274, + 562 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 275, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 275, + 563 + ], + "score": 1.0, + "content": "A.1 MODEL ARCHITECTURE DETAILS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 575, + 415, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 417, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 417, + 588 + ], + "score": 1.0, + "content": "A summary of all model architectures used in this paper can be seen in Tbl 1.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 108, + 609, + 223, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 224, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 224, + 621 + ], + "score": 1.0, + "content": "A.2 INFOGAN TRAINING", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "To train the InfoGAN network described in Tbl. 1 on the 2D shapes dataset (Fig. 7), we followed", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "the training paradigm described in Chen et al. (2016) with the following modifications. For the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 415, + 668 + ], + "score": 1.0, + "content": "mutual information regularised latent code, we used 5 continuous variables", + "type": "text" + }, + { + "bbox": [ + 415, + 657, + 424, + 666 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "sampled uniformly", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 130, + 678 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 130, + 666, + 160, + 678 + ], + "score": 0.92, + "content": "( - 1 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 665, + 279, + 678 + ], + "score": 1.0, + "content": ". We used 5 noise variables", + "type": "text" + }, + { + "bbox": [ + 279, + 668, + 288, + 677 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 665, + 506, + 678 + ], + "score": 1.0, + "content": ", as we found that using a reduced number of noise", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "variables improved the quality of generated samples for this dataset. To help stabilise training, we", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "used the instance noise trick described in Shi et al. (2016), adding Gaussian noise to the discriminator", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 304, + 712 + ], + "score": 1.0, + "content": "inputs (0.2 standard deviation on images scaled to", + "type": "text" + }, + { + "bbox": [ + 304, + 699, + 334, + 711 + ], + "score": 0.87, + "content": "[ - 1 , 1 ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". We followed Radford et al. (2015) for the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "architecture of the convolutional layers, and used batch normalisation in all layers except the last in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 293, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 293, + 733 + ], + "score": 1.0, + "content": "the generator and the first in the discriminator.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 78, + 507, + 489 + ], + "lines": [], + "index": 17.5, + "bbox_fs": [ + 104, + 82, + 507, + 486 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 521, + 183, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 184, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 184, + 536 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 274, + 562 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 275, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 275, + 563 + ], + "score": 1.0, + "content": "A.1 MODEL ARCHITECTURE DETAILS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 575, + 415, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 417, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 417, + 588 + ], + "score": 1.0, + "content": "A summary of all model architectures used in this paper can be seen in Tbl 1.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 573, + 417, + 588 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 609, + 223, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 224, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 224, + 621 + ], + "score": 1.0, + "content": "A.2 INFOGAN TRAINING", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "To train the InfoGAN network described in Tbl. 1 on the 2D shapes dataset (Fig. 7), we followed", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "the training paradigm described in Chen et al. (2016) with the following modifications. For the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 415, + 668 + ], + "score": 1.0, + "content": "mutual information regularised latent code, we used 5 continuous variables", + "type": "text" + }, + { + "bbox": [ + 415, + 657, + 424, + 666 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "sampled uniformly", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 130, + 678 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 130, + 666, + 160, + 678 + ], + "score": 0.92, + "content": "( - 1 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 665, + 279, + 678 + ], + "score": 1.0, + "content": ". We used 5 noise variables", + "type": "text" + }, + { + "bbox": [ + 279, + 668, + 288, + 677 + ], + "score": 0.84, + "content": "z _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 665, + 506, + 678 + ], + "score": 1.0, + "content": ", as we found that using a reduced number of noise", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "variables improved the quality of generated samples for this dataset. To help stabilise training, we", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "used the instance noise trick described in Shi et al. (2016), adding Gaussian noise to the discriminator", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 304, + 712 + ], + "score": 1.0, + "content": "inputs (0.2 standard deviation on images scaled to", + "type": "text" + }, + { + "bbox": [ + 304, + 699, + 334, + 711 + ], + "score": 0.87, + "content": "[ - 1 , 1 ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". We followed Radford et al. (2015) for the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "architecture of the convolutional layers, and used batch normalisation in all layers except the last in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 293, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 293, + 733 + ], + "score": 1.0, + "content": "the generator and the first in the discriminator.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 632, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 77, + 531, + 509 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 77, + 531, + 509 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 77, + 531, + 509 + ], + "spans": [ + { + "bbox": [ + 107, + 77, + 531, + 509 + ], + "score": 0.983, + "html": "
DatasetOptimiserArchitecture
2D shapesAdagradInput4096 (flattened 64x64x1).
(VAE)1e-2EncoderFC 1200,1200.ReLU activation.
Latents Decoder10
FC 1200,1200,1200,4096. Tanh activation.Bernoulli.
2D shapesrmspropInput64x64x1.
(DC-IGN)(as in Kulkarni et al., 2015)EncoderConv 96x3x3,48x3x3,48x3x3 (padding 1). ReLU activation and Max pooling 2x2.
Latents10
DecoderUnpooling,Conv 48x3x3,96x3x3,1x3x3. ReLU activation, Sigmoid.
2D shapes (InfoGAN)Adam 1e-3 (gen)GeneratorFC 256,256,Deconv 128x4x4, 64x4x4 (stride 2). Tanh.
2e-4 (dis)DiscriminatorConv and FC reverse of generator.Leaky ReLU activation. FC1. Sigmoid activation.
Recognition LatentsConv and FC shared with discriminator.FC 128,5.Gaussian 10: z1...5 ~ Unif(-1,1),c1...5 ~ Unif(-1,1)
ChairsAdamInput64x64x1.
(VAE)1e-4EncoderConv 32x4x4 (stride 2),32x4x4 (stride 2), 64x4x4 (stride 2),
Latents64x4x4 (stride 2),FC 256.ReLU activation. 32
DecoderDeconv reverse of encoder. ReLU activation.Bernoulli.
CelebA
(VAE)Adam 1e-4Input Encoder64x64x3. Conv 32x4x4 (stride 2),32x4x4 (stride 2),64x4x4 (stride 2),
Latents64x4x4 (stride 2),FC 256.ReLU activation. 32
DecoderDeconv reverse of encoder.ReLU activation.Gaussian.
3DFacesAdam64x64x1.
(VAE)1e-4Input EncoderConv 32x4x4 (stride 2),32x4x4 (stride 2), 64x4x4 (stride 2),
64x4x4 (stride 2),FC 256.ReLU activation. 32
Latents DecoderDeconv reverse of encoder.ReLU activation.Bernoulli.
", + "type": "table", + "image_path": "d9da1e738460cf74fa403dfbd27f4329fa341c351a36fae056a86f8100279ec8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 77, + 531, + 221.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 221.0, + 531, + 365.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 365.0, + 531, + 509.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 511, + 504, + 534 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "Table 1: Details of model architectures used in the paper. The models were trained using either", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 523, + 392, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 392, + 536 + ], + "score": 1.0, + "content": "adagrad (Duchi et al., 2011) or adam (Kingma & Ba, 2014) optimisers.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 108, + 556, + 246, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 556, + 248, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 248, + 569 + ], + "score": 1.0, + "content": "A.3 ICA AND PCA BASELINES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 634 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "In order to calculate the ICA benchmark, we applied fastICA (Pedregosa et al., 2011) algorithm", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "to the whitened pixel data. Due to memory limitations we had to apply the algorithm to pairwise", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "combinations of the subsets of the dataset corresponding to the transforms of each of the three 2D", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "object identities. We calculated the disentangling metric for all three ICA models trained on each of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 621, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 507, + 635 + ], + "score": 1.0, + "content": "the three pairwise combinations of 2D objects, before presenting the average of these scores in Fig. 6.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 108, + 639, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "We performed PCA on the raw and whitened pixel data. Both approaches resulted in similar", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "disentangling metric scores. Fig. 6 reports the PCA results calculated using whitened pixel data for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 661, + 284, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 284, + 673 + ], + "score": 1.0, + "content": "more direct comparison with the ICA score.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 687, + 293, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 295, + 700 + ], + "score": 1.0, + "content": "A.4 DISENTANGLEMENT METRIC DETAILS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 708, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 703, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 703, + 466, + 724 + ], + "score": 1.0, + "content": "We used a linear classifier to learn the identity of the generative factor that produced", + "type": "text" + }, + { + "bbox": [ + 467, + 707, + 484, + 721 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 703, + 506, + 724 + ], + "score": 1.0, + "content": "(see", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 719, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 339, + 735 + ], + "score": 1.0, + "content": "Equations (5) for the process used to obtain samples of", + "type": "text" + }, + { + "bbox": [ + 339, + 720, + 357, + 733 + ], + "score": 0.9, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } .", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 719, + 506, + 735 + ], + "score": 1.0, + "content": "). We used a fully connected linear", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 77, + 531, + 509 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 77, + 531, + 509 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 77, + 531, + 509 + ], + "spans": [ + { + "bbox": [ + 107, + 77, + 531, + 509 + ], + "score": 0.983, + "html": "
DatasetOptimiserArchitecture
2D shapesAdagradInput4096 (flattened 64x64x1).
(VAE)1e-2EncoderFC 1200,1200.ReLU activation.
Latents Decoder10
FC 1200,1200,1200,4096. Tanh activation.Bernoulli.
2D shapesrmspropInput64x64x1.
(DC-IGN)(as in Kulkarni et al., 2015)EncoderConv 96x3x3,48x3x3,48x3x3 (padding 1). ReLU activation and Max pooling 2x2.
Latents10
DecoderUnpooling,Conv 48x3x3,96x3x3,1x3x3. ReLU activation, Sigmoid.
2D shapes (InfoGAN)Adam 1e-3 (gen)GeneratorFC 256,256,Deconv 128x4x4, 64x4x4 (stride 2). Tanh.
2e-4 (dis)DiscriminatorConv and FC reverse of generator.Leaky ReLU activation. FC1. Sigmoid activation.
Recognition LatentsConv and FC shared with discriminator.FC 128,5.Gaussian 10: z1...5 ~ Unif(-1,1),c1...5 ~ Unif(-1,1)
ChairsAdamInput64x64x1.
(VAE)1e-4EncoderConv 32x4x4 (stride 2),32x4x4 (stride 2), 64x4x4 (stride 2),
Latents64x4x4 (stride 2),FC 256.ReLU activation. 32
DecoderDeconv reverse of encoder. ReLU activation.Bernoulli.
CelebA
(VAE)Adam 1e-4Input Encoder64x64x3. Conv 32x4x4 (stride 2),32x4x4 (stride 2),64x4x4 (stride 2),
Latents64x4x4 (stride 2),FC 256.ReLU activation. 32
DecoderDeconv reverse of encoder.ReLU activation.Gaussian.
3DFacesAdam64x64x1.
(VAE)1e-4Input EncoderConv 32x4x4 (stride 2),32x4x4 (stride 2), 64x4x4 (stride 2),
64x4x4 (stride 2),FC 256.ReLU activation. 32
Latents DecoderDeconv reverse of encoder.ReLU activation.Bernoulli.
", + "type": "table", + "image_path": "d9da1e738460cf74fa403dfbd27f4329fa341c351a36fae056a86f8100279ec8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 77, + 531, + 221.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 221.0, + 531, + 365.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 365.0, + 531, + 509.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 511, + 504, + 534 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "Table 1: Details of model architectures used in the paper. The models were trained using either", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 523, + 392, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 392, + 536 + ], + "score": 1.0, + "content": "adagrad (Duchi et al., 2011) or adam (Kingma & Ba, 2014) optimisers.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 108, + 556, + 246, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 556, + 248, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 248, + 569 + ], + "score": 1.0, + "content": "A.3 ICA AND PCA BASELINES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 634 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "In order to calculate the ICA benchmark, we applied fastICA (Pedregosa et al., 2011) algorithm", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "to the whitened pixel data. Due to memory limitations we had to apply the algorithm to pairwise", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "combinations of the subsets of the dataset corresponding to the transforms of each of the three 2D", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "object identities. We calculated the disentangling metric for all three ICA models trained on each of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 621, + 507, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 507, + 635 + ], + "score": 1.0, + "content": "the three pairwise combinations of 2D objects, before presenting the average of these scores in Fig. 6.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 577, + 507, + 635 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 639, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "We performed PCA on the raw and whitened pixel data. Both approaches resulted in similar", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "disentangling metric scores. Fig. 6 reports the PCA results calculated using whitened pixel data for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 661, + 284, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 284, + 673 + ], + "score": 1.0, + "content": "more direct comparison with the ICA score.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 639, + 505, + 673 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 687, + 293, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 295, + 700 + ], + "score": 1.0, + "content": "A.4 DISENTANGLEMENT METRIC DETAILS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 708, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 703, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 703, + 466, + 724 + ], + "score": 1.0, + "content": "We used a linear classifier to learn the identity of the generative factor that produced", + "type": "text" + }, + { + "bbox": [ + 467, + 707, + 484, + 721 + ], + "score": 0.91, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 703, + 506, + 724 + ], + "score": 1.0, + "content": "(see", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 719, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 339, + 735 + ], + "score": 1.0, + "content": "Equations (5) for the process used to obtain samples of", + "type": "text" + }, + { + "bbox": [ + 339, + 720, + 357, + 733 + ], + "score": 0.9, + "content": "\\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } .", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 719, + 506, + 735 + ], + "score": 1.0, + "content": "). We used a fully connected linear", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 184, + 96 + ], + "score": 1.0, + "content": "classifier to predict", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 184, + 82, + 221, + 95 + ], + "score": 0.93, + "content": "p ( y | \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 222, + 81, + 252, + 96 + ], + "score": 1.0, + "content": ", where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 252, + 84, + 259, + 94 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 259, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "is one of four generative factors (position X, position Y, scale", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "and rotation). We used softmax output nonlinearity and a negative log likelihood loss function. The", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 505, + 119 + ], + "score": 1.0, + "content": "classifier was trained using the Adagrad (Duchi et al., 2011) optimisation algorithm with learning", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 230, + 128 + ], + "score": 1.0, + "content": "rate of 1e-2 until convergence.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 15.5, + "bbox_fs": [ + 104, + 703, + 506, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 184, + 96 + ], + "score": 1.0, + "content": "classifier to predict", + "type": "text" + }, + { + "bbox": [ + 184, + 82, + 221, + 95 + ], + "score": 0.93, + "content": "p ( y | \\mathbf { z } _ { \\mathrm { d i f f } } ^ { b } )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 81, + 252, + 96 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 252, + 84, + 259, + 94 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "is one of four generative factors (position X, position Y, scale", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "and rotation). We used softmax output nonlinearity and a negative log likelihood loss function. The", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 505, + 119 + ], + "score": 1.0, + "content": "classifier was trained using the Adagrad (Duchi et al., 2011) optimisation algorithm with learning", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 230, + 128 + ], + "score": 1.0, + "content": "rate of 1e-2 until convergence.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 137, + 398, + 152 + ], + "lines": [ + { + "bbox": [ + 169, + 137, + 398, + 152 + ], + "spans": [ + { + "bbox": [ + 169, + 137, + 398, + 152 + ], + "score": 0.86, + "content": "\\mathcal { D } = \\{ V \\in \\mathbb { R } ^ { K } , W \\in \\mathbb { R } ^ { H } , X \\in \\mathbb { R } ^ { N } \\} , y \\sim U n i f [ 1 . . . K ]", + "type": "interline_equation", + "image_path": "6036cfc159e05d5bbbf545b650ffd2ebdaa7b04fe7da0a519cffea15b782bf42.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 169, + 137, + 398, + 152 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 129, + 153, + 226, + 164 + ], + "lines": [ + { + "bbox": [ + 128, + 152, + 228, + 166 + ], + "spans": [ + { + "bbox": [ + 128, + 152, + 173, + 166 + ], + "score": 1.0, + "content": "Repeat for", + "type": "text" + }, + { + "bbox": [ + 173, + 153, + 221, + 163 + ], + "score": 0.81, + "content": "b = 1 \\dots B", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 152, + 228, + 166 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 169, + 467, + 262 + ], + "lines": [ + { + "bbox": [ + 150, + 169, + 467, + 262 + ], + "spans": [ + { + "bbox": [ + 150, + 169, + 467, + 262 + ], + "score": 0.92, + "content": "\\begin{array} { l } { { \\displaystyle { \\bf v } _ { 1 , l } \\sim p ( { \\bf v } ) , ~ { \\bf w } _ { 1 , l } \\sim p ( { \\bf w } ) , ~ { \\bf w } _ { 2 , l } \\sim p ( { \\bf w } ) , ~ [ { \\bf v } _ { 2 , l } ] _ { k } = \\left\\{ \\begin{array} { l l } { { \\displaystyle { \\bf v } _ { 1 , l } ] _ { k } , } } & { { \\mathrm { i f ~ } k = y } } \\\\ { { \\displaystyle \\sim p ( v _ { k } ) , } } & { { \\mathrm { o t h e r w i s e } } } \\end{array} \\right. } \\ ~ } \\\\ { { \\displaystyle { \\bf x } _ { 1 , l } \\sim \\mathrm { S i m } ( { \\bf v } _ { 1 , 1 } , { \\bf w } _ { 1 , 1 } ) , ~ { \\bf x } _ { 2 , 1 } \\sim \\mathrm { S i m } ( { \\bf v } _ { 2 , 1 } , { \\bf w } _ { 2 , 1 } ) , } \\ ~ } \\\\ { { \\displaystyle q ( { \\bf z } | { \\bf x } ) \\sim \\mathcal N ( \\mu ( { \\bf x } ) , \\sigma ( { \\bf x } ) ) , ~ { \\bf z } _ { 1 , l } = \\mu ( { \\bf x } _ { 1 , l } ) , ~ { \\bf z } _ { 2 , l } = \\mu ( { \\bf x } _ { 2 , l } ) } \\ ~ } \\\\ { \\displaystyle { \\bf z } _ { \\mathrm { d i f f } } ^ { l } = | { \\bf z } _ { 1 , l } - { \\bf z } _ { 2 , l } | , ~ { \\bf z } _ { \\mathrm { d i f f } } ^ { b } = \\frac { 1 } { L } \\sum _ { l = 1 } ^ { L } { \\bf z } _ { \\mathrm { d i f f } } ^ { l } } \\end{array}", + "type": "interline_equation", + "image_path": "226396db549c26a6e5ed48925ca99618dccde7ac5c2a79fd7af4d7d82ea58a9a.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 150, + 169, + 467, + 200.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 150, + 200.0, + 467, + 231.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 150, + 231.0, + 467, + 262.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 270, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "All disentanglement metric score results reported in the paper were calculated in the following manner.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Ten replicas of each model with the same hyperparameters were trained using different random seeds", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "score": 1.0, + "content": "to obtain disentangled representations. Each of the ten trained model replicas was evaluated three", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "score": 1.0, + "content": "times using the disentanglement metric score algorithm, each time using a different random seed", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 365, + 327 + ], + "score": 1.0, + "content": "to initialise the linear classifier. We then discarded the bottom", + "type": "text" + }, + { + "bbox": [ + 365, + 314, + 385, + 325 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "of the thirty resulting scores", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "and reported the remaining results. This was done to control for the outlier results from the few", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 336, + 276, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 276, + 349 + ], + "score": 1.0, + "content": "experiments that diverged during training.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "The results reported in table in Fig. 6 (left) were calculated using the following data. Ground truth", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 363, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 267, + 378 + ], + "score": 1.0, + "content": "uses independent data generating factors", + "type": "text" + }, + { + "bbox": [ + 267, + 366, + 275, + 374 + ], + "score": 0.29, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 363, + 506, + 378 + ], + "score": 1.0, + "content": "(our dataset did not contain any correlated data generating", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "score": 1.0, + "content": "factors w). PCA and ICA decompositions keep the first ten components (PCA components explain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 134, + 397 + ], + "score": 0.88, + "content": "6 0 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 386, + 187, + 397 + ], + "score": 1.0, + "content": "of variance).", + "type": "text" + }, + { + "bbox": [ + 187, + 386, + 195, + 397 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 386, + 220, + 397 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 220, + 386, + 248, + 397 + ], + "score": 0.85, + "content": "\\beta = 4 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 386, + 277, + 397 + ], + "score": 1.0, + "content": "), VAE", + "type": "text" + }, + { + "bbox": [ + 277, + 386, + 304, + 397 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\beta = 1 , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 386, + 505, + 397 + ], + "score": 1.0, + "content": ") and VAE untrained have the same fully connected", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 397, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 234, + 409 + ], + "score": 1.0, + "content": "architecture with ten latent units", + "type": "text" + }, + { + "bbox": [ + 235, + 399, + 241, + 407 + ], + "score": 0.35, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 397, + 506, + 409 + ], + "score": 1.0, + "content": ". InfoGAN uses “inferred” values of the five continuous latents that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 407, + 392, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 392, + 421 + ], + "score": 1.0, + "content": "were regularised with the mutual information objective during training.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 109, + 438, + 440, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 441, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 441, + 451 + ], + "score": 1.0, + "content": "A.5 CLASSIFYING THE GROUND TRUTH DATA GENERATIVE FACTORS VALUES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "In order to further verify the validity of our proposed disentanglement metric we ran an extra", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "score": 1.0, + "content": "quantitative test: we trained a linear classifier to predict the ground truth value of each of the five", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "data generative factors used to generate the 2D shapes dataset. While this test does not measure", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "disentangling directly (since it does not measure independence of the latent representation), a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "disentangled representation should make such a classification trivial. It can be seen in Table 2 that the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 515, + 203, + 528 + ], + "score": 1.0, + "content": "representation learnt by", + "type": "text" + }, + { + "bbox": [ + 204, + 516, + 211, + 527 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "-VAE is on average the best representation for factor classification across", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 527, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 539 + ], + "score": 1.0, + "content": "all five factors. It is closely followed by DC-IGN. It is interesting to note that ICA does well only at", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 551 + ], + "score": 1.0, + "content": "encoding object identity, while PCA manages to learn a very good representation of object position.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5 + }, + { + "type": "table", + "bbox": [ + 135, + 563, + 476, + 666 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 135, + 563, + 476, + 666 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 563, + 476, + 666 + ], + "spans": [ + { + "bbox": [ + 135, + 563, + 476, + 666 + ], + "score": 0.98, + "html": "
ModelClassification accuracy
idscalerotationposition Xposition Yaverage
PCA43.3836.085.9660.6660.1541.25
ICA59.634.47.6125.9625.1230.54
DC-IGN44.8245.9215.8947.6445.8840.03
InfoGAN44.4740.916.3927.5123.7328.60
VAEuntrained39.4425.336.0916.6914.3920.39
VAE41.5524.07816.518.7221.77
β-VAE50.0843.0320.3652.2549.543.04
", + "type": "table", + "image_path": "1fc15232c2175c30c35222848b77c1ec128ebda68d475aac310c5a1e70db67e5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 135, + 563, + 476, + 597.3333333333334 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 135, + 597.3333333333334, + 476, + 631.6666666666667 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 135, + 631.6666666666667, + 476, + 666.0000000000001 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 506, + 718 + ], + "lines": [ + { + "bbox": [ + 104, + 672, + 506, + 686 + ], + "spans": [ + { + "bbox": [ + 104, + 672, + 506, + 686 + ], + "score": 1.0, + "content": "Table 2: Linear classifier classification accuracy for predicting the ground truth values for each data", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 685, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 104, + 685, + 506, + 696 + ], + "score": 1.0, + "content": "generative factor from different latent representations. Each factor could take a variable number of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 695, + 506, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 506, + 709 + ], + "score": 1.0, + "content": "possible values: 3 for id, 6 for scale, 40 for rotation and 32 for position X or Y. Best performing", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 708, + 303, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 303, + 718 + ], + "score": 1.0, + "content": "model results in each column are printed in bold.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 81, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 104, + 81, + 506, + 128 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 137, + 398, + 152 + ], + "lines": [ + { + "bbox": [ + 169, + 137, + 398, + 152 + ], + "spans": [ + { + "bbox": [ + 169, + 137, + 398, + 152 + ], + "score": 0.86, + "content": "\\mathcal { D } = \\{ V \\in \\mathbb { R } ^ { K } , W \\in \\mathbb { R } ^ { H } , X \\in \\mathbb { R } ^ { N } \\} , y \\sim U n i f [ 1 . . . K ]", + "type": "interline_equation", + "image_path": "6036cfc159e05d5bbbf545b650ffd2ebdaa7b04fe7da0a519cffea15b782bf42.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 169, + 137, + 398, + 152 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 129, + 153, + 226, + 164 + ], + "lines": [ + { + "bbox": [ + 128, + 152, + 228, + 166 + ], + "spans": [ + { + "bbox": [ + 128, + 152, + 173, + 166 + ], + "score": 1.0, + "content": "Repeat for", + "type": "text" + }, + { + "bbox": [ + 173, + 153, + 221, + 163 + ], + "score": 0.81, + "content": "b = 1 \\dots B", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 152, + 228, + 166 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 128, + 152, + 228, + 166 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 169, + 467, + 262 + ], + "lines": [ + { + "bbox": [ + 150, + 169, + 467, + 262 + ], + "spans": [ + { + "bbox": [ + 150, + 169, + 467, + 262 + ], + "score": 0.92, + "content": "\\begin{array} { l } { { \\displaystyle { \\bf v } _ { 1 , l } \\sim p ( { \\bf v } ) , ~ { \\bf w } _ { 1 , l } \\sim p ( { \\bf w } ) , ~ { \\bf w } _ { 2 , l } \\sim p ( { \\bf w } ) , ~ [ { \\bf v } _ { 2 , l } ] _ { k } = \\left\\{ \\begin{array} { l l } { { \\displaystyle { \\bf v } _ { 1 , l } ] _ { k } , } } & { { \\mathrm { i f ~ } k = y } } \\\\ { { \\displaystyle \\sim p ( v _ { k } ) , } } & { { \\mathrm { o t h e r w i s e } } } \\end{array} \\right. } \\ ~ } \\\\ { { \\displaystyle { \\bf x } _ { 1 , l } \\sim \\mathrm { S i m } ( { \\bf v } _ { 1 , 1 } , { \\bf w } _ { 1 , 1 } ) , ~ { \\bf x } _ { 2 , 1 } \\sim \\mathrm { S i m } ( { \\bf v } _ { 2 , 1 } , { \\bf w } _ { 2 , 1 } ) , } \\ ~ } \\\\ { { \\displaystyle q ( { \\bf z } | { \\bf x } ) \\sim \\mathcal N ( \\mu ( { \\bf x } ) , \\sigma ( { \\bf x } ) ) , ~ { \\bf z } _ { 1 , l } = \\mu ( { \\bf x } _ { 1 , l } ) , ~ { \\bf z } _ { 2 , l } = \\mu ( { \\bf x } _ { 2 , l } ) } \\ ~ } \\\\ { \\displaystyle { \\bf z } _ { \\mathrm { d i f f } } ^ { l } = | { \\bf z } _ { 1 , l } - { \\bf z } _ { 2 , l } | , ~ { \\bf z } _ { \\mathrm { d i f f } } ^ { b } = \\frac { 1 } { L } \\sum _ { l = 1 } ^ { L } { \\bf z } _ { \\mathrm { d i f f } } ^ { l } } \\end{array}", + "type": "interline_equation", + "image_path": "226396db549c26a6e5ed48925ca99618dccde7ac5c2a79fd7af4d7d82ea58a9a.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 150, + 169, + 467, + 200.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 150, + 200.0, + 467, + 231.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 150, + 231.0, + 467, + 262.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 270, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "All disentanglement metric score results reported in the paper were calculated in the following manner.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Ten replicas of each model with the same hyperparameters were trained using different random seeds", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "score": 1.0, + "content": "to obtain disentangled representations. Each of the ten trained model replicas was evaluated three", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "score": 1.0, + "content": "times using the disentanglement metric score algorithm, each time using a different random seed", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 365, + 327 + ], + "score": 1.0, + "content": "to initialise the linear classifier. We then discarded the bottom", + "type": "text" + }, + { + "bbox": [ + 365, + 314, + 385, + 325 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "of the thirty resulting scores", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "and reported the remaining results. This was done to control for the outlier results from the few", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 336, + 276, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 276, + 349 + ], + "score": 1.0, + "content": "experiments that diverged during training.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 270, + 506, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "The results reported in table in Fig. 6 (left) were calculated using the following data. Ground truth", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 363, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 267, + 378 + ], + "score": 1.0, + "content": "uses independent data generating factors", + "type": "text" + }, + { + "bbox": [ + 267, + 366, + 275, + 374 + ], + "score": 0.29, + "content": "\\mathbf { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 363, + 506, + 378 + ], + "score": 1.0, + "content": "(our dataset did not contain any correlated data generating", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "score": 1.0, + "content": "factors w). PCA and ICA decompositions keep the first ten components (PCA components explain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 134, + 397 + ], + "score": 0.88, + "content": "6 0 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 386, + 187, + 397 + ], + "score": 1.0, + "content": "of variance).", + "type": "text" + }, + { + "bbox": [ + 187, + 386, + 195, + 397 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 386, + 220, + 397 + ], + "score": 1.0, + "content": "-VAE", + "type": "text" + }, + { + "bbox": [ + 220, + 386, + 248, + 397 + ], + "score": 0.85, + "content": "\\beta = 4 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 386, + 277, + 397 + ], + "score": 1.0, + "content": "), VAE", + "type": "text" + }, + { + "bbox": [ + 277, + 386, + 304, + 397 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\beta = 1 , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 386, + 505, + 397 + ], + "score": 1.0, + "content": ") and VAE untrained have the same fully connected", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 397, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 234, + 409 + ], + "score": 1.0, + "content": "architecture with ten latent units", + "type": "text" + }, + { + "bbox": [ + 235, + 399, + 241, + 407 + ], + "score": 0.35, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 397, + 506, + 409 + ], + "score": 1.0, + "content": ". InfoGAN uses “inferred” values of the five continuous latents that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 407, + 392, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 392, + 421 + ], + "score": 1.0, + "content": "were regularised with the mutual information objective during training.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 353, + 506, + 421 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 438, + 440, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 441, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 441, + 451 + ], + "score": 1.0, + "content": "A.5 CLASSIFYING THE GROUND TRUTH DATA GENERATIVE FACTORS VALUES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "In order to further verify the validity of our proposed disentanglement metric we ran an extra", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "score": 1.0, + "content": "quantitative test: we trained a linear classifier to predict the ground truth value of each of the five", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "data generative factors used to generate the 2D shapes dataset. While this test does not measure", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "disentangling directly (since it does not measure independence of the latent representation), a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "disentangled representation should make such a classification trivial. It can be seen in Table 2 that the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 515, + 203, + 528 + ], + "score": 1.0, + "content": "representation learnt by", + "type": "text" + }, + { + "bbox": [ + 204, + 516, + 211, + 527 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "-VAE is on average the best representation for factor classification across", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 527, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 539 + ], + "score": 1.0, + "content": "all five factors. It is closely followed by DC-IGN. It is interesting to note that ICA does well only at", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 551 + ], + "score": 1.0, + "content": "encoding object identity, while PCA manages to learn a very good representation of object position.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 460, + 506, + 551 + ] + }, + { + "type": "table", + "bbox": [ + 135, + 563, + 476, + 666 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 135, + 563, + 476, + 666 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 563, + 476, + 666 + ], + "spans": [ + { + "bbox": [ + 135, + 563, + 476, + 666 + ], + "score": 0.98, + "html": "
ModelClassification accuracy
idscalerotationposition Xposition Yaverage
PCA43.3836.085.9660.6660.1541.25
ICA59.634.47.6125.9625.1230.54
DC-IGN44.8245.9215.8947.6445.8840.03
InfoGAN44.4740.916.3927.5123.7328.60
VAEuntrained39.4425.336.0916.6914.3920.39
VAE41.5524.07816.518.7221.77
β-VAE50.0843.0320.3652.2549.543.04
", + "type": "table", + "image_path": "1fc15232c2175c30c35222848b77c1ec128ebda68d475aac310c5a1e70db67e5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 135, + 563, + 476, + 597.3333333333334 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 135, + 597.3333333333334, + 476, + 631.6666666666667 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 135, + 631.6666666666667, + 476, + 666.0000000000001 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 506, + 718 + ], + "lines": [ + { + "bbox": [ + 104, + 672, + 506, + 686 + ], + "spans": [ + { + "bbox": [ + 104, + 672, + 506, + 686 + ], + "score": 1.0, + "content": "Table 2: Linear classifier classification accuracy for predicting the ground truth values for each data", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 685, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 104, + 685, + 506, + 696 + ], + "score": 1.0, + "content": "generative factor from different latent representations. Each factor could take a variable number of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 695, + 506, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 506, + 709 + ], + "score": 1.0, + "content": "possible values: 3 for id, 6 for scale, 40 for rotation and 32 for position X or Y. Best performing", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 708, + 303, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 303, + 718 + ], + "score": 1.0, + "content": "model results in each column are printed in bold.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 672, + 506, + 718 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 267, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 266, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 258, + 95 + ], + "score": 1.0, + "content": "A.6 INTERPRETING NORMALISED", + "type": "text" + }, + { + "bbox": [ + 258, + 83, + 266, + 94 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 110, + 102, + 478, + 115 + ], + "lines": [ + { + "bbox": [ + 107, + 102, + 480, + 115 + ], + "spans": [ + { + "bbox": [ + 107, + 102, + 176, + 115 + ], + "score": 1.0, + "content": "We start with the", + "type": "text" + }, + { + "bbox": [ + 176, + 104, + 183, + 115 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 102, + 480, + 115 + ], + "score": 1.0, + "content": "-VAE constrained optimisation formulation that we have derived in Sec. 2.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 120, + 438, + 135 + ], + "lines": [ + { + "bbox": [ + 172, + 120, + 438, + 135 + ], + "spans": [ + { + "bbox": [ + 172, + 120, + 438, + 135 + ], + "score": 0.88, + "content": "\\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\beta ) = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ] - \\beta D _ { K L } ( q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) )", + "type": "interline_equation", + "image_path": "f89b501d346c79a257146ec0a0dd1d1b4e660c1b99ded97f1cb8254ab328f850.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 172, + 120, + 438, + 135 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 142, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 507, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 268, + 157 + ], + "score": 1.0, + "content": "We make the assumption that every pixel", + "type": "text" + }, + { + "bbox": [ + 268, + 145, + 275, + 153 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 140, + 286, + 157 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 286, + 142, + 320, + 153 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 140, + 458, + 157 + ], + "score": 1.0, + "content": "is conditionally independent given", + "type": "text" + }, + { + "bbox": [ + 458, + 146, + 464, + 153 + ], + "score": 0.69, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 140, + 507, + 157 + ], + "score": 1.0, + "content": "(Doersch,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 153, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 286, + 167 + ], + "score": 1.0, + "content": "2016). The first term of Eq. 6 then becomes:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 168, + 466, + 194 + ], + "lines": [ + { + "bbox": [ + 144, + 168, + 466, + 194 + ], + "spans": [ + { + "bbox": [ + 144, + 168, + 466, + 194 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ] = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log \\prod _ { n } p _ { \\theta } ( x _ { n } | \\mathbf { z } ) ] = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\sum _ { n } \\log p _ { \\theta } ( x _ { n } | \\mathbf { z } ) ]", + "type": "interline_equation", + "image_path": "3820f977f925ba0b8dfff99fb0f457039fc987b114a83d492a438c1e83fa2d35.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 144, + 168, + 466, + 194 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 195, + 286, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 234, + 209 + ], + "score": 1.0, + "content": "Dividing both sides of Eq. 6 by", + "type": "text" + }, + { + "bbox": [ + 234, + 196, + 244, + 206 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 194, + 286, + 209 + ], + "score": 1.0, + "content": "produces:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 210, + 449, + 235 + ], + "lines": [ + { + "bbox": [ + 161, + 210, + 449, + 235 + ], + "spans": [ + { + "bbox": [ + 161, + 210, + 449, + 235 + ], + "score": 0.94, + "content": "\\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\beta ) \\propto \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } \\mathbb { E } _ { n } [ \\log p _ { \\boldsymbol { \\theta } } ( x _ { n } | \\mathbf { z } ) ] - \\frac { \\beta } { N } D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) )", + "type": "interline_equation", + "image_path": "c3a908c01da88e7270ac0d7c18eaf95e08914a2168990b8c9f59682ae3374540.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 161, + 210, + 449, + 235 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 152, + 252 + ], + "score": 1.0, + "content": "We design", + "type": "text" + }, + { + "bbox": [ + 152, + 240, + 160, + 251 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 238, + 506, + 252 + ], + "score": 1.0, + "content": "-VAE to learn conditionally independent factors of variation in the data. Hence we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 311, + 263 + ], + "score": 1.0, + "content": "assume conditional independence of every latent", + "type": "text" + }, + { + "bbox": [ + 311, + 252, + 325, + 262 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 250, + 352, + 263 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 353, + 253, + 360, + 261 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 250, + 393, + 263 + ], + "score": 1.0, + "content": "(where", + "type": "text" + }, + { + "bbox": [ + 393, + 251, + 443, + 261 + ], + "score": 0.9, + "content": "m \\in 1 . . . M", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 250, + 466, + 263 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 466, + 251, + 478, + 261 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 179, + 274 + ], + "score": 1.0, + "content": "dimensionality of", + "type": "text" + }, + { + "bbox": [ + 180, + 264, + 186, + 271 + ], + "score": 0.65, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 261, + 256, + 274 + ], + "score": 1.0, + "content": "). Since our prior", + "type": "text" + }, + { + "bbox": [ + 256, + 262, + 275, + 273 + ], + "score": 0.91, + "content": "p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "is an isotropic unit Gaussian, we can re-write the second", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 273, + 175, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 175, + 285 + ], + "score": 1.0, + "content": "term of Eq. 6 as:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 286, + 470, + 316 + ], + "lines": [ + { + "bbox": [ + 140, + 286, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 140, + 286, + 470, + 316 + ], + "score": 0.93, + "content": "D _ { K L } ( q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) ) = \\int _ { z } q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) l o g { \\frac { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } { p ( \\mathbf { z } ) } } = \\sum _ { m } \\int _ { z _ { m } } q _ { \\phi } ( z _ { m } | \\mathbf { x } ) l o g { \\frac { q _ { \\phi } ( z _ { m } | \\mathbf { x } ) } { p ( z _ { m } ) } }", + "type": "interline_equation", + "image_path": "e81c3cac6d5e5e0c267ce4b360bbf68ba427f3f0bbbe1ef87ac2550100f403c7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 140, + 286, + 470, + 296.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 140, + 296.0, + 470, + 306.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 140, + 306.0, + 470, + 316.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 354, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 316, + 357, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 301, + 334 + ], + "score": 1.0, + "content": "Multiplying the second term in Eq. 8 by a factor", + "type": "text" + }, + { + "bbox": [ + 302, + 318, + 313, + 332 + ], + "score": 0.9, + "content": "\\textstyle { \\frac { M } { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 316, + 357, + 334 + ], + "score": 1.0, + "content": "produces:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 340, + 469, + 394 + ], + "lines": [ + { + "bbox": [ + 127, + 340, + 469, + 394 + ], + "spans": [ + { + "bbox": [ + 127, + 340, + 469, + 394 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\displaystyle \\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\boldsymbol { \\beta } ) \\propto \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } \\mathbb { E } _ { n } [ \\log p _ { \\boldsymbol { \\theta } } ( x _ { n } | \\mathbf { z } ) ] - \\frac { \\beta M } { N } \\mathbb { E } _ { m } \\int _ { z _ { m } } [ q _ { \\boldsymbol { \\phi } } ( z _ { m } | \\mathbf { x } ) l o g \\frac { q _ { \\boldsymbol { \\phi } } ( z _ { m } | \\mathbf { x } ) } { p ( z _ { m } ) } ] } \\\\ { \\displaystyle = \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } \\mathbb { E } _ { n } [ \\log p _ { \\boldsymbol { \\theta } } ( x _ { n } | \\mathbf { z } ) ] - \\frac { \\beta M } { N } \\mathbb { E } _ { m } [ D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( z _ { m } | \\mathbf { x } ) | | p ( z _ { m } ) ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "d7e24fc15f48c2257ded2a302733eb54342bc38b94082b52323a09fd1c79deaf.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 127, + 340, + 469, + 358.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 127, + 358.0, + 469, + 376.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 127, + 376.0, + 469, + 394.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 158, + 409 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 159, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 159, + 411 + ], + "score": 1.0, + "content": "Hence using", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 421, + 336, + 446 + ], + "lines": [ + { + "bbox": [ + 275, + 421, + 336, + 446 + ], + "spans": [ + { + "bbox": [ + 275, + 421, + 336, + 446 + ], + "score": 0.92, + "content": "\\beta _ { n o r m } = { \\frac { \\beta M } { N } }", + "type": "interline_equation", + "image_path": "41bbb3f29b256fe799a9f371d32e1145c2c07d986a5c61015fc12a3ede1fb591.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 275, + 421, + 336, + 446 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 313, + 466 + ], + "score": 1.0, + "content": "in Eq. 10 is equivalent to optimising the original", + "type": "text" + }, + { + "bbox": [ + 313, + 454, + 321, + 465 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "-VAE formulation from Sec. 2, but with the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "additional independence assumptions that let us calculate data log likelihood and KL divergence", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 475, + 412, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 293, + 489 + ], + "score": 1.0, + "content": "terms in expectation over the individual pixels", + "type": "text" + }, + { + "bbox": [ + 293, + 477, + 305, + 487 + ], + "score": 0.85, + "content": "x _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 475, + 394, + 489 + ], + "score": 1.0, + "content": "and individual latents", + "type": "text" + }, + { + "bbox": [ + 395, + 478, + 408, + 487 + ], + "score": 0.85, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 475, + 412, + 489 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 278, + 512 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 278, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 242, + 513 + ], + "score": 1.0, + "content": "A.7 RELATIONSHIP BETWEEN", + "type": "text" + }, + { + "bbox": [ + 243, + 501, + 250, + 512 + ], + "score": 0.77, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 499, + 272, + 513 + ], + "score": 1.0, + "content": "AND", + "type": "text" + }, + { + "bbox": [ + 273, + 503, + 278, + 510 + ], + "score": 0.39, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 151, + 534 + ], + "score": 1.0, + "content": "For a given", + "type": "text" + }, + { + "bbox": [ + 152, + 524, + 158, + 531 + ], + "score": 0.61, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 520, + 455, + 534 + ], + "score": 1.0, + "content": "we can solve the constrained optimisation problem in Eq. 3 (find the optimal", + "type": "text" + }, + { + "bbox": [ + 456, + 521, + 502, + 533 + ], + "score": 0.91, + "content": "( \\theta ^ { * } , \\phi ^ { * } , \\beta ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 520, + 506, + 534 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 143, + 545 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 144, + 532, + 229, + 544 + ], + "score": 0.92, + "content": "\\Delta \\mathcal { F } ( \\theta ^ { * } , \\phi ^ { * } , \\beta ^ { * } ) = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 531, + 505, + 545 + ], + "score": 1.0, + "content": ". We can then re-write our optimal solution to the original optimisation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 543, + 250, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 240, + 555 + ], + "score": 1.0, + "content": "problem in Eq. 2 as a function of", + "type": "text" + }, + { + "bbox": [ + 241, + 546, + 245, + 552 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 543, + 250, + 555 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 560, + 400, + 575 + ], + "lines": [ + { + "bbox": [ + 211, + 560, + 400, + 575 + ], + "spans": [ + { + "bbox": [ + 211, + 560, + 400, + 575 + ], + "score": 0.93, + "content": "\\mathcal { G } ( \\theta ^ { * } ( \\epsilon ) , \\phi ^ { * } ( \\epsilon ) ) = \\mathbb { E } _ { q _ { \\phi ^ { * } ( \\epsilon ) } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta ^ { * } ( \\epsilon ) } ( \\mathbf { x } | \\mathbf { z } ) ]", + "type": "interline_equation", + "image_path": "5782f03be4ccac442a47081021c560a4bfb31bb312f502656f5b33bd40ceec47.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 211, + 560, + 400, + 575 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 503, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 127, + 593 + ], + "score": 1.0, + "content": "Now", + "type": "text" + }, + { + "bbox": [ + 128, + 581, + 135, + 591 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 579, + 383, + 593 + ], + "score": 1.0, + "content": "can be interpreted as the rate of change of the optimal solution", + "type": "text" + }, + { + "bbox": [ + 383, + 580, + 415, + 591 + ], + "score": 0.91, + "content": "( \\theta ^ { * } , \\phi ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 579, + 426, + 593 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 426, + 581, + 434, + 591 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "when varying the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 591, + 158, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 148, + 603 + ], + "score": 1.0, + "content": "constraint", + "type": "text" + }, + { + "bbox": [ + 149, + 594, + 154, + 601 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 591, + 158, + 603 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 280, + 598, + 331, + 623 + ], + "lines": [ + { + "bbox": [ + 280, + 598, + 331, + 623 + ], + "spans": [ + { + "bbox": [ + 280, + 598, + 331, + 623 + ], + "score": 0.94, + "content": "\\frac { \\delta \\mathcal { G } } { \\delta \\epsilon } = \\beta ^ { * } ( \\epsilon )", + "type": "interline_equation", + "image_path": "36273b9b9e57ff492cffd7b3a9b88fb73aaeb6e4ff3540ed9ca99e78438d3c02.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 280, + 598, + 331, + 623 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 634, + 214, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 216, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 216, + 647 + ], + "score": 1.0, + "content": "A.8 DATA CONTINUITY", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "We hypothesise that data continuity plays a role in guiding unsupervised models towards learning the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "correct data manifolds. To test this idea we measure how the degree of learnt disentangling changes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 362, + 690 + ], + "score": 1.0, + "content": "with reduced continuity in the 2D shapes dataset. We trained a", + "type": "text" + }, + { + "bbox": [ + 363, + 678, + 370, + 688 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 676, + 414, + 690 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + }, + { + "bbox": [ + 415, + 677, + 441, + 688 + ], + "score": 0.9, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "(Figure 7A) on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "subsamples of the original 2D shapes dataset, where we progressively decreased the generative factor", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "sampling density. Reduction in data continuity negatively correlates with the average pixel wise", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "(Hamming) distance between two consecutive transforms of each object (normalised by the average", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "number of pixels occupied by each of the two adjacent transforms of an object to account for object", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 267, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 266, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 258, + 95 + ], + "score": 1.0, + "content": "A.6 INTERPRETING NORMALISED", + "type": "text" + }, + { + "bbox": [ + 258, + 83, + 266, + 94 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 110, + 102, + 478, + 115 + ], + "lines": [ + { + "bbox": [ + 107, + 102, + 480, + 115 + ], + "spans": [ + { + "bbox": [ + 107, + 102, + 176, + 115 + ], + "score": 1.0, + "content": "We start with the", + "type": "text" + }, + { + "bbox": [ + 176, + 104, + 183, + 115 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 102, + 480, + 115 + ], + "score": 1.0, + "content": "-VAE constrained optimisation formulation that we have derived in Sec. 2.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 107, + 102, + 480, + 115 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 120, + 438, + 135 + ], + "lines": [ + { + "bbox": [ + 172, + 120, + 438, + 135 + ], + "spans": [ + { + "bbox": [ + 172, + 120, + 438, + 135 + ], + "score": 0.88, + "content": "\\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\beta ) = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ] - \\beta D _ { K L } ( q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) )", + "type": "interline_equation", + "image_path": "f89b501d346c79a257146ec0a0dd1d1b4e660c1b99ded97f1cb8254ab328f850.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 172, + 120, + 438, + 135 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 142, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 507, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 268, + 157 + ], + "score": 1.0, + "content": "We make the assumption that every pixel", + "type": "text" + }, + { + "bbox": [ + 268, + 145, + 275, + 153 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 140, + 286, + 157 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 286, + 142, + 320, + 153 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 140, + 458, + 157 + ], + "score": 1.0, + "content": "is conditionally independent given", + "type": "text" + }, + { + "bbox": [ + 458, + 146, + 464, + 153 + ], + "score": 0.69, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 140, + 507, + 157 + ], + "score": 1.0, + "content": "(Doersch,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 153, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 286, + 167 + ], + "score": 1.0, + "content": "2016). The first term of Eq. 6 then becomes:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 140, + 507, + 167 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 168, + 466, + 194 + ], + "lines": [ + { + "bbox": [ + 144, + 168, + 466, + 194 + ], + "spans": [ + { + "bbox": [ + 144, + 168, + 466, + 194 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) ] = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log \\prod _ { n } p _ { \\theta } ( x _ { n } | \\mathbf { z } ) ] = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } [ \\sum _ { n } \\log p _ { \\theta } ( x _ { n } | \\mathbf { z } ) ]", + "type": "interline_equation", + "image_path": "3820f977f925ba0b8dfff99fb0f457039fc987b114a83d492a438c1e83fa2d35.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 144, + 168, + 466, + 194 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 195, + 286, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 234, + 209 + ], + "score": 1.0, + "content": "Dividing both sides of Eq. 6 by", + "type": "text" + }, + { + "bbox": [ + 234, + 196, + 244, + 206 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 194, + 286, + 209 + ], + "score": 1.0, + "content": "produces:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 194, + 286, + 209 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 210, + 449, + 235 + ], + "lines": [ + { + "bbox": [ + 161, + 210, + 449, + 235 + ], + "spans": [ + { + "bbox": [ + 161, + 210, + 449, + 235 + ], + "score": 0.94, + "content": "\\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\beta ) \\propto \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } \\mathbb { E } _ { n } [ \\log p _ { \\boldsymbol { \\theta } } ( x _ { n } | \\mathbf { z } ) ] - \\frac { \\beta } { N } D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) )", + "type": "interline_equation", + "image_path": "c3a908c01da88e7270ac0d7c18eaf95e08914a2168990b8c9f59682ae3374540.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 161, + 210, + 449, + 235 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 152, + 252 + ], + "score": 1.0, + "content": "We design", + "type": "text" + }, + { + "bbox": [ + 152, + 240, + 160, + 251 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 238, + 506, + 252 + ], + "score": 1.0, + "content": "-VAE to learn conditionally independent factors of variation in the data. Hence we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 311, + 263 + ], + "score": 1.0, + "content": "assume conditional independence of every latent", + "type": "text" + }, + { + "bbox": [ + 311, + 252, + 325, + 262 + ], + "score": 0.86, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 250, + 352, + 263 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 353, + 253, + 360, + 261 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 250, + 393, + 263 + ], + "score": 1.0, + "content": "(where", + "type": "text" + }, + { + "bbox": [ + 393, + 251, + 443, + 261 + ], + "score": 0.9, + "content": "m \\in 1 . . . M", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 250, + 466, + 263 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 466, + 251, + 478, + 261 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 179, + 274 + ], + "score": 1.0, + "content": "dimensionality of", + "type": "text" + }, + { + "bbox": [ + 180, + 264, + 186, + 271 + ], + "score": 0.65, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 261, + 256, + 274 + ], + "score": 1.0, + "content": "). Since our prior", + "type": "text" + }, + { + "bbox": [ + 256, + 262, + 275, + 273 + ], + "score": 0.91, + "content": "p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "is an isotropic unit Gaussian, we can re-write the second", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 273, + 175, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 175, + 285 + ], + "score": 1.0, + "content": "term of Eq. 6 as:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 238, + 506, + 285 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 286, + 470, + 316 + ], + "lines": [ + { + "bbox": [ + 140, + 286, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 140, + 286, + 470, + 316 + ], + "score": 0.93, + "content": "D _ { K L } ( q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) | | p ( \\mathbf { z } ) ) = \\int _ { z } q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) l o g { \\frac { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } { p ( \\mathbf { z } ) } } = \\sum _ { m } \\int _ { z _ { m } } q _ { \\phi } ( z _ { m } | \\mathbf { x } ) l o g { \\frac { q _ { \\phi } ( z _ { m } | \\mathbf { x } ) } { p ( z _ { m } ) } }", + "type": "interline_equation", + "image_path": "e81c3cac6d5e5e0c267ce4b360bbf68ba427f3f0bbbe1ef87ac2550100f403c7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 140, + 286, + 470, + 296.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 140, + 296.0, + 470, + 306.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 140, + 306.0, + 470, + 316.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 354, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 316, + 357, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 301, + 334 + ], + "score": 1.0, + "content": "Multiplying the second term in Eq. 8 by a factor", + "type": "text" + }, + { + "bbox": [ + 302, + 318, + 313, + 332 + ], + "score": 0.9, + "content": "\\textstyle { \\frac { M } { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 316, + 357, + 334 + ], + "score": 1.0, + "content": "produces:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 316, + 357, + 334 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 340, + 469, + 394 + ], + "lines": [ + { + "bbox": [ + 127, + 340, + 469, + 394 + ], + "spans": [ + { + "bbox": [ + 127, + 340, + 469, + 394 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\displaystyle \\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\boldsymbol { \\beta } ) \\propto \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } \\mathbb { E } _ { n } [ \\log p _ { \\boldsymbol { \\theta } } ( x _ { n } | \\mathbf { z } ) ] - \\frac { \\beta M } { N } \\mathbb { E } _ { m } \\int _ { z _ { m } } [ q _ { \\boldsymbol { \\phi } } ( z _ { m } | \\mathbf { x } ) l o g \\frac { q _ { \\boldsymbol { \\phi } } ( z _ { m } | \\mathbf { x } ) } { p ( z _ { m } ) } ] } \\\\ { \\displaystyle = \\mathbb { E } _ { q _ { \\boldsymbol { \\phi } } ( \\mathbf { z } | \\mathbf { x } ) } \\mathbb { E } _ { n } [ \\log p _ { \\boldsymbol { \\theta } } ( x _ { n } | \\mathbf { z } ) ] - \\frac { \\beta M } { N } \\mathbb { E } _ { m } [ D _ { K L } ( q _ { \\boldsymbol { \\phi } } ( z _ { m } | \\mathbf { x } ) | | p ( z _ { m } ) ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "d7e24fc15f48c2257ded2a302733eb54342bc38b94082b52323a09fd1c79deaf.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 127, + 340, + 469, + 358.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 127, + 358.0, + 469, + 376.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 127, + 376.0, + 469, + 394.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 158, + 409 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 159, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 159, + 411 + ], + "score": 1.0, + "content": "Hence using", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 395, + 159, + 411 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 421, + 336, + 446 + ], + "lines": [ + { + "bbox": [ + 275, + 421, + 336, + 446 + ], + "spans": [ + { + "bbox": [ + 275, + 421, + 336, + 446 + ], + "score": 0.92, + "content": "\\beta _ { n o r m } = { \\frac { \\beta M } { N } }", + "type": "interline_equation", + "image_path": "41bbb3f29b256fe799a9f371d32e1145c2c07d986a5c61015fc12a3ede1fb591.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 275, + 421, + 336, + 446 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 313, + 466 + ], + "score": 1.0, + "content": "in Eq. 10 is equivalent to optimising the original", + "type": "text" + }, + { + "bbox": [ + 313, + 454, + 321, + 465 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "-VAE formulation from Sec. 2, but with the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "additional independence assumptions that let us calculate data log likelihood and KL divergence", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 475, + 412, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 293, + 489 + ], + "score": 1.0, + "content": "terms in expectation over the individual pixels", + "type": "text" + }, + { + "bbox": [ + 293, + 477, + 305, + 487 + ], + "score": 0.85, + "content": "x _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 475, + 394, + 489 + ], + "score": 1.0, + "content": "and individual latents", + "type": "text" + }, + { + "bbox": [ + 395, + 478, + 408, + 487 + ], + "score": 0.85, + "content": "z _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 475, + 412, + 489 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 453, + 505, + 489 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 278, + 512 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 278, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 242, + 513 + ], + "score": 1.0, + "content": "A.7 RELATIONSHIP BETWEEN", + "type": "text" + }, + { + "bbox": [ + 243, + 501, + 250, + 512 + ], + "score": 0.77, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 499, + 272, + 513 + ], + "score": 1.0, + "content": "AND", + "type": "text" + }, + { + "bbox": [ + 273, + 503, + 278, + 510 + ], + "score": 0.39, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 151, + 534 + ], + "score": 1.0, + "content": "For a given", + "type": "text" + }, + { + "bbox": [ + 152, + 524, + 158, + 531 + ], + "score": 0.61, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 520, + 455, + 534 + ], + "score": 1.0, + "content": "we can solve the constrained optimisation problem in Eq. 3 (find the optimal", + "type": "text" + }, + { + "bbox": [ + 456, + 521, + 502, + 533 + ], + "score": 0.91, + "content": "( \\theta ^ { * } , \\phi ^ { * } , \\beta ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 520, + 506, + 534 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 143, + 545 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 144, + 532, + 229, + 544 + ], + "score": 0.92, + "content": "\\Delta \\mathcal { F } ( \\theta ^ { * } , \\phi ^ { * } , \\beta ^ { * } ) = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 531, + 505, + 545 + ], + "score": 1.0, + "content": ". We can then re-write our optimal solution to the original optimisation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 543, + 250, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 240, + 555 + ], + "score": 1.0, + "content": "problem in Eq. 2 as a function of", + "type": "text" + }, + { + "bbox": [ + 241, + 546, + 245, + 552 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 543, + 250, + 555 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 520, + 506, + 555 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 560, + 400, + 575 + ], + "lines": [ + { + "bbox": [ + 211, + 560, + 400, + 575 + ], + "spans": [ + { + "bbox": [ + 211, + 560, + 400, + 575 + ], + "score": 0.93, + "content": "\\mathcal { G } ( \\theta ^ { * } ( \\epsilon ) , \\phi ^ { * } ( \\epsilon ) ) = \\mathbb { E } _ { q _ { \\phi ^ { * } ( \\epsilon ) } ( \\mathbf { z } | \\mathbf { x } ) } [ \\log p _ { \\theta ^ { * } ( \\epsilon ) } ( \\mathbf { x } | \\mathbf { z } ) ]", + "type": "interline_equation", + "image_path": "5782f03be4ccac442a47081021c560a4bfb31bb312f502656f5b33bd40ceec47.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 211, + 560, + 400, + 575 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 503, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 127, + 593 + ], + "score": 1.0, + "content": "Now", + "type": "text" + }, + { + "bbox": [ + 128, + 581, + 135, + 591 + ], + "score": 0.86, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 579, + 383, + 593 + ], + "score": 1.0, + "content": "can be interpreted as the rate of change of the optimal solution", + "type": "text" + }, + { + "bbox": [ + 383, + 580, + 415, + 591 + ], + "score": 0.91, + "content": "( \\theta ^ { * } , \\phi ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 579, + 426, + 593 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 426, + 581, + 434, + 591 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 579, + 505, + 593 + ], + "score": 1.0, + "content": "when varying the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 591, + 158, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 148, + 603 + ], + "score": 1.0, + "content": "constraint", + "type": "text" + }, + { + "bbox": [ + 149, + 594, + 154, + 601 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 591, + 158, + 603 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 579, + 505, + 603 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 280, + 598, + 331, + 623 + ], + "lines": [ + { + "bbox": [ + 280, + 598, + 331, + 623 + ], + "spans": [ + { + "bbox": [ + 280, + 598, + 331, + 623 + ], + "score": 0.94, + "content": "\\frac { \\delta \\mathcal { G } } { \\delta \\epsilon } = \\beta ^ { * } ( \\epsilon )", + "type": "interline_equation", + "image_path": "36273b9b9e57ff492cffd7b3a9b88fb73aaeb6e4ff3540ed9ca99e78438d3c02.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 280, + 598, + 331, + 623 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 634, + 214, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 216, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 216, + 647 + ], + "score": 1.0, + "content": "A.8 DATA CONTINUITY", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "We hypothesise that data continuity plays a role in guiding unsupervised models towards learning the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "correct data manifolds. To test this idea we measure how the degree of learnt disentangling changes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 362, + 690 + ], + "score": 1.0, + "content": "with reduced continuity in the 2D shapes dataset. We trained a", + "type": "text" + }, + { + "bbox": [ + 363, + 678, + 370, + 688 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 676, + 414, + 690 + ], + "score": 1.0, + "content": "-VAE with", + "type": "text" + }, + { + "bbox": [ + 415, + 677, + 441, + 688 + ], + "score": 0.9, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "(Figure 7A) on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "subsamples of the original 2D shapes dataset, where we progressively decreased the generative factor", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "sampling density. Reduction in data continuity negatively correlates with the average pixel wise", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "(Hamming) distance between two consecutive transforms of each object (normalised by the average", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "number of pixels occupied by each of the two adjacent transforms of an object to account for object", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 83, + 504, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 504, + 94 + ], + "score": 1.0, + "content": "scale). Figure 8 demonstrates that as the continuity in the data reduces, the degree of disentanglement", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "in the learnt representations also drops. This effect holds after additional hyperparameter tuning and", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "can not solely be explained by the decrease in dataset size, since the same VAE can learn disentangled", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 457, + 128 + ], + "score": 1.0, + "content": "representations from a data subset that preserves data continuity but is approximately", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 457, + 115, + 478, + 126 + ], + "score": 0.89, + "content": "55 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 478, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "of the", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 238, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 238, + 139 + ], + "score": 1.0, + "content": "original size (results not shown).", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 654, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 504, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 504, + 94 + ], + "score": 1.0, + "content": "scale). Figure 8 demonstrates that as the continuity in the data reduces, the degree of disentanglement", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "in the learnt representations also drops. This effect holds after additional hyperparameter tuning and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "can not solely be explained by the decrease in dataset size, since the same VAE can learn disentangled", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 457, + 128 + ], + "score": 1.0, + "content": "representations from a data subset that preserves data continuity but is approximately", + "type": "text" + }, + { + "bbox": [ + 457, + 115, + 478, + 126 + ], + "score": 0.89, + "content": "55 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 238, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 238, + 139 + ], + "score": 1.0, + "content": "original size (results not shown).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 184, + 163, + 424, + 367 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 184, + 163, + 424, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 163, + 424, + 367 + ], + "spans": [ + { + "bbox": [ + 184, + 163, + 424, + 367 + ], + "score": 0.971, + "type": "image", + "image_path": "459a24497b2b3896af58ec63ae32465f95812d23a12e44b5ccde1d283f06622a.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 184, + 163, + 424, + 176.6 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 184, + 176.6, + 424, + 190.2 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 184, + 190.2, + 424, + 203.79999999999998 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 184, + 203.79999999999998, + 424, + 217.39999999999998 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 184, + 217.39999999999998, + 424, + 230.99999999999997 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 184, + 230.99999999999997, + 424, + 244.59999999999997 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 184, + 244.59999999999997, + 424, + 258.2 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 184, + 258.2, + 424, + 271.8 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 184, + 271.8, + 424, + 285.40000000000003 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 184, + 285.40000000000003, + 424, + 299.00000000000006 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 184, + 299.00000000000006, + 424, + 312.6000000000001 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 184, + 312.6000000000001, + 424, + 326.2000000000001 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 184, + 326.2000000000001, + 424, + 339.8000000000001 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 184, + 339.8000000000001, + 424, + 353.40000000000015 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 184, + 353.40000000000015, + 424, + 367.00000000000017 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 374, + 506, + 462 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "Figure 8: Negative correlation between data transform continuity and the degree of disentangling", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 158, + 397 + ], + "score": 1.0, + "content": "achieved by", + "type": "text" + }, + { + "bbox": [ + 158, + 385, + 166, + 396 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "-VAE. Abscissa is the average normalized Hamming distance between each of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "two consecutive transforms of each object. Ordinate is disentanglement metric score. Disentangling", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 406, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 406, + 506, + 421 + ], + "score": 1.0, + "content": "performance is robust to Bernoulli noise added to the data at test time, as shown by slowly degrading", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 225, + 432 + ], + "score": 1.0, + "content": "classification accuracy up to", + "type": "text" + }, + { + "bbox": [ + 225, + 418, + 245, + 429 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 417, + 506, + 432 + ], + "score": 1.0, + "content": "noise level, considering that the 2D objects occupy on average", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 142, + 441 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 143, + 429, + 166, + 439 + ], + "score": 0.85, + "content": "2 - 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "of the image depending on scale. Fluctuations in classification accuracy for similar", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "Hamming distances are due the different nature of subsampled generative factors (i.e. symmetries are", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 294, + 463 + ], + "score": 1.0, + "content": "present in rotation but are lacking in position).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + } + ], + "index": 17.75 + }, + { + "type": "title", + "bbox": [ + 108, + 482, + 207, + 494 + ], + "lines": [ + { + "bbox": [ + 105, + 481, + 208, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 132, + 496 + ], + "score": 1.0, + "content": "A.9", + "type": "text" + }, + { + "bbox": [ + 133, + 483, + 140, + 494 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 481, + 208, + 496 + ], + "score": 1.0, + "content": "-VAE SAMPLES", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 164, + 516 + ], + "score": 1.0, + "content": "Samples from", + "type": "text" + }, + { + "bbox": [ + 164, + 504, + 172, + 515 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 502, + 293, + 516 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled", + "type": "text" + }, + { + "bbox": [ + 293, + 504, + 320, + 515 + ], + "score": 0.86, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 502, + 384, + 516 + ], + "score": 1.0, + "content": ") and entangled", + "type": "text" + }, + { + "bbox": [ + 384, + 504, + 411, + 515 + ], + "score": 0.87, + "content": "\\beta = 1 \\mathrm { \\hbar }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 502, + 505, + 516 + ], + "score": 1.0, + "content": ") representations can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 514, + 176, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 176, + 527 + ], + "score": 1.0, + "content": "seen in Figure 9.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 108, + 539, + 284, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 539, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 171, + 552 + ], + "score": 1.0, + "content": "A.10 EXTRA", + "type": "text" + }, + { + "bbox": [ + 171, + 540, + 178, + 551 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 539, + 286, + 552 + ], + "score": 1.0, + "content": "-VAE TRAVERSAL PLOTS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 285, + 573 + ], + "score": 1.0, + "content": "We present extra latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 285, + 560, + 292, + 571 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations of 3D", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "chairs (Figures 10-11) and CelebA (Figures 12-14) datasets. Here we show traversals from all", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 580, + 331, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 331, + 595 + ], + "score": 1.0, + "content": "informative latents from a large number of seed images.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 138 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 104, + 83, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 184, + 163, + 424, + 367 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 184, + 163, + 424, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 163, + 424, + 367 + ], + "spans": [ + { + "bbox": [ + 184, + 163, + 424, + 367 + ], + "score": 0.971, + "type": "image", + "image_path": "459a24497b2b3896af58ec63ae32465f95812d23a12e44b5ccde1d283f06622a.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 184, + 163, + 424, + 176.6 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 184, + 176.6, + 424, + 190.2 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 184, + 190.2, + 424, + 203.79999999999998 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 184, + 203.79999999999998, + 424, + 217.39999999999998 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 184, + 217.39999999999998, + 424, + 230.99999999999997 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 184, + 230.99999999999997, + 424, + 244.59999999999997 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 184, + 244.59999999999997, + 424, + 258.2 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 184, + 258.2, + 424, + 271.8 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 184, + 271.8, + 424, + 285.40000000000003 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 184, + 285.40000000000003, + 424, + 299.00000000000006 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 184, + 299.00000000000006, + 424, + 312.6000000000001 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 184, + 312.6000000000001, + 424, + 326.2000000000001 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 184, + 326.2000000000001, + 424, + 339.8000000000001 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 184, + 339.8000000000001, + 424, + 353.40000000000015 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 184, + 353.40000000000015, + 424, + 367.00000000000017 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 374, + 506, + 462 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "score": 1.0, + "content": "Figure 8: Negative correlation between data transform continuity and the degree of disentangling", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 158, + 397 + ], + "score": 1.0, + "content": "achieved by", + "type": "text" + }, + { + "bbox": [ + 158, + 385, + 166, + 396 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "-VAE. Abscissa is the average normalized Hamming distance between each of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "two consecutive transforms of each object. Ordinate is disentanglement metric score. Disentangling", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 406, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 406, + 506, + 421 + ], + "score": 1.0, + "content": "performance is robust to Bernoulli noise added to the data at test time, as shown by slowly degrading", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 225, + 432 + ], + "score": 1.0, + "content": "classification accuracy up to", + "type": "text" + }, + { + "bbox": [ + 225, + 418, + 245, + 429 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 417, + 506, + 432 + ], + "score": 1.0, + "content": "noise level, considering that the 2D objects occupy on average", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 142, + 441 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 143, + 429, + 166, + 439 + ], + "score": 0.85, + "content": "2 - 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "of the image depending on scale. Fluctuations in classification accuracy for similar", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "Hamming distances are due the different nature of subsampled generative factors (i.e. symmetries are", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 294, + 463 + ], + "score": 1.0, + "content": "present in rotation but are lacking in position).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + } + ], + "index": 17.75 + }, + { + "type": "title", + "bbox": [ + 108, + 482, + 207, + 494 + ], + "lines": [ + { + "bbox": [ + 105, + 481, + 208, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 132, + 496 + ], + "score": 1.0, + "content": "A.9", + "type": "text" + }, + { + "bbox": [ + 133, + 483, + 140, + 494 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 481, + 208, + 496 + ], + "score": 1.0, + "content": "-VAE SAMPLES", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 164, + 516 + ], + "score": 1.0, + "content": "Samples from", + "type": "text" + }, + { + "bbox": [ + 164, + 504, + 172, + 515 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 502, + 293, + 516 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled", + "type": "text" + }, + { + "bbox": [ + 293, + 504, + 320, + 515 + ], + "score": 0.86, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 502, + 384, + 516 + ], + "score": 1.0, + "content": ") and entangled", + "type": "text" + }, + { + "bbox": [ + 384, + 504, + 411, + 515 + ], + "score": 0.87, + "content": "\\beta = 1 \\mathrm { \\hbar }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 502, + 505, + 516 + ], + "score": 1.0, + "content": ") representations can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 514, + 176, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 176, + 527 + ], + "score": 1.0, + "content": "seen in Figure 9.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 502, + 505, + 527 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 539, + 284, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 539, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 171, + 552 + ], + "score": 1.0, + "content": "A.10 EXTRA", + "type": "text" + }, + { + "bbox": [ + 171, + 540, + 178, + 551 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 539, + 286, + 552 + ], + "score": 1.0, + "content": "-VAE TRAVERSAL PLOTS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 285, + 573 + ], + "score": 1.0, + "content": "We present extra latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 285, + 560, + 292, + 571 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations of 3D", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "chairs (Figures 10-11) and CelebA (Figures 12-14) datasets. Here we show traversals from all", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 580, + 331, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 331, + 595 + ], + "score": 1.0, + "content": "informative latents from a large number of seed images.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 559, + 505, + 595 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 258, + 483, + 491 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 258, + 483, + 491 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 258, + 483, + 491 + ], + "spans": [ + { + "bbox": [ + 125, + 258, + 483, + 491 + ], + "score": 0.975, + "type": "image", + "image_path": "ae67406e8913a55260a5562a058ee2699d2b4069b4ef385b295db2de9abbca03.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 258, + 483, + 335.6666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 335.6666666666667, + 483, + 413.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 413.33333333333337, + 483, + 491.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 495, + 506, + 550 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 204, + 506 + ], + "score": 1.0, + "content": "Figure 9: Samples from", + "type": "text" + }, + { + "bbox": [ + 204, + 495, + 212, + 506 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "-VAE trained on the dataset of 2D shapes that learnt either a disentangled", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 128, + 518 + ], + "score": 1.0, + "content": "(left,", + "type": "text" + }, + { + "bbox": [ + 128, + 506, + 155, + 517 + ], + "score": 0.89, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 506, + 252, + 518 + ], + "score": 1.0, + "content": ") or an entangled (right,", + "type": "text" + }, + { + "bbox": [ + 253, + 506, + 279, + 517 + ], + "score": 0.9, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 506, + 505, + 518 + ], + "score": 1.0, + "content": ") representation of the data generative factors. It can be", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "seen that sampling from an entangled representation results in some unrealistic looking samples. A", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 526, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 541 + ], + "score": 1.0, + "content": "disentangled representation that inverts the original data generation process does not suffer from such", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 538, + 135, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 135, + 551 + ], + "score": 1.0, + "content": "errors.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 258, + 483, + 491 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 258, + 483, + 491 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 258, + 483, + 491 + ], + "spans": [ + { + "bbox": [ + 125, + 258, + 483, + 491 + ], + "score": 0.975, + "type": "image", + "image_path": "ae67406e8913a55260a5562a058ee2699d2b4069b4ef385b295db2de9abbca03.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 258, + 483, + 335.6666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 335.6666666666667, + 483, + 413.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 413.33333333333337, + 483, + 491.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 495, + 506, + 550 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 204, + 506 + ], + "score": 1.0, + "content": "Figure 9: Samples from", + "type": "text" + }, + { + "bbox": [ + 204, + 495, + 212, + 506 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "-VAE trained on the dataset of 2D shapes that learnt either a disentangled", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 128, + 518 + ], + "score": 1.0, + "content": "(left,", + "type": "text" + }, + { + "bbox": [ + 128, + 506, + 155, + 517 + ], + "score": 0.89, + "content": "\\beta = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 506, + 252, + 518 + ], + "score": 1.0, + "content": ") or an entangled (right,", + "type": "text" + }, + { + "bbox": [ + 253, + 506, + 279, + 517 + ], + "score": 0.9, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 506, + 505, + 518 + ], + "score": 1.0, + "content": ") representation of the data generative factors. It can be", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "seen that sampling from an entangled representation results in some unrealistic looking samples. A", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 526, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 541 + ], + "score": 1.0, + "content": "disentangled representation that inverts the original data generation process does not suffer from such", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 538, + 135, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 135, + 551 + ], + "score": 1.0, + "content": "errors.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 109, + 480, + 674 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 130, + 109, + 480, + 674 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 109, + 480, + 674 + ], + "spans": [ + { + "bbox": [ + 130, + 109, + 480, + 674 + ], + "score": 0.917, + "type": "image", + "image_path": "9049f21d4258d2516eb480744efc9e5f4109fd0f5a53c4fd5d1945d693c566ff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 130, + 109, + 480, + 297.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 130, + 297.33333333333337, + 480, + 485.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 130, + 485.66666666666674, + 480, + 674.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 680, + 505, + 703 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 266, + 693 + ], + "score": 1.0, + "content": "Figure 10: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 266, + 681, + 273, + 692 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the 3D", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 691, + 165, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 165, + 703 + ], + "score": 1.0, + "content": "chairs dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 109, + 480, + 674 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 130, + 109, + 480, + 674 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 109, + 480, + 674 + ], + "spans": [ + { + "bbox": [ + 130, + 109, + 480, + 674 + ], + "score": 0.917, + "type": "image", + "image_path": "9049f21d4258d2516eb480744efc9e5f4109fd0f5a53c4fd5d1945d693c566ff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 130, + 109, + 480, + 297.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 130, + 297.33333333333337, + 480, + 485.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 130, + 485.66666666666674, + 480, + 674.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 680, + 505, + 703 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 266, + 693 + ], + "score": 1.0, + "content": "Figure 10: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 266, + 681, + 273, + 692 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the 3D", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 691, + 165, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 165, + 703 + ], + "score": 1.0, + "content": "chairs dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 105, + 480, + 672 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 105, + 480, + 672 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 105, + 480, + 672 + ], + "spans": [ + { + "bbox": [ + 128, + 105, + 480, + 672 + ], + "score": 0.938, + "type": "image", + "image_path": "1a2cc13dc4e5fa96eb6370dc49a989426f7b720c781798e806809d71f5d4675c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 105, + 480, + 294.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 294.0, + 480, + 483.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 483.0, + 480, + 672.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 678, + 505, + 701 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 266, + 692 + ], + "score": 1.0, + "content": "Figure 11: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 266, + 680, + 273, + 691 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 677, + 505, + 692 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the 3D", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 690, + 165, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 165, + 702 + ], + "score": 1.0, + "content": "chairs dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 105, + 480, + 672 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 105, + 480, + 672 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 105, + 480, + 672 + ], + "spans": [ + { + "bbox": [ + 128, + 105, + 480, + 672 + ], + "score": 0.938, + "type": "image", + "image_path": "1a2cc13dc4e5fa96eb6370dc49a989426f7b720c781798e806809d71f5d4675c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 105, + 480, + 294.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 294.0, + 480, + 483.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 483.0, + 480, + 672.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 678, + 505, + 701 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 266, + 692 + ], + "score": 1.0, + "content": "Figure 11: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 266, + 680, + 273, + 691 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 677, + 505, + 692 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the 3D", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 690, + 165, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 165, + 702 + ], + "score": 1.0, + "content": "chairs dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 128, + 489, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 128, + 489, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 128, + 489, + 654 + ], + "spans": [ + { + "bbox": [ + 118, + 128, + 489, + 654 + ], + "score": 0.972, + "type": "image", + "image_path": "ef3fae64fc29c55f9c1a18226beb0eb9b8bf0d442519e028e1cb69113cd260a7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 128, + 489, + 303.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 303.33333333333337, + 489, + 478.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 478.66666666666674, + 489, + 654.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 655, + 504, + 678 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 258, + 668 + ], + "score": 1.0, + "content": "Figure 12: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 258, + 657, + 265, + 667 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the CelebA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 128, + 489, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 128, + 489, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 128, + 489, + 654 + ], + "spans": [ + { + "bbox": [ + 118, + 128, + 489, + 654 + ], + "score": 0.972, + "type": "image", + "image_path": "ef3fae64fc29c55f9c1a18226beb0eb9b8bf0d442519e028e1cb69113cd260a7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 128, + 489, + 303.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 303.33333333333337, + 489, + 478.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 478.66666666666674, + 489, + 654.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 655, + 504, + 678 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 258, + 668 + ], + "score": 1.0, + "content": "Figure 12: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 258, + 657, + 265, + 667 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the CelebA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 131, + 488, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 131, + 488, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 131, + 488, + 654 + ], + "spans": [ + { + "bbox": [ + 119, + 131, + 488, + 654 + ], + "score": 0.971, + "type": "image", + "image_path": "01703dbdb4f3532ca41bf9156e8f6ca60335af5d67cbb56aab3078a7e2c7b3c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 131, + 488, + 305.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 305.33333333333337, + 488, + 479.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 479.66666666666674, + 488, + 654.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 655, + 504, + 678 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 258, + 668 + ], + "score": 1.0, + "content": "Figure 13: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 258, + 657, + 265, + 667 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the CelebA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 131, + 488, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 131, + 488, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 131, + 488, + 654 + ], + "spans": [ + { + "bbox": [ + 119, + 131, + 488, + 654 + ], + "score": 0.971, + "type": "image", + "image_path": "01703dbdb4f3532ca41bf9156e8f6ca60335af5d67cbb56aab3078a7e2c7b3c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 131, + 488, + 305.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 305.33333333333337, + 488, + 479.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 479.66666666666674, + 488, + 654.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 655, + 504, + 678 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 258, + 668 + ], + "score": 1.0, + "content": "Figure 13: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 258, + 657, + 265, + 667 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the CelebA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 141, + 680 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 114, + 443, + 685 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 114, + 443, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 114, + 443, + 685 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 443, + 685 + ], + "score": 0.967, + "type": "image", + "image_path": "f61801f2b3e9448756df3fb7e5d52bdc561084346f95faae387be2246ff24200.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 165, + 114, + 443, + 126.68888888888888 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 165, + 126.68888888888888, + 443, + 139.37777777777777 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 165, + 139.37777777777777, + 443, + 152.06666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 165, + 152.06666666666666, + 443, + 164.75555555555556 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 165, + 164.75555555555556, + 443, + 177.44444444444446 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 165, + 177.44444444444446, + 443, + 190.13333333333335 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 165, + 190.13333333333335, + 443, + 202.82222222222225 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 165, + 202.82222222222225, + 443, + 215.51111111111115 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 165, + 215.51111111111115, + 443, + 228.20000000000005 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 165, + 228.20000000000005, + 443, + 240.88888888888894 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 165, + 240.88888888888894, + 443, + 253.57777777777784 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 165, + 253.57777777777784, + 443, + 266.2666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 165, + 266.2666666666667, + 443, + 278.9555555555556 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 165, + 278.9555555555556, + 443, + 291.64444444444445 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 165, + 291.64444444444445, + 443, + 304.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 165, + 304.3333333333333, + 443, + 317.0222222222222 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 165, + 317.0222222222222, + 443, + 329.71111111111105 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 165, + 329.71111111111105, + 443, + 342.3999999999999 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 165, + 342.3999999999999, + 443, + 355.0888888888888 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 165, + 355.0888888888888, + 443, + 367.77777777777766 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 165, + 367.77777777777766, + 443, + 380.4666666666665 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 165, + 380.4666666666665, + 443, + 393.1555555555554 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 165, + 393.1555555555554, + 443, + 405.84444444444426 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 165, + 405.84444444444426, + 443, + 418.53333333333313 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 165, + 418.53333333333313, + 443, + 431.222222222222 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 165, + 431.222222222222, + 443, + 443.91111111111087 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 165, + 443.91111111111087, + 443, + 456.59999999999974 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 165, + 456.59999999999974, + 443, + 469.2888888888886 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 165, + 469.2888888888886, + 443, + 481.9777777777775 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 165, + 481.9777777777775, + 443, + 494.66666666666634 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 165, + 494.66666666666634, + 443, + 507.3555555555552 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 165, + 507.3555555555552, + 443, + 520.0444444444441 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 165, + 520.0444444444441, + 443, + 532.733333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 165, + 532.733333333333, + 443, + 545.4222222222219 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 165, + 545.4222222222219, + 443, + 558.1111111111107 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 165, + 558.1111111111107, + 443, + 570.7999999999996 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 165, + 570.7999999999996, + 443, + 583.4888888888885 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 165, + 583.4888888888885, + 443, + 596.1777777777774 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 165, + 596.1777777777774, + 443, + 608.8666666666662 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 165, + 608.8666666666662, + 443, + 621.5555555555551 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 165, + 621.5555555555551, + 443, + 634.244444444444 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 165, + 634.244444444444, + 443, + 646.9333333333328 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 165, + 646.9333333333328, + 443, + 659.6222222222217 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 165, + 659.6222222222217, + 443, + 672.3111111111106 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 165, + 672.3111111111106, + 443, + 684.9999999999994 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 689, + 505, + 712 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 258, + 702 + ], + "score": 1.0, + "content": "Figure 14: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 258, + 690, + 265, + 701 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the CelebA", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 700, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 140, + 712 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "index": 33.75 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "22", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 114, + 443, + 685 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 114, + 443, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 114, + 443, + 685 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 443, + 685 + ], + "score": 0.967, + "type": "image", + "image_path": "f61801f2b3e9448756df3fb7e5d52bdc561084346f95faae387be2246ff24200.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 165, + 114, + 443, + 126.68888888888888 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 165, + 126.68888888888888, + 443, + 139.37777777777777 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 165, + 139.37777777777777, + 443, + 152.06666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 165, + 152.06666666666666, + 443, + 164.75555555555556 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 165, + 164.75555555555556, + 443, + 177.44444444444446 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 165, + 177.44444444444446, + 443, + 190.13333333333335 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 165, + 190.13333333333335, + 443, + 202.82222222222225 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 165, + 202.82222222222225, + 443, + 215.51111111111115 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 165, + 215.51111111111115, + 443, + 228.20000000000005 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 165, + 228.20000000000005, + 443, + 240.88888888888894 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 165, + 240.88888888888894, + 443, + 253.57777777777784 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 165, + 253.57777777777784, + 443, + 266.2666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 165, + 266.2666666666667, + 443, + 278.9555555555556 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 165, + 278.9555555555556, + 443, + 291.64444444444445 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 165, + 291.64444444444445, + 443, + 304.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 165, + 304.3333333333333, + 443, + 317.0222222222222 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 165, + 317.0222222222222, + 443, + 329.71111111111105 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 165, + 329.71111111111105, + 443, + 342.3999999999999 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 165, + 342.3999999999999, + 443, + 355.0888888888888 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 165, + 355.0888888888888, + 443, + 367.77777777777766 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 165, + 367.77777777777766, + 443, + 380.4666666666665 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 165, + 380.4666666666665, + 443, + 393.1555555555554 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 165, + 393.1555555555554, + 443, + 405.84444444444426 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 165, + 405.84444444444426, + 443, + 418.53333333333313 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 165, + 418.53333333333313, + 443, + 431.222222222222 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 165, + 431.222222222222, + 443, + 443.91111111111087 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 165, + 443.91111111111087, + 443, + 456.59999999999974 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 165, + 456.59999999999974, + 443, + 469.2888888888886 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 165, + 469.2888888888886, + 443, + 481.9777777777775 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 165, + 481.9777777777775, + 443, + 494.66666666666634 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 165, + 494.66666666666634, + 443, + 507.3555555555552 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 165, + 507.3555555555552, + 443, + 520.0444444444441 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 165, + 520.0444444444441, + 443, + 532.733333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 165, + 532.733333333333, + 443, + 545.4222222222219 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 165, + 545.4222222222219, + 443, + 558.1111111111107 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 165, + 558.1111111111107, + 443, + 570.7999999999996 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 165, + 570.7999999999996, + 443, + 583.4888888888885 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 165, + 583.4888888888885, + 443, + 596.1777777777774 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 165, + 596.1777777777774, + 443, + 608.8666666666662 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 165, + 608.8666666666662, + 443, + 621.5555555555551 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 165, + 621.5555555555551, + 443, + 634.244444444444 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 165, + 634.244444444444, + 443, + 646.9333333333328 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 165, + 646.9333333333328, + 443, + 659.6222222222217 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 165, + 659.6222222222217, + 443, + 672.3111111111106 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 165, + 672.3111111111106, + 443, + 684.9999999999994 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 689, + 505, + 712 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 258, + 702 + ], + "score": 1.0, + "content": "Figure 14: Latent traversal plots from", + "type": "text" + }, + { + "bbox": [ + 258, + 690, + 265, + 701 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "-VAE that learnt disentangled representations on the CelebA", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 700, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 140, + 712 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "index": 33.75 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/TR-Nj6nFx42/images/18df541cfe669e305832cff69e6cf0118df269b237d1cab46980e21554ebbbc1.jpg b/parse/train/TR-Nj6nFx42/images/18df541cfe669e305832cff69e6cf0118df269b237d1cab46980e21554ebbbc1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..96177b71522bf5afc0129b17c5eb86fb3d9918b7 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/18df541cfe669e305832cff69e6cf0118df269b237d1cab46980e21554ebbbc1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb247879b4664b0d552855b59a9b3b8ae66f67cdb7e2a8fed4bb7f83425abc47 +size 12166 diff --git a/parse/train/TR-Nj6nFx42/images/324b7454be40ccdbb26b950c68d159de50527bb16041c9bcef150f22511bb899.jpg b/parse/train/TR-Nj6nFx42/images/324b7454be40ccdbb26b950c68d159de50527bb16041c9bcef150f22511bb899.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3c5e4921a0fd23220b36269721b1921c90013f6d --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/324b7454be40ccdbb26b950c68d159de50527bb16041c9bcef150f22511bb899.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c40c79b51603ed830db05ace718b2610dd6b8211ac923eb06b11767db075715 +size 8977 diff --git a/parse/train/TR-Nj6nFx42/images/6d8d1d796b7d45aaef598ae8923862b8eba24eac1424fe03bbbfa0985906dd90.jpg b/parse/train/TR-Nj6nFx42/images/6d8d1d796b7d45aaef598ae8923862b8eba24eac1424fe03bbbfa0985906dd90.jpg new file mode 100644 index 0000000000000000000000000000000000000000..128661b215d35b45bcf44b2373372592e10748dd --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/6d8d1d796b7d45aaef598ae8923862b8eba24eac1424fe03bbbfa0985906dd90.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84513c6e8ae64096d369da1cfa3da8c78391043e90b5c7800031fe9a99b3aff9 +size 10099 diff --git a/parse/train/TR-Nj6nFx42/images/7621f5366256f6ae0811ed0acc4a828e58d33cc2c27fd39249c3084bba6cb23e.jpg b/parse/train/TR-Nj6nFx42/images/7621f5366256f6ae0811ed0acc4a828e58d33cc2c27fd39249c3084bba6cb23e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a1b800b135ba6e29c312bcbb1199f4a611c8d643 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/7621f5366256f6ae0811ed0acc4a828e58d33cc2c27fd39249c3084bba6cb23e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b416dbe545aa1ca8ee2ce3c514888643baf6a453281547eb3973325d5b00a624 +size 8041 diff --git a/parse/train/TR-Nj6nFx42/images/7be05ceaf80ce285415f86cec3ffff19ebae7deedb0d23cb76fedb4c9744dea7.jpg b/parse/train/TR-Nj6nFx42/images/7be05ceaf80ce285415f86cec3ffff19ebae7deedb0d23cb76fedb4c9744dea7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..66fcac0450599dd65c9e41bacc8a9bf39ea69358 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/7be05ceaf80ce285415f86cec3ffff19ebae7deedb0d23cb76fedb4c9744dea7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5d75b7349c3629e78b27530b2f4bfa9d39438a815c38f795bec2464f25afa0a +size 44904 diff --git a/parse/train/TR-Nj6nFx42/images/8b52a3bd00dad5ba7dc94d5f3a344da9cf66434da72159a4be8504309b499f06.jpg b/parse/train/TR-Nj6nFx42/images/8b52a3bd00dad5ba7dc94d5f3a344da9cf66434da72159a4be8504309b499f06.jpg new file mode 100644 index 0000000000000000000000000000000000000000..76081645d1339e7325dd9eb16f3c3042d8fe50ae --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/8b52a3bd00dad5ba7dc94d5f3a344da9cf66434da72159a4be8504309b499f06.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fa5cfce7ece4617e401f478128a5b258a964f3d4dea8b61245edbed45e56c45 +size 7329 diff --git a/parse/train/TR-Nj6nFx42/images/95be847b2d36083b196f396f056385780a53c2f8a4d13d7087e4c5159d026b28.jpg b/parse/train/TR-Nj6nFx42/images/95be847b2d36083b196f396f056385780a53c2f8a4d13d7087e4c5159d026b28.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9d108c434cfdeebf6534209b1510902c47bdeabd --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/95be847b2d36083b196f396f056385780a53c2f8a4d13d7087e4c5159d026b28.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f15042c8b9f23baa4bf111926e368edf0fac8b497227c769e96d662dcab8ab4 +size 8386 diff --git a/parse/train/TR-Nj6nFx42/images/a4641c320cb6eb6531f4c0ca323f0711393f5965299ab8e6c068d2f45d8fca16.jpg b/parse/train/TR-Nj6nFx42/images/a4641c320cb6eb6531f4c0ca323f0711393f5965299ab8e6c068d2f45d8fca16.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e00566d69a06089c99ce991948af1a372aba3396 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/a4641c320cb6eb6531f4c0ca323f0711393f5965299ab8e6c068d2f45d8fca16.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9f745558adeb0a38a2e634326040edb1a24b7f730b5289c87af93f4accb048c +size 16217 diff --git a/parse/train/TR-Nj6nFx42/images/c6a5037a42fcb9e9bd85c963360d2f6a82bedcbda00a3c65110067636e1a3c28.jpg b/parse/train/TR-Nj6nFx42/images/c6a5037a42fcb9e9bd85c963360d2f6a82bedcbda00a3c65110067636e1a3c28.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f43fee3438e66a131ae88ef404510d227a19febf --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/c6a5037a42fcb9e9bd85c963360d2f6a82bedcbda00a3c65110067636e1a3c28.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29304db19fbbc293bf4e71e9e3880e654b24e22a0886b53317b6e6cfdd9b09c0 +size 18636 diff --git a/parse/train/TR-Nj6nFx42/images/c7ac79a96a7a58fce444c5b2e9b5ebec923e705321f9aff773a02688d6eb2d63.jpg b/parse/train/TR-Nj6nFx42/images/c7ac79a96a7a58fce444c5b2e9b5ebec923e705321f9aff773a02688d6eb2d63.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4f0c3623715d6605ae364453a672abc8147cd444 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/c7ac79a96a7a58fce444c5b2e9b5ebec923e705321f9aff773a02688d6eb2d63.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9650de5c98a9a7202d451ebb3e8c50dfaf6155375528051f559d38f0b9b6a1a +size 25320 diff --git a/parse/train/TR-Nj6nFx42/images/cf18a0a173328712ae1d21a00076e376dcd3fc5e2728d54f05aca4e3eb535a40.jpg b/parse/train/TR-Nj6nFx42/images/cf18a0a173328712ae1d21a00076e376dcd3fc5e2728d54f05aca4e3eb535a40.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bff4f5c7200b8a26687d658a0d86506ecedd8886 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/cf18a0a173328712ae1d21a00076e376dcd3fc5e2728d54f05aca4e3eb535a40.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:410a8f8f6fa9a49d3713b5f05734c2580e48b69013372fdc85f78fbf5701de3a +size 53238 diff --git a/parse/train/TR-Nj6nFx42/images/d72335541cd2921528b3b3daf831ca2b8c2fad831da58a4b5006826ef2eb31f5.jpg b/parse/train/TR-Nj6nFx42/images/d72335541cd2921528b3b3daf831ca2b8c2fad831da58a4b5006826ef2eb31f5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ca0013234f105185cb565afe9f24a3a69f22b10f --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/d72335541cd2921528b3b3daf831ca2b8c2fad831da58a4b5006826ef2eb31f5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55bb63ca2392c27f112c8dc19e1f6541dc2ea0a121fa566c59047df3dd77a4b0 +size 9524 diff --git a/parse/train/TR-Nj6nFx42/images/ddacf20645ab403baeb8121a9f3062c336868b88a79648be603d121a4934e518.jpg b/parse/train/TR-Nj6nFx42/images/ddacf20645ab403baeb8121a9f3062c336868b88a79648be603d121a4934e518.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6f8aa3057dfc0811a18bd11f04ecc1a540b3afc0 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/ddacf20645ab403baeb8121a9f3062c336868b88a79648be603d121a4934e518.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4b68d6786dca3f144d2852cb4b84276074ad1cc2ec7e4293986ca7ed0c6ea66 +size 120470 diff --git a/parse/train/TR-Nj6nFx42/images/e0a0db331f90bf249f3e076f3a85f6f14f76ef3a0854655f4aff9d815dc3413a.jpg b/parse/train/TR-Nj6nFx42/images/e0a0db331f90bf249f3e076f3a85f6f14f76ef3a0854655f4aff9d815dc3413a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8d2801845ecc0139fd2fa1c13f00c9543b862c65 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/e0a0db331f90bf249f3e076f3a85f6f14f76ef3a0854655f4aff9d815dc3413a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67482859598c318a1b4394752910d52fe8f42b40343fe9294da4ca6f865a1640 +size 15896 diff --git a/parse/train/TR-Nj6nFx42/images/e53c3b96f760803c4c0195874a9b2045e06575d941ffc91ef3426e2da002f21f.jpg b/parse/train/TR-Nj6nFx42/images/e53c3b96f760803c4c0195874a9b2045e06575d941ffc91ef3426e2da002f21f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d9676ea015556270479509eeedded525bd75c0f5 --- /dev/null +++ b/parse/train/TR-Nj6nFx42/images/e53c3b96f760803c4c0195874a9b2045e06575d941ffc91ef3426e2da002f21f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a181387e4682d2f40ccbd3a654817e1de7ff909f627120409b205981ec95763 +size 11239 diff --git a/parse/train/YUEFlzlG_0c/images/0afcf2cad957bb91d713ca1d977d180eabe894afd8197cdb69554ea594d79514.jpg b/parse/train/YUEFlzlG_0c/images/0afcf2cad957bb91d713ca1d977d180eabe894afd8197cdb69554ea594d79514.jpg new file mode 100644 index 0000000000000000000000000000000000000000..884bc18800797f532419849d2555117926daecbf --- /dev/null +++ b/parse/train/YUEFlzlG_0c/images/0afcf2cad957bb91d713ca1d977d180eabe894afd8197cdb69554ea594d79514.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9b328662a8e9ead9943d4f581f7a888de59f47ddf76810d08ec8b69a9a4b725 +size 7163 diff --git a/parse/train/YUEFlzlG_0c/images/3f716bc0fdf329c8651b34a0ee8151d109ef2249803f50f834c077a2687e7548.jpg b/parse/train/YUEFlzlG_0c/images/3f716bc0fdf329c8651b34a0ee8151d109ef2249803f50f834c077a2687e7548.jpg new file mode 100644 index 0000000000000000000000000000000000000000..aaaae60eeb8518c50d7fb99f6fc3159854baed22 --- /dev/null +++ b/parse/train/YUEFlzlG_0c/images/3f716bc0fdf329c8651b34a0ee8151d109ef2249803f50f834c077a2687e7548.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6a4c2ba525c574f42786afb12e392cbac00c1a4160e01f3063662786e640e97 +size 16385 diff --git a/parse/train/YUEFlzlG_0c/images/40186f46e1519fa4d25e8379f3069f26be6dc308e2e04c8e31c4467cc0aa5026.jpg b/parse/train/YUEFlzlG_0c/images/40186f46e1519fa4d25e8379f3069f26be6dc308e2e04c8e31c4467cc0aa5026.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ee5747c6e8cc642cc5e262df47541fe03d92e889 --- /dev/null +++ b/parse/train/YUEFlzlG_0c/images/40186f46e1519fa4d25e8379f3069f26be6dc308e2e04c8e31c4467cc0aa5026.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:464d7bda72e7272933dcbb2c6c86bc5e64ba56cfe6cbde5e41de814ccdcfaa14 +size 5938 diff --git a/parse/train/YUEFlzlG_0c/images/520ca044f1ba8cc0ca59bc40e01f2d02aa400e93eb6f6741a3ef7a14ee11a80c.jpg b/parse/train/YUEFlzlG_0c/images/520ca044f1ba8cc0ca59bc40e01f2d02aa400e93eb6f6741a3ef7a14ee11a80c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..78be8ef80d4cb61f5f26e61c03b9b4050c9e36fc --- /dev/null +++ b/parse/train/YUEFlzlG_0c/images/520ca044f1ba8cc0ca59bc40e01f2d02aa400e93eb6f6741a3ef7a14ee11a80c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0c0d6c2c1c76747ea6cde53f3147d017d5d004ae6b91a5c73f9f72102b2e0a2 +size 6143 diff --git a/parse/train/YUEFlzlG_0c/images/5b5d797713a6469acad5fb0bf6e7bc946bba80013ead1410fa91caa14bcccc0b.jpg b/parse/train/YUEFlzlG_0c/images/5b5d797713a6469acad5fb0bf6e7bc946bba80013ead1410fa91caa14bcccc0b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..545a91e8a65c4f134fab20d7fec38a6886bed84c --- /dev/null +++ b/parse/train/YUEFlzlG_0c/images/5b5d797713a6469acad5fb0bf6e7bc946bba80013ead1410fa91caa14bcccc0b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb42798046a6367b883eb44578b37403f68ff46281e8bbf0ee215fdd2722d7d4 +size 5106 diff --git a/parse/train/YUEFlzlG_0c/images/924ee084bc57d7be88147e6536f4ee59feef084ad947ea7a4fcf28f3831f4841.jpg b/parse/train/YUEFlzlG_0c/images/924ee084bc57d7be88147e6536f4ee59feef084ad947ea7a4fcf28f3831f4841.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2ad3008720593aeae9bee84fc3a94d8b4ac8e734 --- /dev/null +++ b/parse/train/YUEFlzlG_0c/images/924ee084bc57d7be88147e6536f4ee59feef084ad947ea7a4fcf28f3831f4841.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:248976f51cfdb40da74a4427a778be55ca17a44cdf0155ba6ad9f610ff907115 +size 5592 diff --git a/parse/train/YUEFlzlG_0c/images/ce99269200b8311cd63ef3c46c41140dd45c6a22a0c985b5bfc62b9be77124a6.jpg b/parse/train/YUEFlzlG_0c/images/ce99269200b8311cd63ef3c46c41140dd45c6a22a0c985b5bfc62b9be77124a6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..59d579741a85c3922c69b7195cd23bdd381d1bf9 --- /dev/null +++ b/parse/train/YUEFlzlG_0c/images/ce99269200b8311cd63ef3c46c41140dd45c6a22a0c985b5bfc62b9be77124a6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0095e650d8e3010c83db40071c835eb2f3d0559dfbe82fa00dac79979f3dfa66 +size 11639 diff --git a/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd.md b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd.md new file mode 100644 index 0000000000000000000000000000000000000000..d2c565ffab1030db0cb63aec9be6c79645797794 --- /dev/null +++ b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd.md @@ -0,0 +1,398 @@ +# Direct then Diffuse: Incremental Unsupervised Skill Discovery for State Covering and Goal Reaching + +Anonymous Author(s) +Affiliation +Address +email + +# Abstract + +1 Learning meaningful behaviors in the absence of a task-specific reward function +2 is a challenging problem in reinforcement learning. A desirable unsupervised +3 objective is to learn a set of diverse skills that provide a thorough coverage of +4 the state space while being directed, i.e., reliably reaching distinct regions of the +5 environment. At test time, an agent could then leverage these skills to solve sparse +6 reward problems by performing efficient exploration and finding an effective goal +7 directed policy with little-to-no additional learning. Unfortunately, it is challenging +8 to learn skills with such properties, as diffusing (e.g., stochastic policies performing +9 good coverage) skills are not reliable in targeting specific states, whereas directed +10 (e.g., goal-based policies) skills provide limited coverage. In this paper, inspired +11 by the mutual information framework, we propose a novel algorithm designed +12 to maximize coverage while ensuring a constraint on the directedness of each +13 skill. In particular, we design skills with a decoupled policy structure, with a first +14 part trained to be directed and a second diffusing part that ensures local coverage. +15 Furthermore, we leverage the directedness constraint to adaptively add or remove +16 skills as well as incrementally compose them along a tree that is grown to achieve a +17 thorough coverage of the environment. We illustrate how our learned skills enables +18 to efficiently solve sparse-reward downstream tasks in navigation and continuous +19 control environments, where it compares favorably with existing baselines. + +# 20 1 Introduction + +21 Deep reinforcement learning (RL) algorithms have been shown to effectively solve a wide variety of +22 complex problems [e.g., 23, 6, 31, 12, 2, 28]. However, they are often designed to solve one single +23 task at a time and they need to restart the learning process from scratch for any new problem, even +24 when it is defined on the very same environment (e.g., navigating to different locations in the same +25 apartment). Recently, unsupervised RL (URL) has been proposed as an approach to address this +26 limitation. In URL, the agent first interacts with the environment without any extrinsic reward signal. +27 Afterward, the agent leverages the experience accumulated during the unsupervised learning phase to +28 efficiently solve a variety of downstream tasks defined on the same environment. +29 In this paper, we consider the URL setting where the agent starts from an initial state $s _ { 0 }$ and it resets +30 to it every time the policy terminates. We focus on sparse-reward downstream tasks, which require +31 effective exploration (i.e., via a thorough coverage of the state space) to find the goal as well as +32 learning a policy reliably reaching the goal (i.e., a directed policy). +33 We build on the insight that mutual information (MI) effectively formalizes the dual objective of +34 learning skills that both cover and navigate the environment efficiently [e.g., 11]. Specifically, given +35 the state variable $S$ and some variables $Z$ on which the skill policies are conditioned, MI is defined as + +$$ +\begin{array} { r } { \boldsymbol { \mathcal { Z } } ( \boldsymbol { S } ; Z ) = \boldsymbol { \mathcal { H } } ( \boldsymbol { S } ) } \\ { \boldsymbol { \mathcal { Z } } ( \boldsymbol { S } ; Z ) = \boldsymbol { \mathcal { H } } ( \boldsymbol { S } ) } \end{array} - \boldsymbol { \mathcal { H } } ( \boldsymbol { S } | Z ) = \boldsymbol { \mathcal { H } } ( Z ) - \boldsymbol { \mathcal { H } } ( Z | \boldsymbol { S } ) , +$$ + +Submitted to 35th Conference on Neural Information Processing Systems (NeurIPS 2021). Do not distribute. + +![](images/f3a6a7f8b3e7aa8d82fa62e9127f2a972085e7545768b2da5305611566b631fe.jpg) +Figure 1: Overview of UPSIDE. The black dot corresponds to the initial state $s _ { 0 }$ . (A) A set of random skills is initialized, each skill being composed of a directed part (illustrated as a black arrow) and a diffusing part (red arrows), which induces a local coverage (colored circles). $( B )$ The policies associated to the directed part of each skill are then updated to maximize the discriminability of the states reached by their diffusing part (Sect. 3.1). $( C )$ The least discriminable skills are iteratively removed while the policies of the remaining skills are re-optimized. This is executed until the discriminability of each skill satisfies a given constraint (see Sect. 3.2). In this example three skills are kept. $( D )$ One of these learned skill is then used as basis to add new skills, which are then optimized following the same procedure. For the “red” and “purple” skills, UPSIDE is not able to find sub-skills of sufficient quality and thus they are not expanded any further. $( E )$ At the end of the process, UPSIDE has created a tree of directed skills covering the state space (Sect. 3.3). These covering skills can then be used to solve downstream tasks. Moreover, the discriminator learned together with the skills can be used to select the skill to reach any specific goal region, where the directed parts get close to the goal, while the diffusing part provides the local coverage to attain the goal. The complete algorithm is detailed in Sect. 3.4 and Appendix. + +36 where $\mathcal { T }$ denotes the MI and $\mathcal { H }$ is the entropy function. The first expression, known as the forward +37 form of MI, explicitly balances the two sought-after properties of coverage — captured by the entropy +38 over the state space $\mathcal { H } ( S )$ — and directedness, i.e., the ability to reach specific states $S$ depending +39 on $Z$ — captured by the negative conditional entropy $- { \mathcal { H } } ( S | Z )$ . The second expression of (1), often +40 easier to optimize and referred to as the reverse form, stipulates that the skills should be sampled as +41 diversely as possible while being discriminable. +42 Maximizing (1) has been shown to be a powerful approach for encouraging exploration in RL [16, 25] +43 and for unsupervised skill discovery [e.g., 11, 9, 1, 30, 8]. Nonetheless, learning skills that maximize +44 the MI is a challenging optimization problem. Several approximations have been proposed to simplify +45 the problem at the cost of possibly deviating from the original objective of coverage and directedness +46 (see Sect. 4 for a review of related work). In this paper, we propose UPSIDE (UnsuPervised Skills +47 that dIrect then DiffusE) to learn skills that can be effectively used to solve goal-based downstream +48 tasks. Our solution builds on the following components (see Fig. 1 for an illustration of UPSIDE): + +• Skill structure. In order to balance coverage and directedness, we design skills composed of two parts: 1) a directed part that is trained to reach a distinct region of the environment, and 2) a diffusing part that covers the states around the region attained by the first part. Optimization. We further strengthen the coverage and directedness properties of the skills by turning the MI objective into a constrained optimization problem designed to maximize coverage under the constraint that each skill achieves a minimum level of discriminability. This in turn enables UPSIDE to adaptively add skills to improve coverage, when all the initial skills meet the constraint, or remove those that violate the constraint to guarantee that each skill is directed and reaches a distinct region of the environment. • Tree structure. When the agent starts from a fixed initial state, the skills’ length is a crucial parameter, where short skills do not allow for proper coverage, and long skills are difficult to train. In UPSIDE we consider short skills to make the optimization easier, while composing them along a tree structure that ensures an adaptive and deep coverage of the environment. + +62 We study how our learned skill structure enables to both perform efficient exploration and learn +63 effective goal-reaching policies in a variety of navigation and continuous control environments +64 (including MuJoCo’s reacher) and we compare its performance to relevant baselines. + +# 2 Setting + +66 We consider the URL setting where the agent interacts with a Markov decision process (MDP) $M$ with state space 67 $s$ , action space $\mathcal { A }$ , dynamics $p ( s ^ { \prime } | s , a )$ , and no reward. The agent starts each + +68 episode from a designated initial state $s _ { 0 } \in S$ . Upon termination of the chosen policy, the agent is +69 then reset to $s _ { 0 }$ . This setting is particularly challenging from an exploration point of view since the +70 agent cannot rely on the initial distribution to cover the state space. +71 We recall the MI-based unsupervised skill discovery approach [see e.g., 11]. Denote by $Z$ some +72 (latent) variables on which the skills of length $T$ are conditioned. There are three optimization +73 variables: $( i )$ the support of the skills denoted by $| Z |$ (we consider it to be discrete so $| Z |$ is the +74 number of skills), $( i i )$ the policy $\pi ( z )$ associated to skill $z$ , and $( i i i )$ the sampling rule $\rho$ (i.e., $\rho ( z )$ +75 is the probability of sampling skill $z$ at the beginning of the episode). Let the variable $S _ { T }$ be the +76 random (final) state induced by sampling a skill $z$ from $\rho$ and executing the associated policy $\pi ( z )$ +77 from $s _ { 0 }$ for an episode. We denote by $p _ { \pi ( z ) } ( s _ { T } )$ the distribution over (final) states induced by +78 executing the policy of skill $z$ , by $p ( z | s _ { T } )$ the probability of $z$ being the skill to induce state $s _ { T }$ , and +79 let $\begin{array} { r } { \bar { p } ( s _ { T } ) \bar { ( s _ { T } ) } = \bar { \sum _ { z \in Z } \bar { \rho ( z ) } } p _ { \pi ( z ) } ( s _ { T } ) } \end{array}$ . Then maximizing the MI between $Z$ and $S _ { T }$ can be written as + +$$ +\begin{array} { r l r } & { } & { \displaystyle \underset { | Z | , \rho , \pi } { \operatorname* { m a x } } \ : \mathcal { Z } ( S _ { T } ; Z ) = \mathcal { H } ( S _ { T } ) - \mathcal { H } ( S _ { T } | Z ) = - { \displaystyle \sum _ { s _ { T } } } \overline { { p } } ( s _ { T } ) \log \overline { { p } } ( s _ { T } ) + { \displaystyle \sum _ { z \in Z } } \rho ( z ) \mathbb { E } _ { s _ { T } } \left[ \log p _ { \pi ( z ) } ( s _ { T } ) \right] } \\ & { } & { = \mathcal { H } ( Z ) - \mathcal { H } ( Z | S _ { T } ) = - { \displaystyle \sum _ { z \in Z } } \rho ( z ) \log \rho ( z ) + { \displaystyle \sum _ { z \in Z } } \rho ( z ) \mathbb { E } _ { s _ { T } } \left[ \log p ( z | s _ { T } ) \right] , \quad ( 1 ) } \end{array} +$$ + +80 where in the expectations $s _ { T } \sim p _ { \pi ( z ) } ( s _ { T } )$ . As discussed in Sect. 1, learning the optimal $| Z | , \rho ,$ and $\pi$ +81 is a challenging problem [see e.g., 11, 9, 8]. + +# 3 Algorithm Structure + +UPSIDE is based on three main components: a) the skill learning corresponding to stage $A$ and $B$ of Fig. 1 and described in Sect. 3.1, b) a constrained optimization problem used to optimize the number of skills (stage $C$ and Sect. 3.2) and c) a tree-building procedure (stage $D$ and Sect. 3.3). Together, these components allow UPSIDE to discover skills that combine coverage and directedness. + +# 3.1 Skill Structure and Optimization + +As shown in e.g., [9, 30, 37], the level of stochasticity of each skill (e.g., induced via a regularization on the entropy over the actions) plays a key role in trading off coverage and directedness. In fact, while randomness promotes broader coverage, it may compromise the directedness of the skills. In fact, a highly stochastic skill tends to induce a distribution $p _ { \pi ( z ) } ( s _ { T } )$ over final states with high entropy (thus decreasing $- \mathcal { H } ( S _ { T } | Z ) )$ , which prevents the skill to be reusable in solving sparse-reward downstream tasks where the objective is to reliably reach a specific goal state of the environment. Determining how much stochasticity to inject to adequately balance both objectives and optimize (1) is a difficult problem.1 + +96 We propose to design skills with a decoupled policy structure: + +• A directed part (of length $T$ ) with low stochasticity and trained to reach a specific region of the environment. It is responsible for increasing the $- { \mathcal { H } } ( S | Z )$ term in (1). • A diffusing part (of length $H$ ) with high stochasticity to promote local coverage of the states around the region reached by the directed part. It is responsible for increasing the $\mathcal { H } ( S )$ term in (1). + +![](images/cd135dcbe8681b37eac698ab1aa41d695eeae5d2288d124ad1c5924bd7a5fec1.jpg) +Figure 2: Directed and diffusing parts of the skill. + +98 Similar to prior work [e.g., 11, 9], the policy associated to the directed part of skill $z$ is trained to max +99 imize an intrinsic reward $r _ { z } ( s ) \approx p ( z | s )$ ,2 where $p ( z | s )$ measures the “discriminability” of the skill $z$ +100 given the state $s$ . More formally, $\pi ( z )$ maximizes the cumulative reward $\begin{array} { r } { \mathbb { E } _ { \pi ( z ) } \left[ \sum _ { t = T + 1 } ^ { T + H } r _ { z } ( s _ { t } ) \right] } \end{array}$ +101 over the states traversed by the policy during the diffusing part. In practice, we also add a small +102 entropy regularization $\mathcal { H } ( \pi ( \cdot | z , s _ { t } ) )$ to the directed policy in order to ensure a minimum level of +103 exploration and make the learning more robust. For the diffusing part, we rely on a simple random +104 walk policy (i.e., a stochastic policy with uniform distribution over actions). +105 Intuitively, the diffusing part defines a cluster of states that is used as a goal for the directed part. +106 This allows us to “ground” the latent variable representations of the skills $Z$ to specific regions of +107 the environment (i.e., the clusters). As a result, maximizing the MI over such skills can be seen as +108 learning a set of “cluster-conditioned”, and thus directed, policies. + +# 3.2 Skill Support and Sampling Rule + +110 The MI objective (1) crucially depends on the number of skills $( | Z | )$ and the distribution $\rho ( z )$ +111 Unfortunately, it is been shown [e.g., 8] that solving (1) is particularly challenging. In order to +112 simplify the optimization and the associated learning problem, we modify (1) in two ways. +113 First, coherently with the skill optimization detailed in Sect. 3.1, the random variable $S$ in the +114 conditional entropy is any state reached during the diffusing part of the skill and not just the terminal +115 state. More formally, we denote by $S _ { \mathrm { d i f f } }$ the random variable and its distribution for a specific skill $z$ +116 is $p _ { \pi ( z ) } ( s _ { \mathrm { d i f f } } ) = 1 / \dot { H } \sum _ { t = T + 1 } ^ { T + H } p _ { \pi ( z ) } ( s _ { t } )$ , i.e., the distribution over states obtained by averaging the +117 distributions at any of the steps in the diffusing part. Similarly, $p ( z | s _ { \mathrm { d i f f } } )$ now denotes the probability +118 of $z$ being the skill to traverse $s _ { \mathrm { d i f f } }$ during its diffusing part. As a result, training the skills to maximize +119 MI naturally leads the diffusing parts to “push” the directed parts away so as to reach diverse regions +120 of the environment. The combination of “global” coverage of the directed parts and “local” coverage +121 of the diffusing part ensures that the whole environment is properly visited with $| Z | \ll S$ skills.3 +122 Second, we introduce an alternative problem that simplifies the optimization while preserving the +123 coverage and directedness properties of MI. This is achieved by introducing a stronger requirement +124 on the discriminability. While the conditional entropy term $- { \mathcal { H } } ( Z | S )$ in (1) promotes the discrim +125 inability of skills on average, we argue that a more suitable objective is to constrain each skill to +126 achieve a minimum level of discriminability. First, we move from the average to the minimum over +127 skills by lower bounding the conditional entropy as + +$$ +- \mathcal { H } ( Z | S _ { \mathrm { d i f f } } ) = \sum _ { z \in Z } \rho ( z ) \mathbb { E } _ { s _ { \mathrm { d i f f } } } \left[ \log p ( z | s _ { \mathrm { d i f f } } ) \right] \geq \operatorname* { m i n } _ { z \in Z } \mathbb { E } _ { s _ { \mathrm { d i f f } } } \left[ \log p ( z | s _ { \mathrm { d i f f } } ) \right] , +$$ + +128 which leads to the following optimization (assuming $\pi$ is fixed for convenience) + +$$ +\operatorname* { m a x } _ { | Z | = N , \rho } \bigg \{ \mathcal { H } ( Z ) + \operatorname* { m i n } _ { z \in [ N ] } \mathbb { E } _ { s _ { \mathrm { d i f f } } } \left[ \log p ( z | s _ { \mathrm { d i f f } } ) \right] \bigg \} , +$$ + +129 where with an abuse of notation we use $z \in [ N ]$ to denote all skills in a set $Z$ with cardinality $N$ +130 Since (3) is a lower bound to MI, it tends to promote the same type of covering and directed skills. +131 Furthermore, (2) no longer depends on the distribution over skills and the entropy term $\mathcal { H } ( Z )$ is +132 maximized by setting $\rho$ to the uniform distribution over $N$ skills (i.e., $\begin{array} { r } { \operatorname* { m a x } _ { \rho } \mathcal { H } ( Z ) \bar { = } \log ( N ) ) } \end{array}$ , thus +133 simplifying the optimization, which now only depends on $N$ . +134 While optimizing (3) promotes a cardinality $N$ such that all skills have good discriminability, a more +135 convenient formulation is to explicitly set a minimum level of discriminability for all skills through +136 the following constrained optimization problem: + +$$ +\operatorname* { m a x } _ { N \geq 1 } \log ( N ) \qquad \mathrm { s . t . } \qquad \operatorname* { m i n } _ { z \in [ N ] } \mathbb { E } _ { s _ { \mathrm { d i f f } } } \left[ \log p ( z | s _ { \mathrm { d i f f } } ) \right] \geq \log \eta . +$$ + +137 where $\eta$ is a parameter that defines the discriminability threshold. A skill $z$ is said to be $\eta$ -consolidated +138 if it satisfies the constraint. Crucially, let $\begin{array} { r } { P _ { N } : = \operatorname* { m i n } _ { z \in [ N ] } \mathbb { E } _ { s _ { \mathrm { d i f f } } } \left[ \log p ( z | s _ { \mathrm { d i f f } } ) \right] } \end{array}$ , then the sequence +139 $( P _ { N } ) _ { N \ge 1 }$ is non-increasing with $P _ { 1 } = 0$ (i.e., the more skills the harder it is to meet the constraint). +140 As a result, (4) can be optimized following a simple greedy strategy incrementally adding skills until +141 the constraint is violated. The optimal $N$ thus defines the effective number of $\eta$ -consolidated skills and +142 it corresponds to the largest number of skills that is guaranteed to display sufficient discriminability. +143 Alternatively, we can interpret (4) as finding the largest number of clusters (i.e., the region reached +144 by the directed part of a skill and covered by its associated diffusing part) with a minimum level of +145 inter-cluster distance. This effect is qualitatively illustrated in Fig. 1, where the states attained by the +146 directed part of the skills attain different regions that are locally covered by their diffusing parts. + +# Algorithm 1: UPSIDE + +Initialize: Discriminability threshold $\eta \in ( 0 , 1 )$ , branching factor $N _ { 0 } \geq 1$ , patience $K$ +Initialize: Tree $\tau$ initialized as a root node indexed by 0, queue of parent nodes $\mathcal { W } = \{ 0 \}$ . +while $\mathcal { W } \neq \emptyset$ do // tree expansion +Dequeue a node/skill $w \in \mathcal W$ and expand $\tau$ at $w$ by adding a set $\mathcal { C } ( w )$ of $N _ { 0 }$ nodes/skills +2 Create random policies $\pi _ { z }$ , $\forall z \in \mathcal { C } ( \bar { \boldsymbol { w } } )$ +3 Initialize discriminator $q _ { \phi }$ with $| \tau |$ classes +4 Continue $=$ true; Saturated $=$ false +5 while Continue do +6 for $K$ iterations do +7 Sample a skill $z$ from $\tau$ at random +8 Extract the sequence of nodes $z _ { ( 1 ) } , \dotsc , z$ in $\tau$ leading to $z$ +9 Execute the composed (directed part) policy $( \pi _ { z _ { ( 1 ) } } , \ldots , \pi _ { z } )$ followed by the diffusing part +10 Add states observed during the diffusion part to state buffer $B _ { z }$ +11 Update discriminator $q _ { \phi }$ with SGD on $B _ { z }$ to predict label $z$ +12 if $z \in \mathcal { C } ( w )$ then $/ /$ Update only new policies, other polices kept fixed +13 Update policy $\pi _ { z }$ using SAC to optimize the discriminator reward as in Sect. 3.1. +14 Compute the skill-discriminability $\begin{array} { r } { d ( z ) = \widehat { q } _ { \phi } ^ { \scriptscriptstyle ( B ) } ( z ) = \frac { 1 } { | \mathscr { B } _ { z } | } \sum _ { s \in \mathscr { B } _ { z } } q _ { \phi } ( z | s ) } \end{array}$ for all $z \in \mathcal { C } ( w )$ +15 if $\begin{array} { r } { \operatorname* { m i n } _ { z \in \mathcal { C } ( w ) } d ( z ) < \eta } \end{array}$ then $/ /$ Node removal +16 Remove the node/skill $z = \arg \operatorname* { m i n } _ { z \in { \mathcal { C } } ( w ) } d ( z )$ from $\mathcal { C } ( w )$ and $\tau$ +17 Set Saturate $=$ true +18 else if not Saturated then +19 Add one new node/skill to $\mathcal { C } ( w )$ and $\tau$ +20 else +21 Set Continue $=$ false +22 Enqueue in $\mathcal { W }$ the consolidated nodes $\mathcal { C } ( w )$ + +# 147 3.3 Composing Skills in a Tree Structure + +The MI optimization problem as well as our constrained variant (4) depend on the initial state $s _ { 0 }$ and on the length of each skill. Although these quantities are usually predefined and only appear implicitly in the equations, they have a crucial impact on the obtained behavior. In fact, resetting after each skill execution unavoidably restricts the coverage to a radius of at most $T + H$ steps around $s _ { 0 }$ This may suggest to set $T$ and $H$ to a large value. However, increasing the horizon makes the training of the skills more challenging, as learning $\pi$ would require solving a difficult RL problem itself. + +Instead, we propose to “extend” the length of the skills through composition. Indeed, the decoupled skill structure and the constraint in (4) entail that the directed part of each of the $\eta$ -consolidated skills reliably reach a specific (and distinct) region of the environment and it is thus re-usable and amenable to composition. We propose to chain the directed part of the skills in order to reach further and further parts of the state space. Specifically, we build a growing tree, where the root is the initial state $s _ { 0 }$ , the edges represent the directed part of the skills, and the nodes represent the diffusing part of skills. As such, whenever a skill $z$ is selected, the directed part of all the policies associated to its predecessor skills in the tree are executed first (see Fig. 1 for an illustration of the tree structure). + +As a result, the agent naturally builds a curriculum on the episode lengths, which grow as the sequence $( i T + H ) _ { i \geq 1 }$ . As such, it does not require prior knowledge on an adequate horizon of the downstream goal-based task.4 Here this knowledge is replaced by $T$ and $H$ which are more environment-agnostic and task-agnostic quantities, as their choice rather has an impact on the size and shape of the learned tree (e.g., the smaller $T$ and $H$ the bigger the tree). + +# 3.4 The UPSIDE Algorithm + +We are now ready to introduce UPSIDE, which provides a specific implementation of the components described before (see Fig. 1 for a qualitative illustration and Algorithm 1 for the detailed pseudo-code). + +0 We perform standard approximations to make the constraint in (4) easier to estimate. We approximate the unknown posterior $p ( z | s )$ with a learned discriminator $q _ { \phi } ( z | s )$ with parameters $\phi$ . We also + +172 remove the logarithm from the constraint to have an estimation range of $[ 0 , 1 ]$ and thus lower +173 variance2. Finally, we replace the expectation over $s$ with an empirical estimate $\widehat { q } _ { \phi } ^ { ( B ) } ( z )$ averaging the +174 value of the discriminator evaluated on the last $B$ states observed while executing the diffusing part +175 of $z$ . Integrating these approximations in (4) leads to + +$$ +\operatorname* { m a x } _ { N \geq 1 , \pi } N \qquad \mathrm { s . t . } \qquad \operatorname* { m i n } _ { z \in [ N ] } \widehat { q } _ { \phi } ^ { ( B ) } ( z ) \geq \eta . +$$ + +176 As discussed in Sect. 3.2, this problem can be conveniently optimized using a greedy strategy. We +177 then integrate the optimization of (5) into an adaptive tree expansion strategy: (Generating new +178 skills) Given a tree structure as described in Sect. 3.3, we expand the tree at a leaf $w$ by adding $N _ { 0 }$ +179 new nodes/skills following a breadth-first-search approach (lines 1, 2). Then (Skill Learning) the +180 new skills are optimized by: i) sampling random skills in the tree to update the discriminator (lines +181 7-11), and ii) by updating the policies to optimize the discriminability reward (Sect. 3.1) computed +182 using the discriminator (lines 13). To speed-up convergence, we only update the policies that have be +183 added to the tree structure, keeping all the previous policies fixed (line 12). Note that in the update of +184 the discriminator we leverage the states observed in previous phases of the algorithm by maintaining +185 a (small) replay buffer of states for each skill. (Node Consolidation) After a patience period (line 6), +186 if all skills are $\eta$ -consolidated, we tentatively add more skills to the leaf $w$ (line 18). On the other +187 hand, if any skill does not meet the discriminability threshold, we remove it and consolidate the +188 remaining skills into the tree (lines 16, 17) and we repeat the process. +189 Model selection. A core aspect of any RL algorithm is model selection, i.e., finding the best +190 configuration of hyperparameters. In URL with no prior knowledge of the downstream task(s), it +191 is non-trivial to devise an adequate criterion for model selection and this aspect is rarely addressed, +192 despite being crucial in practice. For instance, while the coverage of the state space may be a +193 good proxy for the performance of a URL algorithm [see e.g., 8], it may be difficult to measure in +194 continuous problems. Interestingly, our optimization problem directly provides a single, task-agnostic +195 and environment-agnostic criterion for model selection, which is the number $N$ of $\eta$ -consolidated +196 skills discovered by the agent. Indeed in all of our experiments we simply select the model (i.e., set +197 of hyperparameters) that maximizes $N$ . This is a significant advantage w.r.t. existing methods, such +198 as VIC and DIAYN, for which no principled approach to model selection is provided. + +# 199 4 Related work + +Unsupervised Reinforcement Learning methods can be broadly decomposed according to the way they summarize the experience accumulated during the unsupervised phase into reusable knowledge to solve downstream tasks. This includes both off-policy model-free [e.g., 27] and model-based [e.g., 29] methods that seek to populate a representative replay buffer and build accurate value or model estimates, that are used to solve a given downstream task in a zero- or few-shot manner. The accumulated experience during train time can also be compressed into a low-dimensional representation for value functions as well as policies and to improve exploration [e.g., 36]. An alternative line of work focuses on the discovery of a set of skills in an unsupervised manner. Our approach falls in this category, on which we now focus our related work review. + +209 Skill discovery based on MI maximization was first proposed in VIC [11], where only the final +210 states of each trajectory are considered in the reverse form of (1) and where both the skills and +211 their sampling rules are simultaneously learned (with a fixed support $| Z |$ , i.e., a fixed number of +212 skills). DIAYN [9] fixes the sampling rule to be uniform, and weighs the skills with an action-entropy +213 coefficient (i.e., it additionally minimizes the MI between actions and skills given the state), so as +214 to push the skills away from each other and enhance coverage. DADS [30] learns skills that are not +215 only diverse but also predictable by learned dynamics models, by using a generative model over +216 observations (rather than over skills) and optimizing a forward form of MI, namely $\mathcal { T } ( s ^ { \prime } ; z | s )$ between +217 the next state $s ^ { \prime }$ and current skill $z$ (with continuous latent) conditioned on the current state $s$ . EDL [8] +218 shows that existing skill discovery approaches can provide insufficient coverage, and instead proposes +219 to rely on a fixed distribution over states $p ( s )$ which is either provided by an oracle or learned. In +220 SMM [19], the MI formalism is used to learn a policy for which the state marginal distribution matches +221 a given target state distribution (e.g., uniform), which can be seen as a more scalable way of tackling +222 the problem of maximum entropy over the state space [15], and as a way to encourage skills to go +223 through unknown state regions. Other MI-based skill discovery methods include [10, 14, 24, 5, 34], +224 as well as [35, 20] which investigate skill discovery in non-episodic settings. +225 Our approach shares a similar motivation to prior MI-based works of targeting skills that are both +226 directed and state-covering. In particular, the decoupled structure introduced in Sect. 3.1 can be seen +227 as a more suitable way to achieve the objective of improving the coverage of VIC as done in DIAYN +228 and SMM, without compromising the directedness of the skills. +229 While most skill discovery approaches consider a fixed number of skills, a curriculum with increasing +230 number of skills is studied in [1, 3]. Our discriminability constraint is what enables skills to be +231 composed along a tree structure, which allows increases or decreases the support of available skills +232 depending on the region of the state space. + +![](images/e9c4d365719fb312ceafdef45970c4f57a1b44e9db7d7229c05847b9c4b09596.jpg) +Figure 3: UPSIDE, DIAYN-curriculum and SMM-10 skills learned in a bottleneck maze (Top) and a U-maze (Bottom). For both DIAYN and SMM we report the stochastic execution of the learned skills and for UPSIDE we report the deterministic directed parts (that are composed) followed by the (stochastic) diffusing part, which is the same protocol used to evaluate coverage. + +Recently, [37] proposed a hierarchical RL method that discovers abstract and task-agnostic skills while jointly learning a higher-level policy which is trained to maximize environment reward. Our approach builds on a similar promise of composing skills instead of resetting to $s _ { 0 }$ after each execution, yet we articulate the composition differently, by exploiting the direct-then-diffuse structure to ground learned skills to the state space instead of being abstract. + +238 In addition, approaches such as DISCERN [33] and Skew-Fit [27] learn a goal-conditioned policy in +239 an unsupervised way with an MI objective. As explained in [8, Sect. 5], this can be interpreted as a +240 skill discovery approach with latent $Z = S$ , i.e., where each goal state can define a different skill. +241 Conditioning on either goal states or abstract latent skills forms two extremes of the spectrum of +242 unsupervised RL. We target an intermediate approach, seeking to benefit from the groundedness of +243 the latent skill $Z$ and the states $S$ (and thus amenability to composition) of goal-conditioned RL, and +244 from the reduced search space and sampling ease of skill-based RL. + +An alternative approach to skill discovery builds on “spectral” properties of the dynamics of the environment. This includes eigenoptions [21, 22] and covering options [17, 18], as well as the algorithm of [4] that builds a discrete graph representation which learns and composes spectral skills. + +# 5 Experiments + +In this section, we investigate the following questions: i) Can the adaptive tree structure of UPSIDE incrementally cover an unknown environment while preserving directedness of the skills? ii) Following the unsupervised phase, how can UPSIDE be leveraged to solve goal-based downstream tasks? + +We report results on: a) Navigation problems in continuous mazes, where actions represent the desired shift in $x$ and $y$ coordinates; b) A difficult instance of CartPole, where the cart starts with zero speed and the pole is oriented downside; c) The Reacher [32] problem using the MuJoCo implementation in Gym [7]. In all environments, the per-dimension action space is in $[ - 1 ; + 1 ]$ . + +![](images/f67a926a679d5569359c50d0e257b1aa3baa0708a7b5b489b2d719b79dee97b3.jpg) +Figure 4: Normalized coverage in U-maze and bottleneck. + +256 We compare to different baselines. DIAYN-K, where $K$ is a fixed number of skills, is the original +257 algorithm proposed in [9]. DIAYN-Curriculum is a variant where the number of skills is automatically +258 tuned following the same procedure as in UPSIDE ensuring a good discriminability. We also compare +259 to SMM [19], which is similar to DIAYN, but it includes an exploration bonus encouraging the policies +260 to visit rarely encountered states. In our implementation, the exploration bonus is obtained by +261 maintaining a multinomial distribution over “buckets of states” obtained by discretization, resulting +262 in an computation-efficient and stable implementation that is more stable than the original VAE-based +263 method. UPSIDE and all baselines are implemented with Soft-Actor Critic (SAC) [13]. +264 Unsupervised Phase. We run all methods until convergence. We then do model selection according +265 to the criterion of either the final number of skills for UPSIDE and DIAYN-curriculum and the final +266 average discriminability for DIAYN-K and SMM. To compute the coverage, we perform rollouts by +267 first sampling a skill uniformly at random and executing its associated policy until termination. We +268 discretize states into buckets (50 interval per dimension for mazes and 10 for control environments) +269 and report the proportion of buckets reached by each method as a function of the total number of +270 steps executed in the environment over multiple rollouts. Since only a small portion of the discretized +271 states can be reached, we normalize the coverage such that the best method obtains 1. +272 We consider two topologies of mazes with size (height and width) 50 such that exploration is non +273 trivial (i.e., a random policy is only able to cover a small part of the state space): a U-shaped maze +274 and a Bottleneck maze (which is a harder version of the one in [8, Fig. 1] which is only of size 10 +275 for the same action space). In Fig. 3 we show that UPSIDE succeeds in covering the near-entirety +276 of the state space by creating a tree of directed skills. Moreover, UPSIDE created directed skills +277 with a low entropy, while the two baselines tend to create skills that are more stochastic. This is +278 particularly evident for SMM, due to the state-entropy exploration bonus, that while it encourages +279 broader coverage makes skills less directed. + +In Fig. 4 we report the coverage on the Bottleneck maze and U-Maze. For UPSIDE, executing a skill corresponds to executing the directed part of all the “parent” skills in the tree and concluding with the diffusion part of the skill. SMM achieves better coverage than DIAYN thanks to the increased level of stochasticity (diffusion) of its skills. UPSIDE outperforms both by reaching regions of the environment that are not be achieved by other methods. Here, we plot UPSIDE with $T = 1 0$ and $H = 1 0$ , but we found UPSIDE to be robust to these parameters as shown in the supplementary. + +Results are similar in the CartPole problem (see Fig. 5) where UPSIDE (with $T = 2 0$ and $H = 2 0$ ) obtains better coverage than baselines. On the other hand, in Reacher (see Fig. 5), DIAYN-50 outperforms UPSIDE in terms of coverage. This can be explained by the fact that, in this environment, highly stochastic skills provide a good coverage. Nonetheless, this comes at the cost of very low discriminability (rightmost plot), which suggests DIAYN-50 skills have poor directedness. On the other hand, UPSIDE (and DIAYN-curriculum) achieves much larger discriminability by removing redundant skills and favoring more directed policies. + +Downstream Tasks. Following the unsupervised phase, UPSIDE has learned a tree of skills. We now investigate how these skills are used to tackle a downstream task. In that setting, we propose to use skill-based approaches (i.e UPSIDE, DIAYN and SMM) in the following way: a) (exploration) first we sample rollouts over the different skills. b) We then select the best skill based on the maximum cumulative reward collected and c) we fine-tune this skill to maximize the reward. We report results on mazes (additional results are provided in the supplementary). We consider a sparse positive reward + +![](images/c0fef9fb524cc8a68d444c1ca195519b14d66a4947d40fb5df6c83c782ac91ca.jpg) +Figure 5: Normalized coverage in Cartpole (Left) and Reacher (Middle). (Right) Average discriminability of the skills during training in Reacher. + +299 when reaching a particular defined goal.5 We consider goals at different distances from the initial +300 state $s _ { 0 }$ , the further, the harder. Fig. 6 shows the learning curves obtained when fine-tuning the best +301 skill for the different models and compare to a classical SAC algorithm where a single policy is +302 learned from scratch. DIAYN/SMM means we use the best state-covering policies between DIAYN and +303 SMM. For the “close” goal setting, both UPSIDE and DIAYN/SMM are able to learn to reach this goal +304 efficiently while SAC solves the task only for some of the training runs. Note that we do not show +305 DIAYN performance since it is lower than the SMM one. For the “far” goal setting, only UPSIDE learns +306 to reach this goal. Obtained trajectories are illustrated in Fig. 6. + +![](images/5c8600e6316f678077840de51c561851a9e26dcd890e6d78ce9628180fe6febb.jpg) +(Left): Learning curves for “short” distance and (Middle) Figure 6: “medium” distance goals. (Right): Learned policies after fine-tuning (Top) U-maze. (Bottom): Bottleneck maze. + +# 6 Conclusion + +We introduced UPSIDE, a novel algorithm for unsupervised skill discovery designed to trade off between coverage and directedness and develop a tree of skills that can be used to both perform efficient exploration of the environment and learn effective goal-directed policies. Natural venues for future investigation are: 1) The diffusing part of each skill could be explicitly trained to maximize local coverage; 2) UPSIDE assumes a good representation of the state is provided as input, it would be interesting to pair UPSIDE with effective representation learning techniques to tackle problems with high-dimensional input (e.g., image-based RL); 3) While UPSIDE is grounded on the solid principle of MI maximization, a more thorough theoretical investigation is needed to explicitly link the optimization problem and its approximations to the downstream performance. + +References [1] J. Achiam, H. Edwards, D. Amodei, and P. Abbeel. Variational option discovery algorithms. arXiv preprint arXiv:1807.10299, 2018. [2] M. Andrychowicz, D. Crow, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin, P. Abbeel, and W. Zaremba. Hindsight experience replay. In NIPS, 2017. [3] A. Aubret, L. Matignon, and S. Hassas. Elsim: End-to-end learning of reusable skills through intrinsic motivation. arXiv preprint arXiv:2006.12903, 2020. [4] A. Bagaria, J. Crowley, J. W. N. Lim, and G. Konidaris. Skill discovery for exploration and planning using deep skill graphs. 2021. [5] K. Baumli, D. Warde-Farley, S. Hansen, and V. Mnih. Relative variational intrinsic control. arXiv preprint arXiv:2012.07827, 2020. [6] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279, 2013. [7] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba. Openai gym, 2016. [8] V. Campos, A. Trott, C. Xiong, R. Socher, X. Giro-i Nieto, and J. Torres. Explore, discover and learn: Unsupervised discovery of state-covering skills. In International Conference on Machine Learning, 2020. [9] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills without a reward function. In International Conference on Learning Representations, 2019. [10] C. Florensa, Y. Duan, and P. Abbeel. Stochastic neural networks for hierarchical reinforcement learning. arXiv preprint arXiv:1704.03012, 2017. [11] K. Gregor, D. J. Rezende, and D. Wierstra. Variational intrinsic control. arXiv preprint arXiv:1611.07507, 2016. [12] S. Gu, E. Holly, T. Lillicrap, and S. Levine. Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In 2017 IEEE international conference on robotics and automation (ICRA), pages 3389–3396. IEEE, 2017. +45 [13] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. CoRR, abs/1801.01290, 2018. +47 [14] S. Hansen, W. Dabney, A. Barreto, D. Warde-Farley, T. Van de Wiele, and V. Mnih. Fast task inference with variational intrinsic successor features. In International Conference on Learning Representations, 2019. +50 [15] E. Hazan, S. Kakade, K. Singh, and A. Van Soest. Provably efficient maximum entropy exploration. In International Conference on Machine Learning, pages 2681–2691, 2019. [16] R. Houthooft, X. Chen, Y. Duan, J. Schulman, F. De Turck, and P. Abbeel. Vime: variational information maximizing exploration. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 1117–1125, 2016. [17] Y. Jinnai, J. W. Park, D. Abel, and G. Konidaris. Discovering options for exploration by minimizing cover time. In International Conference on Machine Learning, pages 3130–3139. PMLR, 2019. [18] Y. Jinnai, J. W. Park, M. C. Machado, and G. Konidaris. Exploration in reinforcement learning with deep covering options. In International Conference on Learning Representations, 2020. [19] L. Lee, B. Eysenbach, E. Parisotto, E. Xing, S. Levine, and R. Salakhutdinov. Efficient exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019. + +[20] K. Lu, A. Grover, P. Abbeel, and I. Mordatch. Reset-free lifelong learning with skill-space planning. arXiv preprint arXiv:2012.03548, 2020. [21] M. C. Machado, M. G. Bellemare, and M. Bowling. A laplacian framework for option discovery in reinforcement learning. In International Conference on Machine Learning, pages 2295–2304. PMLR, 2017. [22] M. C. Machado, C. Rosenbaum, X. Guo, M. Liu, G. Tesauro, and M. Campbell. Eigenoption discovery through the deep successor representation. In International Conference on Learning Representations, 2018. [23] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529–533, 2015. [24] N. Modhe, P. Chattopadhyay, M. Sharma, A. Das, D. Parikh, D. Batra, and R. Vedantam. Ir-vic: Unsupervised discovery of sub-goals for transfer in rl. In Proceedings of the TwentyNinth International Joint Conference on Artificial Intelligence, IJCAI-20. International Joint Conferences on Artificial Intelligence Organization, 2020. [25] S. Mohamed and D. J. Rezende. Variational information maximisation for intrinsically motivated reinforcement learning. In Advances in neural information processing systems, pages 2125– 2133, 2015. [26] M. Mutti, L. Pratissoli, and M. Restelli. A policy gradient method for task-agnostic exploration. arXiv preprint arXiv:2007.04640, 2020. [27] V. H. Pong, M. Dalal, S. Lin, A. Nair, S. Bahl, and S. Levine. Skew-fit: State-covering self-supervised reinforcement learning. In International Conference on Machine Learning, 2020. [28] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. [29] R. Sekar, O. Rybkin, K. Daniilidis, P. Abbeel, D. Hafner, and D. Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, pages 8583–8592. PMLR, 2020. [30] A. Sharma, S. Gu, S. Levine, V. Kumar, and K. Hausman. Dynamics-aware unsupervised discovery of skills. In International Conference on Learning Representations, 2020. [31] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, et al. Mastering the game of go without human knowledge. nature, 550(7676):354–359, 2017. [32] E. Todorov, T. Erez, and Y. Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 50pou26–5033, 2012. [33] D. Warde-Farley, T. Van de Wiele, T. Kulkarni, C. Ionescu, S. Hansen, and V. Mnih. Unsupervised control through non-parametric discriminative rewards. In International Conference on Learning Representations, 2019. [34] K. Xie, H. Bharadhwaj, D. Hafner, A. Garg, and F. Shkurti. Skill transfer via partially amortized hierarchical planning. In International Conference on Learning Representations, 2021. [35] K. Xu, S. Verma, C. Finn, and S. Levine. Continual learning of control primitives: Skill discovery via reset-games. arXiv preprint arXiv:2011.05286, 2020. [36] D. Yarats, R. Fergus, A. Lazaric, and L. Pinto. Reinforcement learning with prototypical representations. arXiv preprint arXiv:2102.11271, 2021. [37] J. Zhang, H. Yu, and W. Xu. Hierarchical reinforcement learning by discovering intrinsic 08 options. In International Conference on Learning Representations, 2021. + +1. For all authors... + +(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] +(b) Did you describe the limitations of your work? [Yes] We discuss the limitations and directions of further investigation in the conclusion. +(c) Did you discuss any potential negative societal impacts of your work? [N/A] We do not foresee any obvious negative societal impacts from our work, which focuses on the fundamentals of reinforcement learning and proposes a new algorithm for unsupervised skill discovery. +(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] + +2. If you are including theoretical results... + +(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] + +3. If you ran experiments... + +(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] We plan to release our code upon acceptance of this work. +(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Appendix. +(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] Yes when possible. +(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No] + +4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... + +(a) If your work uses existing assets, did you cite the creators? [Yes] See Sect. 5. +(b) Did you mention the license of the assets? [N/A] +(c) Did you include any new assets either in the supplemental material or as a URL? [N/A] +(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] +(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] + +5. If you used crowdsourcing or conducted research with human subjects... + +(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] +(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] +(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] \ No newline at end of file diff --git a/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_content_list.json b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98363821998789a082104a819e64197611ff82eb --- /dev/null +++ b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_content_list.json @@ -0,0 +1,1041 @@ +[ + { + "type": "text", + "text": "Direct then Diffuse: Incremental Unsupervised Skill Discovery for State Covering and Goal Reaching ", + "text_level": 1, + "bbox": [ + 184, + 122, + 813, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous Author(s) \nAffiliation \nAddress \nemail ", + "bbox": [ + 423, + 226, + 578, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 318, + 535, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Learning meaningful behaviors in the absence of a task-specific reward function \n2 is a challenging problem in reinforcement learning. A desirable unsupervised \n3 objective is to learn a set of diverse skills that provide a thorough coverage of \n4 the state space while being directed, i.e., reliably reaching distinct regions of the \n5 environment. At test time, an agent could then leverage these skills to solve sparse \n6 reward problems by performing efficient exploration and finding an effective goal \n7 directed policy with little-to-no additional learning. Unfortunately, it is challenging \n8 to learn skills with such properties, as diffusing (e.g., stochastic policies performing \n9 good coverage) skills are not reliable in targeting specific states, whereas directed \n10 (e.g., goal-based policies) skills provide limited coverage. In this paper, inspired \n11 by the mutual information framework, we propose a novel algorithm designed \n12 to maximize coverage while ensuring a constraint on the directedness of each \n13 skill. In particular, we design skills with a decoupled policy structure, with a first \n14 part trained to be directed and a second diffusing part that ensures local coverage. \n15 Furthermore, we leverage the directedness constraint to adaptively add or remove \n16 skills as well as incrementally compose them along a tree that is grown to achieve a \n17 thorough coverage of the environment. We illustrate how our learned skills enables \n18 to efficiently solve sparse-reward downstream tasks in navigation and continuous \n19 control environments, where it compares favorably with existing baselines. ", + "bbox": [ + 148, + 347, + 766, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "20 1 Introduction ", + "text_level": 1, + "bbox": [ + 148, + 623, + 312, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "21 Deep reinforcement learning (RL) algorithms have been shown to effectively solve a wide variety of \n22 complex problems [e.g., 23, 6, 31, 12, 2, 28]. However, they are often designed to solve one single \n23 task at a time and they need to restart the learning process from scratch for any new problem, even \n24 when it is defined on the very same environment (e.g., navigating to different locations in the same \n25 apartment). Recently, unsupervised RL (URL) has been proposed as an approach to address this \n26 limitation. In URL, the agent first interacts with the environment without any extrinsic reward signal. \n27 Afterward, the agent leverages the experience accumulated during the unsupervised learning phase to \n28 efficiently solve a variety of downstream tasks defined on the same environment. \n29 In this paper, we consider the URL setting where the agent starts from an initial state $s _ { 0 }$ and it resets \n30 to it every time the policy terminates. We focus on sparse-reward downstream tasks, which require \n31 effective exploration (i.e., via a thorough coverage of the state space) to find the goal as well as \n32 learning a policy reliably reaching the goal (i.e., a directed policy). \n33 We build on the insight that mutual information (MI) effectively formalizes the dual objective of \n34 learning skills that both cover and navigate the environment efficiently [e.g., 11]. Specifically, given \n35 the state variable $S$ and some variables $Z$ on which the skill policies are conditioned, MI is defined as ", + "bbox": [ + 147, + 650, + 825, + 761 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 765, + 825, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 825, + 825, + 867 + ], + "page_idx": 0 + }, + { + "type": "equation", + "img_path": "images/c3bfc0f38cdfab89e1181d0c4924a0c8dc8268d8d34cb919b94fb7bd9ae886a6.jpg", + "text": "$$\n\\begin{array} { r } { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\\\ { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\end{array} - \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } | Z ) = \\boldsymbol { \\mathcal { H } } ( Z ) - \\boldsymbol { \\mathcal { H } } ( Z | \\boldsymbol { S } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 312, + 869, + 683, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Submitted to 35th Conference on Neural Information Processing Systems (NeurIPS 2021). Do not distribute. ", + "bbox": [ + 168, + 922, + 815, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3a6a7f8b3e7aa8d82fa62e9127f2a972085e7545768b2da5305611566b631fe.jpg", + "image_caption": [ + "Figure 1: Overview of UPSIDE. The black dot corresponds to the initial state $s _ { 0 }$ . (A) A set of random skills is initialized, each skill being composed of a directed part (illustrated as a black arrow) and a diffusing part (red arrows), which induces a local coverage (colored circles). $( B )$ The policies associated to the directed part of each skill are then updated to maximize the discriminability of the states reached by their diffusing part (Sect. 3.1). $( C )$ The least discriminable skills are iteratively removed while the policies of the remaining skills are re-optimized. This is executed until the discriminability of each skill satisfies a given constraint (see Sect. 3.2). In this example three skills are kept. $( D )$ One of these learned skill is then used as basis to add new skills, which are then optimized following the same procedure. For the “red” and “purple” skills, UPSIDE is not able to find sub-skills of sufficient quality and thus they are not expanded any further. $( E )$ At the end of the process, UPSIDE has created a tree of directed skills covering the state space (Sect. 3.3). These covering skills can then be used to solve downstream tasks. Moreover, the discriminator learned together with the skills can be used to select the skill to reach any specific goal region, where the directed parts get close to the goal, while the diffusing part provides the local coverage to attain the goal. The complete algorithm is detailed in Sect. 3.4 and Appendix. " + ], + "image_footnote": [], + "bbox": [ + 209, + 93, + 789, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "36 where $\\mathcal { T }$ denotes the MI and $\\mathcal { H }$ is the entropy function. The first expression, known as the forward \n37 form of MI, explicitly balances the two sought-after properties of coverage — captured by the entropy \n38 over the state space $\\mathcal { H } ( S )$ — and directedness, i.e., the ability to reach specific states $S$ depending \n39 on $Z$ — captured by the negative conditional entropy $- { \\mathcal { H } } ( S | Z )$ . The second expression of (1), often \n40 easier to optimize and referred to as the reverse form, stipulates that the skills should be sampled as \n41 diversely as possible while being discriminable. \n42 Maximizing (1) has been shown to be a powerful approach for encouraging exploration in RL [16, 25] \n43 and for unsupervised skill discovery [e.g., 11, 9, 1, 30, 8]. Nonetheless, learning skills that maximize \n44 the MI is a challenging optimization problem. Several approximations have been proposed to simplify \n45 the problem at the cost of possibly deviating from the original objective of coverage and directedness \n46 (see Sect. 4 for a review of related work). In this paper, we propose UPSIDE (UnsuPervised Skills \n47 that dIrect then DiffusE) to learn skills that can be effectively used to solve goal-based downstream \n48 tasks. Our solution builds on the following components (see Fig. 1 for an illustration of UPSIDE): ", + "bbox": [ + 147, + 433, + 825, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 521, + 825, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• Skill structure. In order to balance coverage and directedness, we design skills composed of two parts: 1) a directed part that is trained to reach a distinct region of the environment, and 2) a diffusing part that covers the states around the region attained by the first part. Optimization. We further strengthen the coverage and directedness properties of the skills by turning the MI objective into a constrained optimization problem designed to maximize coverage under the constraint that each skill achieves a minimum level of discriminability. This in turn enables UPSIDE to adaptively add skills to improve coverage, when all the initial skills meet the constraint, or remove those that violate the constraint to guarantee that each skill is directed and reaches a distinct region of the environment. • Tree structure. When the agent starts from a fixed initial state, the skills’ length is a crucial parameter, where short skills do not allow for proper coverage, and long skills are difficult to train. In UPSIDE we consider short skills to make the optimization easier, while composing them along a tree structure that ensures an adaptive and deep coverage of the environment. ", + "bbox": [ + 163, + 622, + 825, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "62 We study how our learned skill structure enables to both perform efficient exploration and learn \n63 effective goal-reaching policies in a variety of navigation and continuous control environments \n64 (including MuJoCo’s reacher) and we compare its performance to relevant baselines. ", + "bbox": [ + 148, + 809, + 825, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Setting ", + "text_level": 1, + "bbox": [ + 165, + 861, + 264, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "66 We consider the URL setting where the agent interacts with a Markov decision process (MDP) $M$ with state space 67 $s$ , action space $\\mathcal { A }$ , dynamics $p ( s ^ { \\prime } | s , a )$ , and no reward. The agent starts each ", + "bbox": [ + 148, + 883, + 825, + 912 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "68 episode from a designated initial state $s _ { 0 } \\in S$ . Upon termination of the chosen policy, the agent is \n69 then reset to $s _ { 0 }$ . This setting is particularly challenging from an exploration point of view since the \n70 agent cannot rely on the initial distribution to cover the state space. \n71 We recall the MI-based unsupervised skill discovery approach [see e.g., 11]. Denote by $Z$ some \n72 (latent) variables on which the skills of length $T$ are conditioned. There are three optimization \n73 variables: $( i )$ the support of the skills denoted by $| Z |$ (we consider it to be discrete so $| Z |$ is the \n74 number of skills), $( i i )$ the policy $\\pi ( z )$ associated to skill $z$ , and $( i i i )$ the sampling rule $\\rho$ (i.e., $\\rho ( z )$ \n75 is the probability of sampling skill $z$ at the beginning of the episode). Let the variable $S _ { T }$ be the \n76 random (final) state induced by sampling a skill $z$ from $\\rho$ and executing the associated policy $\\pi ( z )$ \n77 from $s _ { 0 }$ for an episode. We denote by $p _ { \\pi ( z ) } ( s _ { T } )$ the distribution over (final) states induced by \n78 executing the policy of skill $z$ , by $p ( z | s _ { T } )$ the probability of $z$ being the skill to induce state $s _ { T }$ , and \n79 let $\\begin{array} { r } { \\bar { p } ( s _ { T } ) \\bar { ( s _ { T } ) } = \\bar { \\sum _ { z \\in Z } \\bar { \\rho ( z ) } } p _ { \\pi ( z ) } ( s _ { T } ) } \\end{array}$ . Then maximizing the MI between $Z$ and $S _ { T }$ can be written as ", + "bbox": [ + 151, + 90, + 823, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 145, + 138, + 825, + 268 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/12ce697ef699aff72c90ae32fd9cf68f47bc811b82b8a62f65f86a70a2e9cb19.jpg", + "text": "$$\n\\begin{array} { r l r } & { } & { \\displaystyle \\underset { | Z | , \\rho , \\pi } { \\operatorname* { m a x } } \\ : \\mathcal { Z } ( S _ { T } ; Z ) = \\mathcal { H } ( S _ { T } ) - \\mathcal { H } ( S _ { T } | Z ) = - { \\displaystyle \\sum _ { s _ { T } } } \\overline { { p } } ( s _ { T } ) \\log \\overline { { p } } ( s _ { T } ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p _ { \\pi ( z ) } ( s _ { T } ) \\right] } \\\\ & { } & { = \\mathcal { H } ( Z ) - \\mathcal { H } ( Z | S _ { T } ) = - { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\log \\rho ( z ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p ( z | s _ { T } ) \\right] , \\quad ( 1 ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 275, + 820, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "80 where in the expectations $s _ { T } \\sim p _ { \\pi ( z ) } ( s _ { T } )$ . As discussed in Sect. 1, learning the optimal $| Z | , \\rho ,$ and $\\pi$ \n81 is a challenging problem [see e.g., 11, 9, 8]. ", + "bbox": [ + 151, + 354, + 825, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Algorithm Structure ", + "text_level": 1, + "bbox": [ + 171, + 396, + 379, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UPSIDE is based on three main components: a) the skill learning corresponding to stage $A$ and $B$ of Fig. 1 and described in Sect. 3.1, b) a constrained optimization problem used to optimize the number of skills (stage $C$ and Sect. 3.2) and c) a tree-building procedure (stage $D$ and Sect. 3.3). Together, these components allow UPSIDE to discover skills that combine coverage and directedness. ", + "bbox": [ + 173, + 420, + 825, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Skill Structure and Optimization ", + "text_level": 1, + "bbox": [ + 171, + 486, + 442, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in e.g., [9, 30, 37], the level of stochasticity of each skill (e.g., induced via a regularization on the entropy over the actions) plays a key role in trading off coverage and directedness. In fact, while randomness promotes broader coverage, it may compromise the directedness of the skills. In fact, a highly stochastic skill tends to induce a distribution $p _ { \\pi ( z ) } ( s _ { T } )$ over final states with high entropy (thus decreasing $- \\mathcal { H } ( S _ { T } | Z ) )$ , which prevents the skill to be reusable in solving sparse-reward downstream tasks where the objective is to reliably reach a specific goal state of the environment. Determining how much stochasticity to inject to adequately balance both objectives and optimize (1) is a difficult problem.1 ", + "bbox": [ + 165, + 503, + 826, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "96 We propose to design skills with a decoupled policy structure: ", + "bbox": [ + 147, + 621, + 580, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "• A directed part (of length $T$ ) with low stochasticity and trained to reach a specific region of the environment. It is responsible for increasing the $- { \\mathcal { H } } ( S | Z )$ term in (1). • A diffusing part (of length $H$ ) with high stochasticity to promote local coverage of the states around the region reached by the directed part. It is responsible for increasing the $\\mathcal { H } ( S )$ term in (1). ", + "bbox": [ + 178, + 641, + 629, + 727 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cd135dcbe8681b37eac698ab1aa41d695eeae5d2288d124ad1c5924bd7a5fec1.jpg", + "image_caption": [ + "Figure 2: Directed and diffusing parts of the skill. " + ], + "image_footnote": [], + "bbox": [ + 640, + 613, + 831, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "98 Similar to prior work [e.g., 11, 9], the policy associated to the directed part of skill $z$ is trained to max \n99 imize an intrinsic reward $r _ { z } ( s ) \\approx p ( z | s )$ ,2 where $p ( z | s )$ measures the “discriminability” of the skill $z$ \n100 given the state $s$ . More formally, $\\pi ( z )$ maximizes the cumulative reward $\\begin{array} { r } { \\mathbb { E } _ { \\pi ( z ) } \\left[ \\sum _ { t = T + 1 } ^ { T + H } r _ { z } ( s _ { t } ) \\right] } \\end{array}$ \n101 over the states traversed by the policy during the diffusing part. In practice, we also add a small \n102 entropy regularization $\\mathcal { H } ( \\pi ( \\cdot | z , s _ { t } ) )$ to the directed policy in order to ensure a minimum level of \n103 exploration and make the learning more robust. For the diffusing part, we rely on a simple random \n104 walk policy (i.e., a stochastic policy with uniform distribution over actions). \n105 Intuitively, the diffusing part defines a cluster of states that is used as a goal for the directed part. \n106 This allows us to “ground” the latent variable representations of the skills $Z$ to specific regions of \n107 the environment (i.e., the clusters). As a result, maximizing the MI over such skills can be seen as \n108 learning a set of “cluster-conditioned”, and thus directed, policies. ", + "bbox": [ + 145, + 731, + 826, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 142, + 90, + 825, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Skill Support and Sampling Rule ", + "text_level": 1, + "bbox": [ + 171, + 160, + 442, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "110 The MI objective (1) crucially depends on the number of skills $( | Z | )$ and the distribution $\\rho ( z )$ \n111 Unfortunately, it is been shown [e.g., 8] that solving (1) is particularly challenging. In order to \n112 simplify the optimization and the associated learning problem, we modify (1) in two ways. \n113 First, coherently with the skill optimization detailed in Sect. 3.1, the random variable $S$ in the \n114 conditional entropy is any state reached during the diffusing part of the skill and not just the terminal \n115 state. More formally, we denote by $S _ { \\mathrm { d i f f } }$ the random variable and its distribution for a specific skill $z$ \n116 is $p _ { \\pi ( z ) } ( s _ { \\mathrm { d i f f } } ) = 1 / \\dot { H } \\sum _ { t = T + 1 } ^ { T + H } p _ { \\pi ( z ) } ( s _ { t } )$ , i.e., the distribution over states obtained by averaging the \n117 distributions at any of the steps in the diffusing part. Similarly, $p ( z | s _ { \\mathrm { d i f f } } )$ now denotes the probability \n118 of $z$ being the skill to traverse $s _ { \\mathrm { d i f f } }$ during its diffusing part. As a result, training the skills to maximize \n119 MI naturally leads the diffusing parts to “push” the directed parts away so as to reach diverse regions \n120 of the environment. The combination of “global” coverage of the directed parts and “local” coverage \n121 of the diffusing part ensures that the whole environment is properly visited with $| Z | \\ll S$ skills.3 \n122 Second, we introduce an alternative problem that simplifies the optimization while preserving the \n123 coverage and directedness properties of MI. This is achieved by introducing a stronger requirement \n124 on the discriminability. While the conditional entropy term $- { \\mathcal { H } } ( Z | S )$ in (1) promotes the discrim \n125 inability of skills on average, we argue that a more suitable objective is to constrain each skill to \n126 achieve a minimum level of discriminability. First, we move from the average to the minimum over \n127 skills by lower bounding the conditional entropy as ", + "bbox": [ + 142, + 181, + 825, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 140, + 231, + 825, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 140, + 366, + 825, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/3157110c6c064eebeada077a520a63d063d2f20ec053138229dd1dff3acd2c56.jpg", + "text": "$$\n- \\mathcal { H } ( Z | S _ { \\mathrm { d i f f } } ) = \\sum _ { z \\in Z } \\rho ( z ) \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\operatorname* { m i n } _ { z \\in Z } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 253, + 457, + 741, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "128 which leads to the following optimization (assuming $\\pi$ is fixed for convenience) ", + "bbox": [ + 142, + 500, + 697, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/2f057bb1c2ed4a8288cb1135168369eb9b6e22ba2a4b80fd54634b61436bd511.jpg", + "text": "$$\n\\operatorname* { m a x } _ { | Z | = N , \\rho } \\bigg \\{ \\mathcal { H } ( Z ) + \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\bigg \\} ,\n$$", + "text_format": "latex", + "bbox": [ + 346, + 522, + 650, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "129 where with an abuse of notation we use $z \\in [ N ]$ to denote all skills in a set $Z$ with cardinality $N$ \n130 Since (3) is a lower bound to MI, it tends to promote the same type of covering and directed skills. \n131 Furthermore, (2) no longer depends on the distribution over skills and the entropy term $\\mathcal { H } ( Z )$ is \n132 maximized by setting $\\rho$ to the uniform distribution over $N$ skills (i.e., $\\begin{array} { r } { \\operatorname* { m a x } _ { \\rho } \\mathcal { H } ( Z ) \\bar { = } \\log ( N ) ) } \\end{array}$ , thus \n133 simplifying the optimization, which now only depends on $N$ . \n134 While optimizing (3) promotes a cardinality $N$ such that all skills have good discriminability, a more \n135 convenient formulation is to explicitly set a minimum level of discriminability for all skills through \n136 the following constrained optimization problem: ", + "bbox": [ + 140, + 561, + 825, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 140, + 638, + 828, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/d4b44d35c1819e2e5cc983ddcde5fbbc82e0a5a49a11f7481ce5dd77a52ac2b5.jpg", + "text": "$$\n\\operatorname* { m a x } _ { N \\geq 1 } \\log ( N ) \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\log \\eta .\n$$", + "text_format": "latex", + "bbox": [ + 305, + 689, + 694, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "137 where $\\eta$ is a parameter that defines the discriminability threshold. A skill $z$ is said to be $\\eta$ -consolidated \n138 if it satisfies the constraint. Crucially, let $\\begin{array} { r } { P _ { N } : = \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] } \\end{array}$ , then the sequence \n139 $( P _ { N } ) _ { N \\ge 1 }$ is non-increasing with $P _ { 1 } = 0$ (i.e., the more skills the harder it is to meet the constraint). \n140 As a result, (4) can be optimized following a simple greedy strategy incrementally adding skills until \n141 the constraint is violated. The optimal $N$ thus defines the effective number of $\\eta$ -consolidated skills and \n142 it corresponds to the largest number of skills that is guaranteed to display sufficient discriminability. \n143 Alternatively, we can interpret (4) as finding the largest number of clusters (i.e., the region reached \n144 by the directed part of a skill and covered by its associated diffusing part) with a minimum level of \n145 inter-cluster distance. This effect is qualitatively illustrated in Fig. 1, where the states attained by the \n146 directed part of the skills attain different regions that are locally covered by their diffusing parts. ", + "bbox": [ + 140, + 722, + 825, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: UPSIDE ", + "text_level": 1, + "bbox": [ + 174, + 94, + 321, + 109 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialize: Discriminability threshold $\\eta \\in ( 0 , 1 )$ , branching factor $N _ { 0 } \\geq 1$ , patience $K$ \nInitialize: Tree $\\tau$ initialized as a root node indexed by 0, queue of parent nodes $\\mathcal { W } = \\{ 0 \\}$ . \nwhile $\\mathcal { W } \\neq \\emptyset$ do // tree expansion \nDequeue a node/skill $w \\in \\mathcal W$ and expand $\\tau$ at $w$ by adding a set $\\mathcal { C } ( w )$ of $N _ { 0 }$ nodes/skills \n2 Create random policies $\\pi _ { z }$ , $\\forall z \\in \\mathcal { C } ( \\bar { \\boldsymbol { w } } )$ \n3 Initialize discriminator $q _ { \\phi }$ with $| \\tau |$ classes \n4 Continue $=$ true; Saturated $=$ false \n5 while Continue do \n6 for $K$ iterations do \n7 Sample a skill $z$ from $\\tau$ at random \n8 Extract the sequence of nodes $z _ { ( 1 ) } , \\dotsc , z$ in $\\tau$ leading to $z$ \n9 Execute the composed (directed part) policy $( \\pi _ { z _ { ( 1 ) } } , \\ldots , \\pi _ { z } )$ followed by the diffusing part \n10 Add states observed during the diffusion part to state buffer $B _ { z }$ \n11 Update discriminator $q _ { \\phi }$ with SGD on $B _ { z }$ to predict label $z$ \n12 if $z \\in \\mathcal { C } ( w )$ then $/ /$ Update only new policies, other polices kept fixed \n13 Update policy $\\pi _ { z }$ using SAC to optimize the discriminator reward as in Sect. 3.1. \n14 Compute the skill-discriminability $\\begin{array} { r } { d ( z ) = \\widehat { q } _ { \\phi } ^ { \\scriptscriptstyle ( B ) } ( z ) = \\frac { 1 } { | \\mathscr { B } _ { z } | } \\sum _ { s \\in \\mathscr { B } _ { z } } q _ { \\phi } ( z | s ) } \\end{array}$ for all $z \\in \\mathcal { C } ( w )$ \n15 if $\\begin{array} { r } { \\operatorname* { m i n } _ { z \\in \\mathcal { C } ( w ) } d ( z ) < \\eta } \\end{array}$ then $/ /$ Node removal \n16 Remove the node/skill $z = \\arg \\operatorname* { m i n } _ { z \\in { \\mathcal { C } } ( w ) } d ( z )$ from $\\mathcal { C } ( w )$ and $\\tau$ \n17 Set Saturate $=$ true \n18 else if not Saturated then \n19 Add one new node/skill to $\\mathcal { C } ( w )$ and $\\tau$ \n20 else \n21 Set Continue $=$ false \n22 Enqueue in $\\mathcal { W }$ the consolidated nodes $\\mathcal { C } ( w )$ ", + "bbox": [ + 158, + 112, + 787, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "147 3.3 Composing Skills in a Tree Structure ", + "text_level": 1, + "bbox": [ + 147, + 477, + 470, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MI optimization problem as well as our constrained variant (4) depend on the initial state $s _ { 0 }$ and on the length of each skill. Although these quantities are usually predefined and only appear implicitly in the equations, they have a crucial impact on the obtained behavior. In fact, resetting after each skill execution unavoidably restricts the coverage to a radius of at most $T + H$ steps around $s _ { 0 }$ This may suggest to set $T$ and $H$ to a large value. However, increasing the horizon makes the training of the skills more challenging, as learning $\\pi$ would require solving a difficult RL problem itself. ", + "bbox": [ + 171, + 498, + 825, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead, we propose to “extend” the length of the skills through composition. Indeed, the decoupled skill structure and the constraint in (4) entail that the directed part of each of the $\\eta$ -consolidated skills reliably reach a specific (and distinct) region of the environment and it is thus re-usable and amenable to composition. We propose to chain the directed part of the skills in order to reach further and further parts of the state space. Specifically, we build a growing tree, where the root is the initial state $s _ { 0 }$ , the edges represent the directed part of the skills, and the nodes represent the diffusing part of skills. As such, whenever a skill $z$ is selected, the directed part of all the policies associated to its predecessor skills in the tree are executed first (see Fig. 1 for an illustration of the tree structure). ", + "bbox": [ + 168, + 588, + 825, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a result, the agent naturally builds a curriculum on the episode lengths, which grow as the sequence $( i T + H ) _ { i \\geq 1 }$ . As such, it does not require prior knowledge on an adequate horizon of the downstream goal-based task.4 Here this knowledge is replaced by $T$ and $H$ which are more environment-agnostic and task-agnostic quantities, as their choice rather has an impact on the size and shape of the learned tree (e.g., the smaller $T$ and $H$ the bigger the tree). ", + "bbox": [ + 173, + 705, + 825, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 The UPSIDE Algorithm ", + "text_level": 1, + "bbox": [ + 176, + 790, + 372, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We are now ready to introduce UPSIDE, which provides a specific implementation of the components described before (see Fig. 1 for a qualitative illustration and Algorithm 1 for the detailed pseudo-code). ", + "bbox": [ + 171, + 810, + 821, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "0 We perform standard approximations to make the constraint in (4) easier to estimate. We approximate the unknown posterior $p ( z | s )$ with a learned discriminator $q _ { \\phi } ( z | s )$ with parameters $\\phi$ . We also ", + "bbox": [ + 158, + 844, + 823, + 873 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "172 remove the logarithm from the constraint to have an estimation range of $[ 0 , 1 ]$ and thus lower \n173 variance2. Finally, we replace the expectation over $s$ with an empirical estimate $\\widehat { q } _ { \\phi } ^ { ( B ) } ( z )$ averaging the \n174 value of the discriminator evaluated on the last $B$ states observed while executing the diffusing part \n175 of $z$ . Integrating these approximations in (4) leads to ", + "bbox": [ + 140, + 90, + 825, + 148 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/3cf11c6f247fdf3cb31c0bd3fbdd8676fa016c31b28947636c19d28d2e51cbd2.jpg", + "text": "$$\n\\operatorname* { m a x } _ { N \\geq 1 , \\pi } N \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\widehat { q } _ { \\phi } ^ { ( B ) } ( z ) \\geq \\eta .\n$$", + "text_format": "latex", + "bbox": [ + 369, + 156, + 629, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "176 As discussed in Sect. 3.2, this problem can be conveniently optimized using a greedy strategy. We \n177 then integrate the optimization of (5) into an adaptive tree expansion strategy: (Generating new \n178 skills) Given a tree structure as described in Sect. 3.3, we expand the tree at a leaf $w$ by adding $N _ { 0 }$ \n179 new nodes/skills following a breadth-first-search approach (lines 1, 2). Then (Skill Learning) the \n180 new skills are optimized by: i) sampling random skills in the tree to update the discriminator (lines \n181 7-11), and ii) by updating the policies to optimize the discriminability reward (Sect. 3.1) computed \n182 using the discriminator (lines 13). To speed-up convergence, we only update the policies that have be \n183 added to the tree structure, keeping all the previous policies fixed (line 12). Note that in the update of \n184 the discriminator we leverage the states observed in previous phases of the algorithm by maintaining \n185 a (small) replay buffer of states for each skill. (Node Consolidation) After a patience period (line 6), \n186 if all skills are $\\eta$ -consolidated, we tentatively add more skills to the leaf $w$ (line 18). On the other \n187 hand, if any skill does not meet the discriminability threshold, we remove it and consolidate the \n188 remaining skills into the tree (lines 16, 17) and we repeat the process. \n189 Model selection. A core aspect of any RL algorithm is model selection, i.e., finding the best \n190 configuration of hyperparameters. In URL with no prior knowledge of the downstream task(s), it \n191 is non-trivial to devise an adequate criterion for model selection and this aspect is rarely addressed, \n192 despite being crucial in practice. For instance, while the coverage of the state space may be a \n193 good proxy for the performance of a URL algorithm [see e.g., 8], it may be difficult to measure in \n194 continuous problems. Interestingly, our optimization problem directly provides a single, task-agnostic \n195 and environment-agnostic criterion for model selection, which is the number $N$ of $\\eta$ -consolidated \n196 skills discovered by the agent. Indeed in all of our experiments we simply select the model (i.e., set \n197 of hyperparameters) that maximizes $N$ . This is a significant advantage w.r.t. existing methods, such \n198 as VIC and DIAYN, for which no principled approach to model selection is provided. ", + "bbox": [ + 140, + 190, + 825, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 140, + 377, + 825, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "199 4 Related work ", + "text_level": 1, + "bbox": [ + 147, + 529, + 316, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unsupervised Reinforcement Learning methods can be broadly decomposed according to the way they summarize the experience accumulated during the unsupervised phase into reusable knowledge to solve downstream tasks. This includes both off-policy model-free [e.g., 27] and model-based [e.g., 29] methods that seek to populate a representative replay buffer and build accurate value or model estimates, that are used to solve a given downstream task in a zero- or few-shot manner. The accumulated experience during train time can also be compressed into a low-dimensional representation for value functions as well as policies and to improve exploration [e.g., 36]. An alternative line of work focuses on the discovery of a set of skills in an unsupervised manner. Our approach falls in this category, on which we now focus our related work review. ", + "bbox": [ + 166, + 560, + 825, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "209 Skill discovery based on MI maximization was first proposed in VIC [11], where only the final \n210 states of each trajectory are considered in the reverse form of (1) and where both the skills and \n211 their sampling rules are simultaneously learned (with a fixed support $| Z |$ , i.e., a fixed number of \n212 skills). DIAYN [9] fixes the sampling rule to be uniform, and weighs the skills with an action-entropy \n213 coefficient (i.e., it additionally minimizes the MI between actions and skills given the state), so as \n214 to push the skills away from each other and enhance coverage. DADS [30] learns skills that are not \n215 only diverse but also predictable by learned dynamics models, by using a generative model over \n216 observations (rather than over skills) and optimizing a forward form of MI, namely $\\mathcal { T } ( s ^ { \\prime } ; z | s )$ between \n217 the next state $s ^ { \\prime }$ and current skill $z$ (with continuous latent) conditioned on the current state $s$ . EDL [8] \n218 shows that existing skill discovery approaches can provide insufficient coverage, and instead proposes \n219 to rely on a fixed distribution over states $p ( s )$ which is either provided by an oracle or learned. In \n220 SMM [19], the MI formalism is used to learn a policy for which the state marginal distribution matches \n221 a given target state distribution (e.g., uniform), which can be seen as a more scalable way of tackling \n222 the problem of maximum entropy over the state space [15], and as a way to encourage skills to go \n223 through unknown state regions. Other MI-based skill discovery methods include [10, 14, 24, 5, 34], \n224 as well as [35, 20] which investigate skill discovery in non-episodic settings. \n225 Our approach shares a similar motivation to prior MI-based works of targeting skills that are both \n226 directed and state-covering. In particular, the decoupled structure introduced in Sect. 3.1 can be seen \n227 as a more suitable way to achieve the objective of improving the coverage of VIC as done in DIAYN \n228 and SMM, without compromising the directedness of the skills. \n229 While most skill discovery approaches consider a fixed number of skills, a curriculum with increasing \n230 number of skills is studied in [1, 3]. Our discriminability constraint is what enables skills to be \n231 composed along a tree structure, which allows increases or decreases the support of available skills \n232 depending on the region of the state space. ", + "bbox": [ + 138, + 690, + 825, + 911 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9c4d365719fb312ceafdef45970c4f57a1b44e9db7d7229c05847b9c4b09596.jpg", + "image_caption": [ + "Figure 3: UPSIDE, DIAYN-curriculum and SMM-10 skills learned in a bottleneck maze (Top) and a U-maze (Bottom). For both DIAYN and SMM we report the stochastic execution of the learned skills and for UPSIDE we report the deterministic directed parts (that are composed) followed by the (stochastic) diffusing part, which is the same protocol used to evaluate coverage. " + ], + "image_footnote": [], + "bbox": [ + 204, + 87, + 807, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 142, + 429, + 825, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 489, + 825, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recently, [37] proposed a hierarchical RL method that discovers abstract and task-agnostic skills while jointly learning a higher-level policy which is trained to maximize environment reward. Our approach builds on a similar promise of composing skills instead of resetting to $s _ { 0 }$ after each execution, yet we articulate the composition differently, by exploiting the direct-then-diffuse structure to ground learned skills to the state space instead of being abstract. ", + "bbox": [ + 160, + 550, + 825, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "238 In addition, approaches such as DISCERN [33] and Skew-Fit [27] learn a goal-conditioned policy in \n239 an unsupervised way with an MI objective. As explained in [8, Sect. 5], this can be interpreted as a \n240 skill discovery approach with latent $Z = S$ , i.e., where each goal state can define a different skill. \n241 Conditioning on either goal states or abstract latent skills forms two extremes of the spectrum of \n242 unsupervised RL. We target an intermediate approach, seeking to benefit from the groundedness of \n243 the latent skill $Z$ and the states $S$ (and thus amenability to composition) of goal-conditioned RL, and \n244 from the reduced search space and sampling ease of skill-based RL. ", + "bbox": [ + 140, + 625, + 825, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An alternative approach to skill discovery builds on “spectral” properties of the dynamics of the environment. This includes eigenoptions [21, 22] and covering options [17, 18], as well as the algorithm of [4] that builds a discrete graph representation which learns and composes spectral skills. ", + "bbox": [ + 165, + 726, + 823, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Experiments ", + "text_level": 1, + "bbox": [ + 171, + 782, + 312, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate the following questions: i) Can the adaptive tree structure of UPSIDE incrementally cover an unknown environment while preserving directedness of the skills? ii) Following the unsupervised phase, how can UPSIDE be leveraged to solve goal-based downstream tasks? ", + "bbox": [ + 174, + 808, + 825, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report results on: a) Navigation problems in continuous mazes, where actions represent the desired shift in $x$ and $y$ coordinates; b) A difficult instance of CartPole, where the cart starts with zero speed and the pole is oriented downside; c) The Reacher [32] problem using the MuJoCo implementation in Gym [7]. In all environments, the per-dimension action space is in $[ - 1 ; + 1 ]$ . ", + "bbox": [ + 168, + 856, + 823, + 911 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f67a926a679d5569359c50d0e257b1aa3baa0708a7b5b489b2d719b79dee97b3.jpg", + "image_caption": [ + "Figure 4: Normalized coverage in U-maze and bottleneck. " + ], + "image_footnote": [], + "bbox": [ + 176, + 88, + 785, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "256 We compare to different baselines. DIAYN-K, where $K$ is a fixed number of skills, is the original \n257 algorithm proposed in [9]. DIAYN-Curriculum is a variant where the number of skills is automatically \n258 tuned following the same procedure as in UPSIDE ensuring a good discriminability. We also compare \n259 to SMM [19], which is similar to DIAYN, but it includes an exploration bonus encouraging the policies \n260 to visit rarely encountered states. In our implementation, the exploration bonus is obtained by \n261 maintaining a multinomial distribution over “buckets of states” obtained by discretization, resulting \n262 in an computation-efficient and stable implementation that is more stable than the original VAE-based \n263 method. UPSIDE and all baselines are implemented with Soft-Actor Critic (SAC) [13]. \n264 Unsupervised Phase. We run all methods until convergence. We then do model selection according \n265 to the criterion of either the final number of skills for UPSIDE and DIAYN-curriculum and the final \n266 average discriminability for DIAYN-K and SMM. To compute the coverage, we perform rollouts by \n267 first sampling a skill uniformly at random and executing its associated policy until termination. We \n268 discretize states into buckets (50 interval per dimension for mazes and 10 for control environments) \n269 and report the proportion of buckets reached by each method as a function of the total number of \n270 steps executed in the environment over multiple rollouts. Since only a small portion of the discretized \n271 states can be reached, we normalize the coverage such that the best method obtains 1. \n272 We consider two topologies of mazes with size (height and width) 50 such that exploration is non \n273 trivial (i.e., a random policy is only able to cover a small part of the state space): a U-shaped maze \n274 and a Bottleneck maze (which is a harder version of the one in [8, Fig. 1] which is only of size 10 \n275 for the same action space). In Fig. 3 we show that UPSIDE succeeds in covering the near-entirety \n276 of the state space by creating a tree of directed skills. Moreover, UPSIDE created directed skills \n277 with a low entropy, while the two baselines tend to create skills that are more stochastic. This is \n278 particularly evident for SMM, due to the state-entropy exploration bonus, that while it encourages \n279 broader coverage makes skills less directed. ", + "bbox": [ + 140, + 284, + 825, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 143, + 400, + 825, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 142, + 517, + 825, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 4 we report the coverage on the Bottleneck maze and U-Maze. For UPSIDE, executing a skill corresponds to executing the directed part of all the “parent” skills in the tree and concluding with the diffusion part of the skill. SMM achieves better coverage than DIAYN thanks to the increased level of stochasticity (diffusion) of its skills. UPSIDE outperforms both by reaching regions of the environment that are not be achieved by other methods. Here, we plot UPSIDE with $T = 1 0$ and $H = 1 0$ , but we found UPSIDE to be robust to these parameters as shown in the supplementary. ", + "bbox": [ + 171, + 635, + 825, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results are similar in the CartPole problem (see Fig. 5) where UPSIDE (with $T = 2 0$ and $H = 2 0$ ) obtains better coverage than baselines. On the other hand, in Reacher (see Fig. 5), DIAYN-50 outperforms UPSIDE in terms of coverage. This can be explained by the fact that, in this environment, highly stochastic skills provide a good coverage. Nonetheless, this comes at the cost of very low discriminability (rightmost plot), which suggests DIAYN-50 skills have poor directedness. On the other hand, UPSIDE (and DIAYN-curriculum) achieves much larger discriminability by removing redundant skills and favoring more directed policies. ", + "bbox": [ + 165, + 724, + 825, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Downstream Tasks. Following the unsupervised phase, UPSIDE has learned a tree of skills. We now investigate how these skills are used to tackle a downstream task. In that setting, we propose to use skill-based approaches (i.e UPSIDE, DIAYN and SMM) in the following way: a) (exploration) first we sample rollouts over the different skills. b) We then select the best skill based on the maximum cumulative reward collected and c) we fine-tune this skill to maximize the reward. We report results on mazes (additional results are provided in the supplementary). We consider a sparse positive reward ", + "bbox": [ + 171, + 828, + 823, + 911 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0fef9fb524cc8a68d444c1ca195519b14d66a4947d40fb5df6c83c782ac91ca.jpg", + "image_caption": [ + "Figure 5: Normalized coverage in Cartpole (Left) and Reacher (Middle). (Right) Average discriminability of the skills during training in Reacher. " + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 818, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "299 when reaching a particular defined goal.5 We consider goals at different distances from the initial \n300 state $s _ { 0 }$ , the further, the harder. Fig. 6 shows the learning curves obtained when fine-tuning the best \n301 skill for the different models and compare to a classical SAC algorithm where a single policy is \n302 learned from scratch. DIAYN/SMM means we use the best state-covering policies between DIAYN and \n303 SMM. For the “close” goal setting, both UPSIDE and DIAYN/SMM are able to learn to reach this goal \n304 efficiently while SAC solves the task only for some of the training runs. Note that we do not show \n305 DIAYN performance since it is lower than the SMM one. For the “far” goal setting, only UPSIDE learns \n306 to reach this goal. Obtained trajectories are illustrated in Fig. 6. ", + "bbox": [ + 140, + 266, + 825, + 378 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5c8600e6316f678077840de51c561851a9e26dcd890e6d78ce9628180fe6febb.jpg", + "image_caption": [ + "(Left): Learning curves for “short” distance and (Middle) Figure 6: “medium” distance goals. (Right): Learned policies after fine-tuning (Top) U-maze. (Bottom): Bottleneck maze. " + ], + "image_footnote": [], + "bbox": [ + 179, + 390, + 808, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Conclusion ", + "text_level": 1, + "bbox": [ + 161, + 729, + 299, + 746 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduced UPSIDE, a novel algorithm for unsupervised skill discovery designed to trade off between coverage and directedness and develop a tree of skills that can be used to both perform efficient exploration of the environment and learn effective goal-directed policies. Natural venues for future investigation are: 1) The diffusing part of each skill could be explicitly trained to maximize local coverage; 2) UPSIDE assumes a good representation of the state is provided as input, it would be interesting to pair UPSIDE with effective representation learning techniques to tackle problems with high-dimensional input (e.g., image-based RL); 3) While UPSIDE is grounded on the solid principle of MI maximization, a more thorough theoretical investigation is needed to explicitly link the optimization problem and its approximations to the downstream performance. ", + "bbox": [ + 173, + 751, + 825, + 876 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References [1] J. Achiam, H. Edwards, D. Amodei, and P. Abbeel. Variational option discovery algorithms. arXiv preprint arXiv:1807.10299, 2018. [2] M. Andrychowicz, D. Crow, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin, P. Abbeel, and W. Zaremba. Hindsight experience replay. In NIPS, 2017. [3] A. Aubret, L. Matignon, and S. Hassas. Elsim: End-to-end learning of reusable skills through intrinsic motivation. arXiv preprint arXiv:2006.12903, 2020. [4] A. Bagaria, J. Crowley, J. W. N. Lim, and G. Konidaris. Skill discovery for exploration and planning using deep skill graphs. 2021. [5] K. Baumli, D. Warde-Farley, S. Hansen, and V. Mnih. Relative variational intrinsic control. arXiv preprint arXiv:2012.07827, 2020. [6] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279, 2013. [7] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba. Openai gym, 2016. [8] V. Campos, A. Trott, C. Xiong, R. Socher, X. Giro-i Nieto, and J. Torres. Explore, discover and learn: Unsupervised discovery of state-covering skills. In International Conference on Machine Learning, 2020. [9] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills without a reward function. In International Conference on Learning Representations, 2019. [10] C. Florensa, Y. Duan, and P. Abbeel. Stochastic neural networks for hierarchical reinforcement learning. arXiv preprint arXiv:1704.03012, 2017. [11] K. Gregor, D. J. Rezende, and D. Wierstra. Variational intrinsic control. arXiv preprint arXiv:1611.07507, 2016. [12] S. Gu, E. Holly, T. Lillicrap, and S. Levine. Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In 2017 IEEE international conference on robotics and automation (ICRA), pages 3389–3396. IEEE, 2017. \n45 [13] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. CoRR, abs/1801.01290, 2018. \n47 [14] S. Hansen, W. Dabney, A. Barreto, D. Warde-Farley, T. Van de Wiele, and V. Mnih. Fast task inference with variational intrinsic successor features. In International Conference on Learning Representations, 2019. \n50 [15] E. Hazan, S. Kakade, K. Singh, and A. Van Soest. Provably efficient maximum entropy exploration. In International Conference on Machine Learning, pages 2681–2691, 2019. [16] R. Houthooft, X. Chen, Y. Duan, J. Schulman, F. De Turck, and P. Abbeel. Vime: variational information maximizing exploration. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 1117–1125, 2016. [17] Y. Jinnai, J. W. Park, D. Abel, and G. Konidaris. Discovering options for exploration by minimizing cover time. In International Conference on Machine Learning, pages 3130–3139. PMLR, 2019. [18] Y. Jinnai, J. W. Park, M. C. Machado, and G. Konidaris. Exploration in reinforcement learning with deep covering options. In International Conference on Learning Representations, 2020. [19] L. Lee, B. Eysenbach, E. Parisotto, E. Xing, S. Levine, and R. Salakhutdinov. Efficient exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019. ", + "bbox": [ + 151, + 85, + 828, + 920 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "[20] K. Lu, A. Grover, P. Abbeel, and I. Mordatch. Reset-free lifelong learning with skill-space planning. arXiv preprint arXiv:2012.03548, 2020. [21] M. C. Machado, M. G. Bellemare, and M. Bowling. A laplacian framework for option discovery in reinforcement learning. In International Conference on Machine Learning, pages 2295–2304. PMLR, 2017. [22] M. C. Machado, C. Rosenbaum, X. Guo, M. Liu, G. Tesauro, and M. Campbell. Eigenoption discovery through the deep successor representation. In International Conference on Learning Representations, 2018. [23] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529–533, 2015. [24] N. Modhe, P. Chattopadhyay, M. Sharma, A. Das, D. Parikh, D. Batra, and R. Vedantam. Ir-vic: Unsupervised discovery of sub-goals for transfer in rl. In Proceedings of the TwentyNinth International Joint Conference on Artificial Intelligence, IJCAI-20. International Joint Conferences on Artificial Intelligence Organization, 2020. [25] S. Mohamed and D. J. Rezende. Variational information maximisation for intrinsically motivated reinforcement learning. In Advances in neural information processing systems, pages 2125– 2133, 2015. [26] M. Mutti, L. Pratissoli, and M. Restelli. A policy gradient method for task-agnostic exploration. arXiv preprint arXiv:2007.04640, 2020. [27] V. H. Pong, M. Dalal, S. Lin, A. Nair, S. Bahl, and S. Levine. Skew-fit: State-covering self-supervised reinforcement learning. In International Conference on Machine Learning, 2020. [28] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. [29] R. Sekar, O. Rybkin, K. Daniilidis, P. Abbeel, D. Hafner, and D. Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, pages 8583–8592. PMLR, 2020. [30] A. Sharma, S. Gu, S. Levine, V. Kumar, and K. Hausman. Dynamics-aware unsupervised discovery of skills. In International Conference on Learning Representations, 2020. [31] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, et al. Mastering the game of go without human knowledge. nature, 550(7676):354–359, 2017. [32] E. Todorov, T. Erez, and Y. Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 50pou26–5033, 2012. [33] D. Warde-Farley, T. Van de Wiele, T. Kulkarni, C. Ionescu, S. Hansen, and V. Mnih. Unsupervised control through non-parametric discriminative rewards. In International Conference on Learning Representations, 2019. [34] K. Xie, H. Bharadhwaj, D. Hafner, A. Garg, and F. Shkurti. Skill transfer via partially amortized hierarchical planning. In International Conference on Learning Representations, 2021. [35] K. Xu, S. Verma, C. Finn, and S. Levine. Continual learning of control primitives: Skill discovery via reset-games. arXiv preprint arXiv:2011.05286, 2020. [36] D. Yarats, R. Fergus, A. Lazaric, and L. Pinto. Reinforcement learning with prototypical representations. arXiv preprint arXiv:2102.11271, 2021. [37] J. Zhang, H. Yu, and W. Xu. Hierarchical reinforcement learning by discovering intrinsic 08 options. In International Conference on Learning Representations, 2021. ", + "bbox": [ + 155, + 71, + 828, + 916 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "1. For all authors... ", + "bbox": [ + 214, + 116, + 339, + 130 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] We discuss the limitations and directions of further investigation in the conclusion. \n(c) Did you discuss any potential negative societal impacts of your work? [N/A] We do not foresee any obvious negative societal impacts from our work, which focuses on the fundamentals of reinforcement learning and proposes a new algorithm for unsupervised skill discovery. \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ", + "bbox": [ + 238, + 136, + 825, + 281 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "2. If you are including theoretical results... ", + "bbox": [ + 215, + 285, + 493, + 300 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ", + "bbox": [ + 236, + 304, + 738, + 335 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3. If you ran experiments... ", + "bbox": [ + 214, + 339, + 393, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] We plan to release our code upon acceptance of this work. \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Appendix. \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] Yes when possible. \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No] ", + "bbox": [ + 238, + 358, + 825, + 491 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ", + "bbox": [ + 217, + 494, + 823, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) If your work uses existing assets, did you cite the creators? [Yes] See Sect. 5. \n(b) Did you mention the license of the assets? [N/A] \n(c) Did you include any new assets either in the supplemental material or as a URL? [N/A] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] ", + "bbox": [ + 238, + 513, + 825, + 635 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5. If you used crowdsourcing or conducted research with human subjects... ", + "bbox": [ + 214, + 638, + 705, + 654 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ", + "bbox": [ + 238, + 657, + 825, + 747 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_middle.json b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..5815dd050cf4f6008448d1588582ca7c347f07ac --- /dev/null +++ b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_middle.json @@ -0,0 +1,42413 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 113, + 97, + 498, + 137 + ], + "lines": [ + { + "bbox": [ + 112, + 97, + 498, + 118 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 498, + 118 + ], + "score": 1.0, + "content": "Direct then Diffuse: Incremental Unsupervised Skill", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 116, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 125, + 116, + 486, + 140 + ], + "score": 1.0, + "content": "Discovery for State Covering and Goal Reaching", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 259, + 179, + 354, + 223 + ], + "lines": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "spans": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "score": 1.0, + "content": "Anonymous Author(s)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "spans": [ + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "score": 1.0, + "content": "Affiliation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "spans": [ + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "score": 1.0, + "content": "Address", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "spans": [ + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "score": 1.0, + "content": "email", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 283, + 252, + 328, + 265 + ], + "lines": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "spans": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 91, + 275, + 469, + 483 + ], + "lines": [ + { + "bbox": [ + 93, + 275, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 93, + 279, + 99, + 285 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 141, + 275, + 469, + 288 + ], + "score": 1.0, + "content": "Learning meaningful behaviors in the absence of a task-specific reward function", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 286, + 470, + 299 + ], + "spans": [ + { + "bbox": [ + 93, + 289, + 99, + 297 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 141, + 286, + 470, + 299 + ], + "score": 1.0, + "content": "is a challenging problem in reinforcement learning. A desirable unsupervised", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 297, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 93, + 300, + 99, + 308 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 141, + 297, + 470, + 309 + ], + "score": 1.0, + "content": "objective is to learn a set of diverse skills that provide a thorough coverage of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 92, + 311, + 99, + 320 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 142, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "the state space while being directed, i.e., reliably reaching distinct regions of the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 93, + 318, + 470, + 332 + ], + "spans": [ + { + "bbox": [ + 93, + 321, + 100, + 330 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 141, + 318, + 470, + 332 + ], + "score": 1.0, + "content": "environment. At test time, an agent could then leverage these skills to solve sparse", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 330, + 470, + 342 + ], + "spans": [ + { + "bbox": [ + 93, + 333, + 99, + 341 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 141, + 330, + 470, + 342 + ], + "score": 1.0, + "content": "reward problems by performing efficient exploration and finding an effective goal-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 92, + 339, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 92, + 343, + 99, + 352 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 141, + 339, + 470, + 354 + ], + "score": 1.0, + "content": "directed policy with little-to-no additional learning. Unfortunately, it is challenging", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 92, + 351, + 470, + 365 + ], + "spans": [ + { + "bbox": [ + 92, + 354, + 99, + 363 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 141, + 351, + 470, + 365 + ], + "score": 1.0, + "content": "to learn skills with such properties, as diffusing (e.g., stochastic policies performing", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 92, + 362, + 470, + 376 + ], + "spans": [ + { + "bbox": [ + 92, + 365, + 100, + 374 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 141, + 362, + 470, + 376 + ], + "score": 1.0, + "content": "good coverage) skills are not reliable in targeting specific states, whereas directed", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 374, + 470, + 387 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 99, + 385 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 141, + 374, + 470, + 387 + ], + "score": 1.0, + "content": "(e.g., goal-based policies) skills provide limited coverage. In this paper, inspired", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 384, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 90, + 387, + 99, + 396 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 141, + 384, + 470, + 397 + ], + "score": 1.0, + "content": "by the mutual information framework, we propose a novel algorithm designed", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 395, + 470, + 408 + ], + "spans": [ + { + "bbox": [ + 89, + 397, + 100, + 407 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 141, + 395, + 470, + 408 + ], + "score": 1.0, + "content": "to maximize coverage while ensuring a constraint on the directedness of each", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 407, + 470, + 418 + ], + "spans": [ + { + "bbox": [ + 90, + 408, + 99, + 418 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 141, + 407, + 470, + 418 + ], + "score": 1.0, + "content": "skill. In particular, we design skills with a decoupled policy structure, with a first", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 89, + 416, + 470, + 430 + ], + "spans": [ + { + "bbox": [ + 89, + 419, + 100, + 429 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 141, + 416, + 470, + 430 + ], + "score": 1.0, + "content": "part trained to be directed and a second diffusing part that ensures local coverage.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 428, + 469, + 441 + ], + "spans": [ + { + "bbox": [ + 89, + 430, + 100, + 440 + ], + "score": 1.0, + "content": "15", + "type": "text" + }, + { + "bbox": [ + 141, + 428, + 469, + 441 + ], + "score": 1.0, + "content": "Furthermore, we leverage the directedness constraint to adaptively add or remove", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 438, + 470, + 452 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 100, + 451 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 141, + 438, + 470, + 452 + ], + "score": 1.0, + "content": "skills as well as incrementally compose them along a tree that is grown to achieve a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 449, + 470, + 462 + ], + "spans": [ + { + "bbox": [ + 89, + 452, + 100, + 462 + ], + "score": 1.0, + "content": "17", + "type": "text" + }, + { + "bbox": [ + 141, + 449, + 470, + 462 + ], + "score": 1.0, + "content": "thorough coverage of the environment. We illustrate how our learned skills enables", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 90, + 461, + 469, + 473 + ], + "spans": [ + { + "bbox": [ + 90, + 463, + 99, + 472 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 142, + 461, + 469, + 473 + ], + "score": 1.0, + "content": "to efficiently solve sparse-reward downstream tasks in navigation and continuous", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 90, + 472, + 443, + 484 + ], + "spans": [ + { + "bbox": [ + 90, + 474, + 99, + 483 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 142, + 472, + 443, + 484 + ], + "score": 1.0, + "content": "control environments, where it compares favorably with existing baselines.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 91, + 494, + 191, + 508 + ], + "lines": [ + { + "bbox": [ + 86, + 493, + 192, + 510 + ], + "spans": [ + { + "bbox": [ + 86, + 493, + 192, + 510 + ], + "score": 1.0, + "content": "20 1 Introduction", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 90, + 515, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 89, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 89, + 516, + 99, + 526 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) algorithms have been shown to effectively solve a wide variety of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 89, + 525, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 528, + 100, + 538 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 105, + 525, + 506, + 539 + ], + "score": 1.0, + "content": "complex problems [e.g., 23, 6, 31, 12, 2, 28]. However, they are often designed to solve one single", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 536, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 89, + 539, + 100, + 549 + ], + "score": 1.0, + "content": "23", + "type": "text" + }, + { + "bbox": [ + 105, + 536, + 506, + 550 + ], + "score": 1.0, + "content": "task at a time and they need to restart the learning process from scratch for any new problem, even", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 88, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 88, + 550, + 100, + 559 + ], + "score": 1.0, + "content": "24", + "type": "text" + }, + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "when it is defined on the very same environment (e.g., navigating to different locations in the same", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 89, + 561, + 100, + 571 + ], + "score": 1.0, + "content": "25", + "type": "text" + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "apartment). Recently, unsupervised RL (URL) has been proposed as an approach to address this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 570, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 88, + 572, + 100, + 581 + ], + "score": 1.0, + "content": "26", + "type": "text" + }, + { + "bbox": [ + 106, + 570, + 506, + 582 + ], + "score": 1.0, + "content": "limitation. In URL, the agent first interacts with the environment without any extrinsic reward signal.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 581, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 89, + 583, + 100, + 592 + ], + "score": 1.0, + "content": "27", + "type": "text" + }, + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "score": 1.0, + "content": "Afterward, the agent leverages the experience accumulated during the unsupervised learning phase to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 592, + 430, + 604 + ], + "spans": [ + { + "bbox": [ + 89, + 594, + 100, + 603 + ], + "score": 1.0, + "content": "28", + "type": "text" + }, + { + "bbox": [ + 105, + 592, + 430, + 604 + ], + "score": 1.0, + "content": "efficiently solve a variety of downstream tasks defined on the same environment.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 90, + 606, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 89, + 608, + 100, + 618 + ], + "score": 1.0, + "content": "29", + "type": "text" + }, + { + "bbox": [ + 105, + 606, + 443, + 619 + ], + "score": 1.0, + "content": "In this paper, we consider the URL setting where the agent starts from an initial state", + "type": "text" + }, + { + "bbox": [ + 444, + 608, + 454, + 618 + ], + "score": 0.86, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "and it resets", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 616, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 619, + 100, + 629 + ], + "score": 1.0, + "content": "30", + "type": "text" + }, + { + "bbox": [ + 105, + 616, + 506, + 631 + ], + "score": 1.0, + "content": "to it every time the policy terminates. We focus on sparse-reward downstream tasks, which require", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 89, + 630, + 100, + 640 + ], + "score": 1.0, + "content": "31", + "type": "text" + }, + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "effective exploration (i.e., via a thorough coverage of the state space) to find the goal as well as", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 89, + 639, + 374, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 641, + 100, + 651 + ], + "score": 1.0, + "content": "32", + "type": "text" + }, + { + "bbox": [ + 105, + 639, + 374, + 653 + ], + "score": 1.0, + "content": "learning a policy reliably reaching the goal (i.e., a directed policy).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 90, + 654, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 90, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 90, + 656, + 100, + 664 + ], + "score": 1.0, + "content": "33", + "type": "text" + }, + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "We build on the insight that mutual information (MI) effectively formalizes the dual objective of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 89, + 667, + 100, + 676 + ], + "score": 1.0, + "content": "34", + "type": "text" + }, + { + "bbox": [ + 104, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "learning skills that both cover and navigate the environment efficiently [e.g., 11]. Specifically, given", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 89, + 678, + 100, + 687 + ], + "score": 1.0, + "content": "35", + "type": "text" + }, + { + "bbox": [ + 105, + 676, + 174, + 688 + ], + "score": 1.0, + "content": "the state variable", + "type": "text" + }, + { + "bbox": [ + 174, + 676, + 182, + 686 + ], + "score": 0.84, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 676, + 260, + 688 + ], + "score": 1.0, + "content": "and some variables", + "type": "text" + }, + { + "bbox": [ + 261, + 676, + 269, + 686 + ], + "score": 0.82, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "on which the skill policies are conditioned, MI is defined as", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 689, + 418, + 704 + ], + "lines": [ + { + "bbox": [ + 191, + 689, + 418, + 704 + ], + "spans": [ + { + "bbox": [ + 191, + 689, + 418, + 704 + ], + "score": 0.8, + "content": "\\begin{array} { r } { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\\\ { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\end{array} - \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } | Z ) = \\boldsymbol { \\mathcal { H } } ( Z ) - \\boldsymbol { \\mathcal { H } } ( Z | \\boldsymbol { S } ) ,", + "type": "interline_equation", + "image_path": "c3bfc0f38cdfab89e1181d0c4924a0c8dc8268d8d34cb919b94fb7bd9ae886a6.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 191, + 689, + 418, + 704 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 103, + 731, + 499, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 500, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 500, + 743 + ], + "score": 1.0, + "content": "Submitted to 35th Conference on Neural Information Processing Systems (NeurIPS 2021). Do not distribute.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 113, + 97, + 498, + 137 + ], + "lines": [ + { + "bbox": [ + 112, + 97, + 498, + 118 + ], + "spans": [ + { + "bbox": [ + 112, + 97, + 498, + 118 + ], + "score": 1.0, + "content": "Direct then Diffuse: Incremental Unsupervised Skill", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 116, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 125, + 116, + 486, + 140 + ], + "score": 1.0, + "content": "Discovery for State Covering and Goal Reaching", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 259, + 179, + 354, + 223 + ], + "lines": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "spans": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "score": 1.0, + "content": "Anonymous Author(s)", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "spans": [ + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "score": 1.0, + "content": "Affiliation", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "spans": [ + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "score": 1.0, + "content": "Address", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "spans": [ + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "score": 1.0, + "content": "email", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + } + ], + "index": 3.5, + "bbox_fs": [ + 258, + 178, + 356, + 222 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 252, + 328, + 265 + ], + "lines": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "spans": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "index", + "bbox": [ + 91, + 275, + 469, + 483 + ], + "lines": [ + { + "bbox": [ + 93, + 275, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 93, + 279, + 99, + 285 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 141, + 275, + 469, + 288 + ], + "score": 1.0, + "content": "Learning meaningful behaviors in the absence of a task-specific reward function", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 286, + 470, + 299 + ], + "spans": [ + { + "bbox": [ + 93, + 289, + 99, + 297 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 141, + 286, + 470, + 299 + ], + "score": 1.0, + "content": "is a challenging problem in reinforcement learning. A desirable unsupervised", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 297, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 93, + 300, + 99, + 308 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 141, + 297, + 470, + 309 + ], + "score": 1.0, + "content": "objective is to learn a set of diverse skills that provide a thorough coverage of", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 92, + 311, + 99, + 320 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 142, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "the state space while being directed, i.e., reliably reaching distinct regions of the", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 318, + 470, + 332 + ], + "spans": [ + { + "bbox": [ + 93, + 321, + 100, + 330 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 141, + 318, + 470, + 332 + ], + "score": 1.0, + "content": "environment. At test time, an agent could then leverage these skills to solve sparse", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 330, + 470, + 342 + ], + "spans": [ + { + "bbox": [ + 93, + 333, + 99, + 341 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 141, + 330, + 470, + 342 + ], + "score": 1.0, + "content": "reward problems by performing efficient exploration and finding an effective goal-", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 339, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 92, + 343, + 99, + 352 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 141, + 339, + 470, + 354 + ], + "score": 1.0, + "content": "directed policy with little-to-no additional learning. Unfortunately, it is challenging", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 351, + 470, + 365 + ], + "spans": [ + { + "bbox": [ + 92, + 354, + 99, + 363 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 141, + 351, + 470, + 365 + ], + "score": 1.0, + "content": "to learn skills with such properties, as diffusing (e.g., stochastic policies performing", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 362, + 470, + 376 + ], + "spans": [ + { + "bbox": [ + 92, + 365, + 100, + 374 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 141, + 362, + 470, + 376 + ], + "score": 1.0, + "content": "good coverage) skills are not reliable in targeting specific states, whereas directed", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 374, + 470, + 387 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 99, + 385 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 141, + 374, + 470, + 387 + ], + "score": 1.0, + "content": "(e.g., goal-based policies) skills provide limited coverage. In this paper, inspired", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 384, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 90, + 387, + 99, + 396 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 141, + 384, + 470, + 397 + ], + "score": 1.0, + "content": "by the mutual information framework, we propose a novel algorithm designed", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 395, + 470, + 408 + ], + "spans": [ + { + "bbox": [ + 89, + 397, + 100, + 407 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 141, + 395, + 470, + 408 + ], + "score": 1.0, + "content": "to maximize coverage while ensuring a constraint on the directedness of each", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 407, + 470, + 418 + ], + "spans": [ + { + "bbox": [ + 90, + 408, + 99, + 418 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 141, + 407, + 470, + 418 + ], + "score": 1.0, + "content": "skill. In particular, we design skills with a decoupled policy structure, with a first", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 416, + 470, + 430 + ], + "spans": [ + { + "bbox": [ + 89, + 419, + 100, + 429 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 141, + 416, + 470, + 430 + ], + "score": 1.0, + "content": "part trained to be directed and a second diffusing part that ensures local coverage.", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 428, + 469, + 441 + ], + "spans": [ + { + "bbox": [ + 89, + 430, + 100, + 440 + ], + "score": 1.0, + "content": "15", + "type": "text" + }, + { + "bbox": [ + 141, + 428, + 469, + 441 + ], + "score": 1.0, + "content": "Furthermore, we leverage the directedness constraint to adaptively add or remove", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 438, + 470, + 452 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 100, + 451 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 141, + 438, + 470, + 452 + ], + "score": 1.0, + "content": "skills as well as incrementally compose them along a tree that is grown to achieve a", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 449, + 470, + 462 + ], + "spans": [ + { + "bbox": [ + 89, + 452, + 100, + 462 + ], + "score": 1.0, + "content": "17", + "type": "text" + }, + { + "bbox": [ + 141, + 449, + 470, + 462 + ], + "score": 1.0, + "content": "thorough coverage of the environment. We illustrate how our learned skills enables", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 461, + 469, + 473 + ], + "spans": [ + { + "bbox": [ + 90, + 463, + 99, + 472 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 142, + 461, + 469, + 473 + ], + "score": 1.0, + "content": "to efficiently solve sparse-reward downstream tasks in navigation and continuous", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 472, + 443, + 484 + ], + "spans": [ + { + "bbox": [ + 90, + 474, + 99, + 483 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 142, + 472, + 443, + 484 + ], + "score": 1.0, + "content": "control environments, where it compares favorably with existing baselines.", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + } + ], + "index": 16, + "bbox_fs": [ + 89, + 275, + 470, + 484 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 494, + 191, + 508 + ], + "lines": [ + { + "bbox": [ + 86, + 493, + 192, + 510 + ], + "spans": [ + { + "bbox": [ + 86, + 493, + 192, + 510 + ], + "score": 1.0, + "content": "20 1 Introduction", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "index", + "bbox": [ + 90, + 515, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 89, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 89, + 516, + 99, + 526 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) algorithms have been shown to effectively solve a wide variety of", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 525, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 528, + 100, + 538 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 105, + 525, + 506, + 539 + ], + "score": 1.0, + "content": "complex problems [e.g., 23, 6, 31, 12, 2, 28]. However, they are often designed to solve one single", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 536, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 89, + 539, + 100, + 549 + ], + "score": 1.0, + "content": "23", + "type": "text" + }, + { + "bbox": [ + 105, + 536, + 506, + 550 + ], + "score": 1.0, + "content": "task at a time and they need to restart the learning process from scratch for any new problem, even", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 88, + 550, + 100, + 559 + ], + "score": 1.0, + "content": "24", + "type": "text" + }, + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "when it is defined on the very same environment (e.g., navigating to different locations in the same", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 89, + 561, + 100, + 571 + ], + "score": 1.0, + "content": "25", + "type": "text" + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "apartment). Recently, unsupervised RL (URL) has been proposed as an approach to address this", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 570, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 88, + 572, + 100, + 581 + ], + "score": 1.0, + "content": "26", + "type": "text" + }, + { + "bbox": [ + 106, + 570, + 506, + 582 + ], + "score": 1.0, + "content": "limitation. In URL, the agent first interacts with the environment without any extrinsic reward signal.", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 581, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 89, + 583, + 100, + 592 + ], + "score": 1.0, + "content": "27", + "type": "text" + }, + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "score": 1.0, + "content": "Afterward, the agent leverages the experience accumulated during the unsupervised learning phase to", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 592, + 430, + 604 + ], + "spans": [ + { + "bbox": [ + 89, + 594, + 100, + 603 + ], + "score": 1.0, + "content": "28", + "type": "text" + }, + { + "bbox": [ + 105, + 592, + 430, + 604 + ], + "score": 1.0, + "content": "efficiently solve a variety of downstream tasks defined on the same environment.", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 89, + 608, + 100, + 618 + ], + "score": 1.0, + "content": "29", + "type": "text" + }, + { + "bbox": [ + 105, + 606, + 443, + 619 + ], + "score": 1.0, + "content": "In this paper, we consider the URL setting where the agent starts from an initial state", + "type": "text" + }, + { + "bbox": [ + 444, + 608, + 454, + 618 + ], + "score": 0.86, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "and it resets", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 616, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 619, + 100, + 629 + ], + "score": 1.0, + "content": "30", + "type": "text" + }, + { + "bbox": [ + 105, + 616, + 506, + 631 + ], + "score": 1.0, + "content": "to it every time the policy terminates. We focus on sparse-reward downstream tasks, which require", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 89, + 630, + 100, + 640 + ], + "score": 1.0, + "content": "31", + "type": "text" + }, + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "effective exploration (i.e., via a thorough coverage of the state space) to find the goal as well as", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 639, + 374, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 641, + 100, + 651 + ], + "score": 1.0, + "content": "32", + "type": "text" + }, + { + "bbox": [ + 105, + 639, + 374, + 653 + ], + "score": 1.0, + "content": "learning a policy reliably reaching the goal (i.e., a directed policy).", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 90, + 656, + 100, + 664 + ], + "score": 1.0, + "content": "33", + "type": "text" + }, + { + "bbox": [ + 105, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "We build on the insight that mutual information (MI) effectively formalizes the dual objective of", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 89, + 667, + 100, + 676 + ], + "score": 1.0, + "content": "34", + "type": "text" + }, + { + "bbox": [ + 104, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "learning skills that both cover and navigate the environment efficiently [e.g., 11]. Specifically, given", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 89, + 678, + 100, + 687 + ], + "score": 1.0, + "content": "35", + "type": "text" + }, + { + "bbox": [ + 105, + 676, + 174, + 688 + ], + "score": 1.0, + "content": "the state variable", + "type": "text" + }, + { + "bbox": [ + 174, + 676, + 182, + 686 + ], + "score": 0.84, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 676, + 260, + 688 + ], + "score": 1.0, + "content": "and some variables", + "type": "text" + }, + { + "bbox": [ + 261, + 676, + 269, + 686 + ], + "score": 0.82, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "on which the skill policies are conditioned, MI is defined as", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + } + ], + "index": 30.5, + "bbox_fs": [ + 88, + 515, + 506, + 604 + ] + }, + { + "type": "index", + "bbox": [ + 90, + 606, + 505, + 651 + ], + "lines": [], + "index": 36.5, + "bbox_fs": [ + 89, + 606, + 506, + 653 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 90, + 654, + 505, + 687 + ], + "lines": [], + "index": 40, + "bbox_fs": [ + 89, + 654, + 506, + 688 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 689, + 418, + 704 + ], + "lines": [ + { + "bbox": [ + 191, + 689, + 418, + 704 + ], + "spans": [ + { + "bbox": [ + 191, + 689, + 418, + 704 + ], + "score": 0.8, + "content": "\\begin{array} { r } { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\\\ { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\end{array} - \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } | Z ) = \\boldsymbol { \\mathcal { H } } ( Z ) - \\boldsymbol { \\mathcal { H } } ( Z | \\boldsymbol { S } ) ,", + "type": "interline_equation", + "image_path": "c3bfc0f38cdfab89e1181d0c4924a0c8dc8268d8d34cb919b94fb7bd9ae886a6.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 191, + 689, + 418, + 704 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 103, + 731, + 499, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 500, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 500, + 743 + ], + "score": 1.0, + "content": "Submitted to 35th Conference on Neural Information Processing Systems (NeurIPS 2021). Do not distribute.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 730, + 500, + 743 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 74, + 483, + 185 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 74, + 483, + 185 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 74, + 483, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 74, + 483, + 185 + ], + "score": 0.975, + "type": "image", + "image_path": "f3a6a7f8b3e7aa8d82fa62e9127f2a972085e7545768b2da5305611566b631fe.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 74, + 483, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 111.0, + 483, + 148.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 148.0, + 483, + 185.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 193, + 506, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 393, + 205 + ], + "score": 1.0, + "content": "Figure 1: Overview of UPSIDE. The black dot corresponds to the initial state", + "type": "text" + }, + { + "bbox": [ + 393, + 195, + 403, + 204 + ], + "score": 0.76, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 193, + 505, + 205 + ], + "score": 1.0, + "content": ". (A) A set of random skills", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "is initialized, each skill being composed of a directed part (illustrated as a black arrow) and a diffusing part", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 216, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 334, + 227 + ], + "score": 1.0, + "content": "(red arrows), which induces a local coverage (colored circles).", + "type": "text" + }, + { + "bbox": [ + 335, + 216, + 347, + 226 + ], + "score": 0.51, + "content": "( B )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 216, + 505, + 227 + ], + "score": 1.0, + "content": "The policies associated to the directed part", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 104, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "of each skill are then updated to maximize the discriminability of the states reached by their diffusing part", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 238, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 147, + 249 + ], + "score": 1.0, + "content": "(Sect. 3.1).", + "type": "text" + }, + { + "bbox": [ + 148, + 238, + 160, + 248 + ], + "score": 0.32, + "content": "( C )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 238, + 505, + 249 + ], + "score": 1.0, + "content": "The least discriminable skills are iteratively removed while the policies of the remaining skills are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "score": 1.0, + "content": "re-optimized. This is executed until the discriminability of each skill satisfies a given constraint (see Sect. 3.2).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 260, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 237, + 270 + ], + "score": 1.0, + "content": "In this example three skills are kept.", + "type": "text" + }, + { + "bbox": [ + 237, + 260, + 250, + 270 + ], + "score": 0.49, + "content": "( D )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 260, + 505, + 270 + ], + "score": 1.0, + "content": "One of these learned skill is then used as basis to add new skills, which", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 270, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 281 + ], + "score": 1.0, + "content": "are then optimized following the same procedure. For the “red” and “purple” skills, UPSIDE is not able to find", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 282, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 368, + 292 + ], + "score": 1.0, + "content": "sub-skills of sufficient quality and thus they are not expanded any further.", + "type": "text" + }, + { + "bbox": [ + 369, + 282, + 381, + 291 + ], + "score": 0.64, + "content": "( E )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 282, + 505, + 292 + ], + "score": 1.0, + "content": "At the end of the process, UPSIDE", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "has created a tree of directed skills covering the state space (Sect. 3.3). These covering skills can then be used to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "solve downstream tasks. Moreover, the discriminator learned together with the skills can be used to select the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "skill to reach any specific goal region, where the directed parts get close to the goal, while the diffusing part", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 324, + 497, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 497, + 337 + ], + "score": 1.0, + "content": "provides the local coverage to attain the goal. The complete algorithm is detailed in Sect. 3.4 and Appendix.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9 + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 90, + 343, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 89, + 344, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 89, + 345, + 100, + 355 + ], + "score": 1.0, + "content": "36", + "type": "text" + }, + { + "bbox": [ + 105, + 344, + 133, + 355 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 344, + 141, + 353 + ], + "score": 0.74, + "content": "\\mathcal { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 344, + 222, + 355 + ], + "score": 1.0, + "content": "denotes the MI and", + "type": "text" + }, + { + "bbox": [ + 223, + 344, + 232, + 354 + ], + "score": 0.8, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 344, + 505, + 355 + ], + "score": 1.0, + "content": "is the entropy function. The first expression, known as the forward", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 353, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 89, + 357, + 99, + 366 + ], + "score": 1.0, + "content": "37", + "type": "text" + }, + { + "bbox": [ + 105, + 353, + 505, + 369 + ], + "score": 1.0, + "content": "form of MI, explicitly balances the two sought-after properties of coverage — captured by the entropy", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 364, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 89, + 367, + 100, + 377 + ], + "score": 1.0, + "content": "38", + "type": "text" + }, + { + "bbox": [ + 104, + 364, + 187, + 379 + ], + "score": 1.0, + "content": "over the state space", + "type": "text" + }, + { + "bbox": [ + 188, + 366, + 212, + 378 + ], + "score": 0.9, + "content": "\\mathcal { H } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 364, + 451, + 379 + ], + "score": 1.0, + "content": "— and directedness, i.e., the ability to reach specific states", + "type": "text" + }, + { + "bbox": [ + 451, + 366, + 459, + 376 + ], + "score": 0.76, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 364, + 506, + 379 + ], + "score": 1.0, + "content": "depending", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 89, + 379, + 100, + 388 + ], + "score": 1.0, + "content": "39", + "type": "text" + }, + { + "bbox": [ + 105, + 377, + 119, + 388 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 376, + 128, + 387 + ], + "score": 0.74, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 377, + 318, + 388 + ], + "score": 1.0, + "content": "— captured by the negative conditional entropy", + "type": "text" + }, + { + "bbox": [ + 318, + 377, + 359, + 388 + ], + "score": 0.92, + "content": "- { \\mathcal { H } } ( S | Z )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 377, + 505, + 388 + ], + "score": 1.0, + "content": ". The second expression of (1), often", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 89, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 89, + 390, + 100, + 399 + ], + "score": 1.0, + "content": "40", + "type": "text" + }, + { + "bbox": [ + 104, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "easier to optimize and referred to as the reverse form, stipulates that the skills should be sampled as", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 399, + 299, + 410 + ], + "spans": [ + { + "bbox": [ + 89, + 401, + 99, + 410 + ], + "score": 1.0, + "content": "41", + "type": "text" + }, + { + "bbox": [ + 105, + 399, + 299, + 410 + ], + "score": 1.0, + "content": "diversely as possible while being discriminable.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 90, + 413, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 89, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 89, + 415, + 100, + 424 + ], + "score": 1.0, + "content": "42", + "type": "text" + }, + { + "bbox": [ + 105, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "Maximizing (1) has been shown to be a powerful approach for encouraging exploration in RL [16, 25]", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 89, + 426, + 100, + 435 + ], + "score": 1.0, + "content": "43", + "type": "text" + }, + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "and for unsupervised skill discovery [e.g., 11, 9, 1, 30, 8]. Nonetheless, learning skills that maximize", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 89, + 437, + 100, + 446 + ], + "score": 1.0, + "content": "44", + "type": "text" + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "the MI is a challenging optimization problem. Several approximations have been proposed to simplify", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 89, + 448, + 100, + 457 + ], + "score": 1.0, + "content": "45", + "type": "text" + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "the problem at the cost of possibly deviating from the original objective of coverage and directedness", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 89, + 459, + 100, + 468 + ], + "score": 1.0, + "content": "46", + "type": "text" + }, + { + "bbox": [ + 106, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "(see Sect. 4 for a review of related work). In this paper, we propose UPSIDE (UnsuPervised Skills", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 89, + 470, + 100, + 479 + ], + "score": 1.0, + "content": "47", + "type": "text" + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "that dIrect then DiffusE) to learn skills that can be effectively used to solve goal-based downstream", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 89, + 479, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 89, + 481, + 100, + 489 + ], + "score": 1.0, + "content": "48", + "type": "text" + }, + { + "bbox": [ + 105, + 479, + 497, + 491 + ], + "score": 1.0, + "content": "tasks. Our solution builds on the following components (see Fig. 1 for an illustration of UPSIDE):", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 100, + 493, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 108, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 108, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "• Skill structure. In order to balance coverage and directedness, we design skills composed of two", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 117, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "parts: 1) a directed part that is trained to reach a distinct region of the environment, and 2) a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 515, + 429, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 429, + 528 + ], + "score": 1.0, + "content": "diffusing part that covers the states around the region attained by the first part.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 114, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "Optimization. We further strengthen the coverage and directedness properties of the skills by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "turning the MI objective into a constrained optimization problem designed to maximize coverage", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 117, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "under the constraint that each skill achieves a minimum level of discriminability. This in turn", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 117, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 117, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "enables UPSIDE to adaptively add skills to improve coverage, when all the initial skills meet the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "constraint, or remove those that violate the constraint to guarantee that each skill is directed and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 582, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 116, + 582, + 296, + 594 + ], + "score": 1.0, + "content": "reaches a distinct region of the environment.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 109, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 109, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "• Tree structure. When the agent starts from a fixed initial state, the skills’ length is a crucial", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "score": 1.0, + "content": "parameter, where short skills do not allow for proper coverage, and long skills are difficult to train.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 614, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 116, + 614, + 506, + 629 + ], + "score": 1.0, + "content": "In UPSIDE we consider short skills to make the optimization easier, while composing them along a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 626, + 426, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 426, + 640 + ], + "score": 1.0, + "content": "tree structure that ensures an adaptive and deep coverage of the environment.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 91, + 641, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 90, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 90, + 644, + 99, + 653 + ], + "score": 1.0, + "content": "62", + "type": "text" + }, + { + "bbox": [ + 105, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "We study how our learned skill structure enables to both perform efficient exploration and learn", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 90, + 653, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 90, + 655, + 100, + 663 + ], + "score": 1.0, + "content": "63", + "type": "text" + }, + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "score": 1.0, + "content": "effective goal-reaching policies in a variety of navigation and continuous control environments", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 90, + 663, + 447, + 676 + ], + "spans": [ + { + "bbox": [ + 90, + 666, + 100, + 675 + ], + "score": 1.0, + "content": "64", + "type": "text" + }, + { + "bbox": [ + 105, + 663, + 447, + 676 + ], + "score": 1.0, + "content": "(including MuJoCo’s reacher) and we compare its performance to relevant baselines.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 101, + 682, + 162, + 696 + ], + "lines": [ + { + "bbox": [ + 104, + 680, + 164, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 680, + 164, + 700 + ], + "score": 1.0, + "content": "2 Setting", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 91, + 700, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 89, + 700, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 89, + 702, + 100, + 711 + ], + "score": 1.0, + "content": "66", + "type": "text" + }, + { + "bbox": [ + 105, + 700, + 491, + 711 + ], + "score": 1.0, + "content": "We consider the URL setting where the agent interacts with a Markov decision process (MDP)", + "type": "text" + }, + { + "bbox": [ + 492, + 700, + 504, + 710 + ], + "score": 0.52, + "content": "M", + "type": "inline_equation" + } + ], + "index": 46 + }, + { + "bbox": [ + 89, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 89, + 711, + 175, + 723 + ], + "score": 1.0, + "content": "with state space 67", + "type": "text" + }, + { + "bbox": [ + 176, + 711, + 183, + 721 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 711, + 243, + 723 + ], + "score": 1.0, + "content": ", action space", + "type": "text" + }, + { + "bbox": [ + 243, + 711, + 252, + 721 + ], + "score": 0.78, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 711, + 299, + 723 + ], + "score": 1.0, + "content": ", dynamics", + "type": "text" + }, + { + "bbox": [ + 300, + 711, + 338, + 723 + ], + "score": 0.94, + "content": "p ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 711, + 505, + 723 + ], + "score": 1.0, + "content": ", and no reward. The agent starts each", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 74, + 483, + 185 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 74, + 483, + 185 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 74, + 483, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 74, + 483, + 185 + ], + "score": 0.975, + "type": "image", + "image_path": "f3a6a7f8b3e7aa8d82fa62e9127f2a972085e7545768b2da5305611566b631fe.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 74, + 483, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 111.0, + 483, + 148.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 148.0, + 483, + 185.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 193, + 506, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 393, + 205 + ], + "score": 1.0, + "content": "Figure 1: Overview of UPSIDE. The black dot corresponds to the initial state", + "type": "text" + }, + { + "bbox": [ + 393, + 195, + 403, + 204 + ], + "score": 0.76, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 193, + 505, + 205 + ], + "score": 1.0, + "content": ". (A) A set of random skills", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "is initialized, each skill being composed of a directed part (illustrated as a black arrow) and a diffusing part", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 216, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 334, + 227 + ], + "score": 1.0, + "content": "(red arrows), which induces a local coverage (colored circles).", + "type": "text" + }, + { + "bbox": [ + 335, + 216, + 347, + 226 + ], + "score": 0.51, + "content": "( B )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 216, + 505, + 227 + ], + "score": 1.0, + "content": "The policies associated to the directed part", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 104, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "of each skill are then updated to maximize the discriminability of the states reached by their diffusing part", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 238, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 147, + 249 + ], + "score": 1.0, + "content": "(Sect. 3.1).", + "type": "text" + }, + { + "bbox": [ + 148, + 238, + 160, + 248 + ], + "score": 0.32, + "content": "( C )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 238, + 505, + 249 + ], + "score": 1.0, + "content": "The least discriminable skills are iteratively removed while the policies of the remaining skills are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "score": 1.0, + "content": "re-optimized. This is executed until the discriminability of each skill satisfies a given constraint (see Sect. 3.2).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 260, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 237, + 270 + ], + "score": 1.0, + "content": "In this example three skills are kept.", + "type": "text" + }, + { + "bbox": [ + 237, + 260, + 250, + 270 + ], + "score": 0.49, + "content": "( D )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 260, + 505, + 270 + ], + "score": 1.0, + "content": "One of these learned skill is then used as basis to add new skills, which", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 270, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 281 + ], + "score": 1.0, + "content": "are then optimized following the same procedure. For the “red” and “purple” skills, UPSIDE is not able to find", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 282, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 368, + 292 + ], + "score": 1.0, + "content": "sub-skills of sufficient quality and thus they are not expanded any further.", + "type": "text" + }, + { + "bbox": [ + 369, + 282, + 381, + 291 + ], + "score": 0.64, + "content": "( E )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 282, + 505, + 292 + ], + "score": 1.0, + "content": "At the end of the process, UPSIDE", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "has created a tree of directed skills covering the state space (Sect. 3.3). These covering skills can then be used to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "solve downstream tasks. Moreover, the discriminator learned together with the skills can be used to select the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "skill to reach any specific goal region, where the directed parts get close to the goal, while the diffusing part", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 324, + 497, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 497, + 337 + ], + "score": 1.0, + "content": "provides the local coverage to attain the goal. The complete algorithm is detailed in Sect. 3.4 and Appendix.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9 + } + ], + "index": 5.0 + }, + { + "type": "index", + "bbox": [ + 90, + 343, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 89, + 344, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 89, + 345, + 100, + 355 + ], + "score": 1.0, + "content": "36", + "type": "text" + }, + { + "bbox": [ + 105, + 344, + 133, + 355 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 344, + 141, + 353 + ], + "score": 0.74, + "content": "\\mathcal { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 344, + 222, + 355 + ], + "score": 1.0, + "content": "denotes the MI and", + "type": "text" + }, + { + "bbox": [ + 223, + 344, + 232, + 354 + ], + "score": 0.8, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 344, + 505, + 355 + ], + "score": 1.0, + "content": "is the entropy function. The first expression, known as the forward", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 353, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 89, + 357, + 99, + 366 + ], + "score": 1.0, + "content": "37", + "type": "text" + }, + { + "bbox": [ + 105, + 353, + 505, + 369 + ], + "score": 1.0, + "content": "form of MI, explicitly balances the two sought-after properties of coverage — captured by the entropy", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 364, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 89, + 367, + 100, + 377 + ], + "score": 1.0, + "content": "38", + "type": "text" + }, + { + "bbox": [ + 104, + 364, + 187, + 379 + ], + "score": 1.0, + "content": "over the state space", + "type": "text" + }, + { + "bbox": [ + 188, + 366, + 212, + 378 + ], + "score": 0.9, + "content": "\\mathcal { H } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 364, + 451, + 379 + ], + "score": 1.0, + "content": "— and directedness, i.e., the ability to reach specific states", + "type": "text" + }, + { + "bbox": [ + 451, + 366, + 459, + 376 + ], + "score": 0.76, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 364, + 506, + 379 + ], + "score": 1.0, + "content": "depending", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 89, + 379, + 100, + 388 + ], + "score": 1.0, + "content": "39", + "type": "text" + }, + { + "bbox": [ + 105, + 377, + 119, + 388 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 376, + 128, + 387 + ], + "score": 0.74, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 377, + 318, + 388 + ], + "score": 1.0, + "content": "— captured by the negative conditional entropy", + "type": "text" + }, + { + "bbox": [ + 318, + 377, + 359, + 388 + ], + "score": 0.92, + "content": "- { \\mathcal { H } } ( S | Z )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 377, + 505, + 388 + ], + "score": 1.0, + "content": ". The second expression of (1), often", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 89, + 390, + 100, + 399 + ], + "score": 1.0, + "content": "40", + "type": "text" + }, + { + "bbox": [ + 104, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "easier to optimize and referred to as the reverse form, stipulates that the skills should be sampled as", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 399, + 299, + 410 + ], + "spans": [ + { + "bbox": [ + 89, + 401, + 99, + 410 + ], + "score": 1.0, + "content": "41", + "type": "text" + }, + { + "bbox": [ + 105, + 399, + 299, + 410 + ], + "score": 1.0, + "content": "diversely as possible while being discriminable.", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 89, + 415, + 100, + 424 + ], + "score": 1.0, + "content": "42", + "type": "text" + }, + { + "bbox": [ + 105, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "Maximizing (1) has been shown to be a powerful approach for encouraging exploration in RL [16, 25]", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 89, + 426, + 100, + 435 + ], + "score": 1.0, + "content": "43", + "type": "text" + }, + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "and for unsupervised skill discovery [e.g., 11, 9, 1, 30, 8]. Nonetheless, learning skills that maximize", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 89, + 437, + 100, + 446 + ], + "score": 1.0, + "content": "44", + "type": "text" + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "the MI is a challenging optimization problem. Several approximations have been proposed to simplify", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 89, + 448, + 100, + 457 + ], + "score": 1.0, + "content": "45", + "type": "text" + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "the problem at the cost of possibly deviating from the original objective of coverage and directedness", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 89, + 459, + 100, + 468 + ], + "score": 1.0, + "content": "46", + "type": "text" + }, + { + "bbox": [ + 106, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "(see Sect. 4 for a review of related work). In this paper, we propose UPSIDE (UnsuPervised Skills", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 89, + 470, + 100, + 479 + ], + "score": 1.0, + "content": "47", + "type": "text" + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "that dIrect then DiffusE) to learn skills that can be effectively used to solve goal-based downstream", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 479, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 89, + 481, + 100, + 489 + ], + "score": 1.0, + "content": "48", + "type": "text" + }, + { + "bbox": [ + 105, + 479, + 497, + 491 + ], + "score": 1.0, + "content": "tasks. Our solution builds on the following components (see Fig. 1 for an illustration of UPSIDE):", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + } + ], + "index": 18.5, + "bbox_fs": [ + 89, + 344, + 506, + 410 + ] + }, + { + "type": "index", + "bbox": [ + 90, + 413, + 505, + 490 + ], + "lines": [], + "index": 25, + "bbox_fs": [ + 89, + 413, + 506, + 491 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 100, + 493, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 108, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 108, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "• Skill structure. In order to balance coverage and directedness, we design skills composed of two", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 117, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "parts: 1) a directed part that is trained to reach a distinct region of the environment, and 2) a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 515, + 429, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 429, + 528 + ], + "score": 1.0, + "content": "diffusing part that covers the states around the region attained by the first part.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 114, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "Optimization. We further strengthen the coverage and directedness properties of the skills by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "turning the MI objective into a constrained optimization problem designed to maximize coverage", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 117, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "under the constraint that each skill achieves a minimum level of discriminability. This in turn", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 117, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 117, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "enables UPSIDE to adaptively add skills to improve coverage, when all the initial skills meet the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 116, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "constraint, or remove those that violate the constraint to guarantee that each skill is directed and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 582, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 116, + 582, + 296, + 594 + ], + "score": 1.0, + "content": "reaches a distinct region of the environment.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 109, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 109, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "• Tree structure. When the agent starts from a fixed initial state, the skills’ length is a crucial", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "score": 1.0, + "content": "parameter, where short skills do not allow for proper coverage, and long skills are difficult to train.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 614, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 116, + 614, + 506, + 629 + ], + "score": 1.0, + "content": "In UPSIDE we consider short skills to make the optimization easier, while composing them along a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 626, + 426, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 426, + 640 + ], + "score": 1.0, + "content": "tree structure that ensures an adaptive and deep coverage of the environment.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35, + "bbox_fs": [ + 108, + 493, + 507, + 640 + ] + }, + { + "type": "index", + "bbox": [ + 91, + 641, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 90, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 90, + 644, + 99, + 653 + ], + "score": 1.0, + "content": "62", + "type": "text" + }, + { + "bbox": [ + 105, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "We study how our learned skill structure enables to both perform efficient exploration and learn", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 653, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 90, + 655, + 100, + 663 + ], + "score": 1.0, + "content": "63", + "type": "text" + }, + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "score": 1.0, + "content": "effective goal-reaching policies in a variety of navigation and continuous control environments", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 663, + 447, + 676 + ], + "spans": [ + { + "bbox": [ + 90, + 666, + 100, + 675 + ], + "score": 1.0, + "content": "64", + "type": "text" + }, + { + "bbox": [ + 105, + 663, + 447, + 676 + ], + "score": 1.0, + "content": "(including MuJoCo’s reacher) and we compare its performance to relevant baselines.", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + } + ], + "index": 43, + "bbox_fs": [ + 90, + 641, + 506, + 676 + ] + }, + { + "type": "title", + "bbox": [ + 101, + 682, + 162, + 696 + ], + "lines": [ + { + "bbox": [ + 104, + 680, + 164, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 680, + 164, + 700 + ], + "score": 1.0, + "content": "2 Setting", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 91, + 700, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 89, + 700, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 89, + 702, + 100, + 711 + ], + "score": 1.0, + "content": "66", + "type": "text" + }, + { + "bbox": [ + 105, + 700, + 491, + 711 + ], + "score": 1.0, + "content": "We consider the URL setting where the agent interacts with a Markov decision process (MDP)", + "type": "text" + }, + { + "bbox": [ + 492, + 700, + 504, + 710 + ], + "score": 0.52, + "content": "M", + "type": "inline_equation" + } + ], + "index": 46 + }, + { + "bbox": [ + 89, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 89, + 711, + 175, + 723 + ], + "score": 1.0, + "content": "with state space 67", + "type": "text" + }, + { + "bbox": [ + 176, + 711, + 183, + 721 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 711, + 243, + 723 + ], + "score": 1.0, + "content": ", action space", + "type": "text" + }, + { + "bbox": [ + 243, + 711, + 252, + 721 + ], + "score": 0.78, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 711, + 299, + 723 + ], + "score": 1.0, + "content": ", dynamics", + "type": "text" + }, + { + "bbox": [ + 300, + 711, + 338, + 723 + ], + "score": 0.94, + "content": "p ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 711, + 505, + 723 + ], + "score": 1.0, + "content": ", and no reward. The agent starts each", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 89, + 700, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 93, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 92, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 92, + 76, + 100, + 83 + ], + "score": 1.0, + "content": "68", + "type": "text" + }, + { + "bbox": [ + 104, + 72, + 262, + 86 + ], + "score": 1.0, + "content": "episode from a designated initial state", + "type": "text" + }, + { + "bbox": [ + 262, + 73, + 291, + 84 + ], + "score": 0.93, + "content": "s _ { 0 } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 72, + 505, + 86 + ], + "score": 1.0, + "content": ". Upon termination of the chosen policy, the agent is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 92, + 84, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 92, + 87, + 99, + 94 + ], + "score": 1.0, + "content": "69", + "type": "text" + }, + { + "bbox": [ + 105, + 84, + 158, + 96 + ], + "score": 1.0, + "content": "then reset to", + "type": "text" + }, + { + "bbox": [ + 158, + 85, + 168, + 95 + ], + "score": 0.84, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 84, + 505, + 96 + ], + "score": 1.0, + "content": ". This setting is particularly challenging from an exploration point of view since the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 92, + 94, + 377, + 108 + ], + "spans": [ + { + "bbox": [ + 92, + 98, + 99, + 105 + ], + "score": 1.0, + "content": "70", + "type": "text" + }, + { + "bbox": [ + 105, + 94, + 377, + 108 + ], + "score": 1.0, + "content": "agent cannot rely on the initial distribution to cover the state space.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 89, + 110, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 89, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 89, + 113, + 99, + 122 + ], + "score": 1.0, + "content": "71", + "type": "text" + }, + { + "bbox": [ + 105, + 110, + 470, + 124 + ], + "score": 1.0, + "content": "We recall the MI-based unsupervised skill discovery approach [see e.g., 11]. Denote by", + "type": "text" + }, + { + "bbox": [ + 471, + 111, + 479, + 121 + ], + "score": 0.82, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "some", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 89, + 124, + 100, + 133 + ], + "score": 1.0, + "content": "72", + "type": "text" + }, + { + "bbox": [ + 105, + 121, + 302, + 135 + ], + "score": 1.0, + "content": "(latent) variables on which the skills of length", + "type": "text" + }, + { + "bbox": [ + 302, + 122, + 311, + 132 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "are conditioned. There are three optimization", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 89, + 135, + 99, + 144 + ], + "score": 1.0, + "content": "73", + "type": "text" + }, + { + "bbox": [ + 105, + 132, + 150, + 145 + ], + "score": 1.0, + "content": "variables:", + "type": "text" + }, + { + "bbox": [ + 150, + 133, + 161, + 144 + ], + "score": 0.41, + "content": "( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 132, + 312, + 145 + ], + "score": 1.0, + "content": "the support of the skills denoted by", + "type": "text" + }, + { + "bbox": [ + 313, + 133, + 327, + 144 + ], + "score": 0.91, + "content": "| Z |", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 132, + 464, + 145 + ], + "score": 1.0, + "content": "(we consider it to be discrete so", + "type": "text" + }, + { + "bbox": [ + 464, + 133, + 478, + 144 + ], + "score": 0.9, + "content": "| Z |", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 142, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 89, + 146, + 100, + 155 + ], + "score": 1.0, + "content": "74", + "type": "text" + }, + { + "bbox": [ + 105, + 142, + 181, + 157 + ], + "score": 1.0, + "content": "number of skills),", + "type": "text" + }, + { + "bbox": [ + 182, + 144, + 195, + 155 + ], + "score": 0.55, + "content": "( i i )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 142, + 240, + 157 + ], + "score": 1.0, + "content": "the policy", + "type": "text" + }, + { + "bbox": [ + 240, + 144, + 260, + 155 + ], + "score": 0.93, + "content": "\\pi ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 142, + 337, + 157 + ], + "score": 1.0, + "content": "associated to skill", + "type": "text" + }, + { + "bbox": [ + 337, + 146, + 343, + 154 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 142, + 365, + 157 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 365, + 144, + 381, + 155 + ], + "score": 0.53, + "content": "( i i i )", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 142, + 456, + 157 + ], + "score": 1.0, + "content": "the sampling rule", + "type": "text" + }, + { + "bbox": [ + 456, + 146, + 463, + 155 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 142, + 486, + 157 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 486, + 144, + 505, + 156 + ], + "score": 0.89, + "content": "\\rho ( z )", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 89, + 156, + 99, + 166 + ], + "score": 1.0, + "content": "75", + "type": "text" + }, + { + "bbox": [ + 105, + 154, + 250, + 167 + ], + "score": 1.0, + "content": "is the probability of sampling skill", + "type": "text" + }, + { + "bbox": [ + 251, + 156, + 257, + 164 + ], + "score": 0.8, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 154, + 462, + 167 + ], + "score": 1.0, + "content": "at the beginning of the episode). Let the variable", + "type": "text" + }, + { + "bbox": [ + 462, + 155, + 476, + 165 + ], + "score": 0.89, + "content": "S _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "be the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 89, + 167, + 99, + 177 + ], + "score": 1.0, + "content": "76", + "type": "text" + }, + { + "bbox": [ + 105, + 164, + 302, + 178 + ], + "score": 1.0, + "content": "random (final) state induced by sampling a skill", + "type": "text" + }, + { + "bbox": [ + 302, + 169, + 308, + 175 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 164, + 331, + 178 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 332, + 167, + 338, + 177 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 164, + 485, + 178 + ], + "score": 1.0, + "content": "and executing the associated policy", + "type": "text" + }, + { + "bbox": [ + 485, + 166, + 505, + 177 + ], + "score": 0.91, + "content": "\\pi ( z )", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 89, + 178, + 99, + 188 + ], + "score": 1.0, + "content": "77", + "type": "text" + }, + { + "bbox": [ + 105, + 176, + 129, + 190 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 130, + 178, + 140, + 188 + ], + "score": 0.83, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 176, + 272, + 190 + ], + "score": 1.0, + "content": "for an episode. We denote by", + "type": "text" + }, + { + "bbox": [ + 272, + 177, + 312, + 190 + ], + "score": 0.92, + "content": "p _ { \\pi ( z ) } ( s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "the distribution over (final) states induced by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 187, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 89, + 190, + 100, + 199 + ], + "score": 1.0, + "content": "78", + "type": "text" + }, + { + "bbox": [ + 105, + 187, + 219, + 202 + ], + "score": 1.0, + "content": "executing the policy of skill", + "type": "text" + }, + { + "bbox": [ + 219, + 190, + 225, + 198 + ], + "score": 0.68, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 187, + 241, + 202 + ], + "score": 1.0, + "content": ", by", + "type": "text" + }, + { + "bbox": [ + 242, + 189, + 274, + 200 + ], + "score": 0.91, + "content": "p ( z | s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 187, + 347, + 202 + ], + "score": 1.0, + "content": "the probability of", + "type": "text" + }, + { + "bbox": [ + 347, + 190, + 353, + 198 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 187, + 472, + 202 + ], + "score": 1.0, + "content": "being the skill to induce state", + "type": "text" + }, + { + "bbox": [ + 473, + 190, + 484, + 199 + ], + "score": 0.81, + "content": "s _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 187, + 506, + 202 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 197, + 498, + 214 + ], + "spans": [ + { + "bbox": [ + 89, + 201, + 99, + 210 + ], + "score": 1.0, + "content": "79", + "type": "text" + }, + { + "bbox": [ + 104, + 197, + 119, + 214 + ], + "score": 1.0, + "content": "let", + "type": "text" + }, + { + "bbox": [ + 119, + 199, + 243, + 213 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\bar { p } ( s _ { T } ) \\bar { ( s _ { T } ) } = \\bar { \\sum _ { z \\in Z } \\bar { \\rho ( z ) } } p _ { \\pi ( z ) } ( s _ { T } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 197, + 385, + 214 + ], + "score": 1.0, + "content": ". Then maximizing the MI between", + "type": "text" + }, + { + "bbox": [ + 385, + 200, + 394, + 209 + ], + "score": 0.82, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 197, + 412, + 214 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 412, + 200, + 425, + 210 + ], + "score": 0.9, + "content": "S _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 197, + 498, + 214 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 218, + 502, + 275 + ], + "lines": [ + { + "bbox": [ + 111, + 218, + 502, + 275 + ], + "spans": [ + { + "bbox": [ + 111, + 218, + 502, + 275 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } & { } & { \\displaystyle \\underset { | Z | , \\rho , \\pi } { \\operatorname* { m a x } } \\ : \\mathcal { Z } ( S _ { T } ; Z ) = \\mathcal { H } ( S _ { T } ) - \\mathcal { H } ( S _ { T } | Z ) = - { \\displaystyle \\sum _ { s _ { T } } } \\overline { { p } } ( s _ { T } ) \\log \\overline { { p } } ( s _ { T } ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p _ { \\pi ( z ) } ( s _ { T } ) \\right] } \\\\ & { } & { = \\mathcal { H } ( Z ) - \\mathcal { H } ( Z | S _ { T } ) = - { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\log \\rho ( z ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p ( z | s _ { T } ) \\right] , \\quad ( 1 ) } \\end{array}", + "type": "interline_equation", + "image_path": "12ce697ef699aff72c90ae32fd9cf68f47bc811b82b8a62f65f86a70a2e9cb19.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 111, + 218, + 502, + 237.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 111, + 237.0, + 502, + 256.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 111, + 256.0, + 502, + 275.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 93, + 281, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 90, + 281, + 504, + 295 + ], + "spans": [ + { + "bbox": [ + 90, + 281, + 208, + 295 + ], + "score": 1.0, + "content": "80 where in the expectations", + "type": "text" + }, + { + "bbox": [ + 208, + 281, + 272, + 294 + ], + "score": 0.92, + "content": "s _ { T } \\sim p _ { \\pi ( z ) } ( s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 281, + 452, + 295 + ], + "score": 1.0, + "content": ". As discussed in Sect. 1, learning the optimal", + "type": "text" + }, + { + "bbox": [ + 452, + 282, + 477, + 294 + ], + "score": 0.58, + "content": "| Z | , \\rho ,", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 281, + 497, + 295 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 497, + 285, + 504, + 291 + ], + "score": 0.69, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 291, + 284, + 306 + ], + "spans": [ + { + "bbox": [ + 89, + 291, + 284, + 306 + ], + "score": 1.0, + "content": "81 is a challenging problem [see e.g., 11, 9, 8].", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 105, + 314, + 232, + 328 + ], + "lines": [ + { + "bbox": [ + 104, + 314, + 232, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 314, + 232, + 331 + ], + "score": 1.0, + "content": "3 Algorithm Structure", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 457, + 345 + ], + "score": 1.0, + "content": "UPSIDE is based on three main components: a) the skill learning corresponding to stage", + "type": "text" + }, + { + "bbox": [ + 457, + 334, + 466, + 343 + ], + "score": 0.79, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 333, + 484, + 345 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 484, + 334, + 493, + 343 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "Fig. 1 and described in Sect. 3.1, b) a constrained optimization problem used to optimize the number", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 167, + 367 + ], + "score": 1.0, + "content": "of skills (stage", + "type": "text" + }, + { + "bbox": [ + 168, + 356, + 177, + 365 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 354, + 394, + 367 + ], + "score": 1.0, + "content": "and Sect. 3.2) and c) a tree-building procedure (stage", + "type": "text" + }, + { + "bbox": [ + 394, + 356, + 404, + 365 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "and Sect. 3.3). Together,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 471, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 471, + 378 + ], + "score": 1.0, + "content": "these components allow UPSIDE to discover skills that combine coverage and directedness.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 105, + 385, + 271, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 272, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 272, + 399 + ], + "score": 1.0, + "content": "3.1 Skill Structure and Optimization", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 101, + 399, + 506, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "score": 1.0, + "content": "As shown in e.g., [9, 30, 37], the level of stochasticity of each skill (e.g., induced via a regularization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 409, + 507, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 507, + 423 + ], + "score": 1.0, + "content": "on the entropy over the actions) plays a key role in trading off coverage and directedness. In fact,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 419, + 507, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 507, + 434 + ], + "score": 1.0, + "content": "while randomness promotes broader coverage, it may compromise the directedness of the skills.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 430, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 356, + 446 + ], + "score": 1.0, + "content": "In fact, a highly stochastic skill tends to induce a distribution", + "type": "text" + }, + { + "bbox": [ + 356, + 432, + 397, + 445 + ], + "score": 0.93, + "content": "p _ { \\pi ( z ) } ( s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 430, + 506, + 446 + ], + "score": 1.0, + "content": "over final states with high", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 204, + 457 + ], + "score": 1.0, + "content": "entropy (thus decreasing", + "type": "text" + }, + { + "bbox": [ + 204, + 443, + 254, + 456 + ], + "score": 0.92, + "content": "- \\mathcal { H } ( S _ { T } | Z ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 443, + 506, + 457 + ], + "score": 1.0, + "content": ", which prevents the skill to be reusable in solving sparse-reward", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 455, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 506, + 467 + ], + "score": 1.0, + "content": "downstream tasks where the objective is to reliably reach a specific goal state of the environment.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "Determining how much stochasticity to inject to adequately balance both objectives and optimize (1)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 476, + 196, + 488 + ], + "spans": [ + { + "bbox": [ + 104, + 476, + 196, + 488 + ], + "score": 1.0, + "content": "is a difficult problem.1", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 90, + 492, + 355, + 504 + ], + "lines": [ + { + "bbox": [ + 87, + 491, + 357, + 507 + ], + "spans": [ + { + "bbox": [ + 87, + 491, + 357, + 507 + ], + "score": 1.0, + "content": "96 We propose to design skills with a decoupled policy structure:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 109, + 508, + 385, + 576 + ], + "lines": [ + { + "bbox": [ + 108, + 507, + 386, + 520 + ], + "spans": [ + { + "bbox": [ + 108, + 507, + 223, + 520 + ], + "score": 1.0, + "content": "• A directed part (of length", + "type": "text" + }, + { + "bbox": [ + 223, + 509, + 232, + 518 + ], + "score": 0.75, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 507, + 386, + 520 + ], + "score": 1.0, + "content": ") with low stochasticity and trained to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 519, + 387, + 532 + ], + "spans": [ + { + "bbox": [ + 116, + 519, + 387, + 532 + ], + "score": 1.0, + "content": "reach a specific region of the environment. It is responsible for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 530, + 266, + 542 + ], + "spans": [ + { + "bbox": [ + 116, + 531, + 175, + 542 + ], + "score": 1.0, + "content": "increasing the", + "type": "text" + }, + { + "bbox": [ + 176, + 530, + 217, + 542 + ], + "score": 0.93, + "content": "- { \\mathcal { H } } ( S | Z )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 531, + 266, + 542 + ], + "score": 1.0, + "content": "term in (1).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 109, + 541, + 386, + 555 + ], + "spans": [ + { + "bbox": [ + 109, + 541, + 227, + 555 + ], + "score": 1.0, + "content": "• A diffusing part (of length", + "type": "text" + }, + { + "bbox": [ + 228, + 542, + 238, + 552 + ], + "score": 0.75, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 541, + 386, + 555 + ], + "score": 1.0, + "content": ") with high stochasticity to promote", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 552, + 387, + 565 + ], + "spans": [ + { + "bbox": [ + 116, + 552, + 387, + 565 + ], + "score": 1.0, + "content": "local coverage of the states around the region reached by the directed", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 564, + 349, + 577 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 276, + 577 + ], + "score": 1.0, + "content": "part. It is responsible for increasing the", + "type": "text" + }, + { + "bbox": [ + 276, + 564, + 300, + 576 + ], + "score": 0.91, + "content": "\\mathcal { H } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 564, + 349, + 577 + ], + "score": 1.0, + "content": "term in (1).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.0 + }, + { + "type": "image", + "bbox": [ + 392, + 486, + 509, + 548 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 392, + 486, + 509, + 548 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 392, + 486, + 509, + 548 + ], + "spans": [ + { + "bbox": [ + 392, + 486, + 509, + 548 + ], + "score": 0.959, + "type": "image", + "image_path": "cd135dcbe8681b37eac698ab1aa41d695eeae5d2288d124ad1c5924bd7a5fec1.jpg" + } + ] + } + ], + "index": 34.0, + "virtual_lines": [ + { + "bbox": [ + 392, + 486, + 509, + 517.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 392, + 517.0, + 509, + 548.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 392, + 551, + 505, + 574 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 392, + 551, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 392, + 551, + 506, + 562 + ], + "score": 1.0, + "content": "Figure 2: Directed and diffus-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 391, + 563, + 468, + 573 + ], + "spans": [ + { + "bbox": [ + 391, + 563, + 468, + 573 + ], + "score": 1.0, + "content": "ing parts of the skill.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.0 + } + ], + "index": 37.0 + }, + { + "type": "text", + "bbox": [ + 89, + 579, + 506, + 659 + ], + "lines": [ + { + "bbox": [ + 89, + 578, + 507, + 592 + ], + "spans": [ + { + "bbox": [ + 89, + 582, + 100, + 591 + ], + "score": 1.0, + "content": "98", + "type": "text" + }, + { + "bbox": [ + 104, + 578, + 428, + 592 + ], + "score": 1.0, + "content": "Similar to prior work [e.g., 11, 9], the policy associated to the directed part of skill", + "type": "text" + }, + { + "bbox": [ + 429, + 581, + 435, + 589 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 578, + 507, + 592 + ], + "score": 1.0, + "content": "is trained to max-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 89, + 590, + 504, + 604 + ], + "spans": [ + { + "bbox": [ + 89, + 592, + 100, + 602 + ], + "score": 1.0, + "content": "99", + "type": "text" + }, + { + "bbox": [ + 105, + 590, + 205, + 604 + ], + "score": 1.0, + "content": "imize an intrinsic reward", + "type": "text" + }, + { + "bbox": [ + 205, + 590, + 267, + 603 + ], + "score": 0.92, + "content": "r _ { z } ( s ) \\approx p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 590, + 300, + 604 + ], + "score": 1.0, + "content": ",2 where", + "type": "text" + }, + { + "bbox": [ + 300, + 590, + 326, + 603 + ], + "score": 0.93, + "content": "p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 590, + 497, + 604 + ], + "score": 1.0, + "content": "measures the “discriminability” of the skill", + "type": "text" + }, + { + "bbox": [ + 497, + 592, + 504, + 600 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 86, + 600, + 504, + 621 + ], + "spans": [ + { + "bbox": [ + 86, + 605, + 100, + 615 + ], + "score": 1.0, + "content": "100", + "type": "text" + }, + { + "bbox": [ + 102, + 600, + 168, + 621 + ], + "score": 1.0, + "content": "given the state", + "type": "text" + }, + { + "bbox": [ + 169, + 606, + 174, + 613 + ], + "score": 0.59, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 600, + 245, + 621 + ], + "score": 1.0, + "content": ". More formally,", + "type": "text" + }, + { + "bbox": [ + 245, + 604, + 265, + 615 + ], + "score": 0.9, + "content": "\\pi ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 600, + 407, + 621 + ], + "score": 1.0, + "content": "maximizes the cumulative reward", + "type": "text" + }, + { + "bbox": [ + 407, + 601, + 504, + 617 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathbb { E } _ { \\pi ( z ) } \\left[ \\sum _ { t = T + 1 } ^ { T + H } r _ { z } ( s _ { t } ) \\right] } \\end{array}", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 86, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 86, + 616, + 100, + 626 + ], + "score": 1.0, + "content": "101", + "type": "text" + }, + { + "bbox": [ + 104, + 614, + 506, + 627 + ], + "score": 1.0, + "content": "over the states traversed by the policy during the diffusing part. In practice, we also add a small", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 86, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 86, + 627, + 100, + 637 + ], + "score": 1.0, + "content": "102", + "type": "text" + }, + { + "bbox": [ + 105, + 625, + 200, + 638 + ], + "score": 1.0, + "content": "entropy regularization", + "type": "text" + }, + { + "bbox": [ + 200, + 626, + 253, + 637 + ], + "score": 0.93, + "content": "\\mathcal { H } ( \\pi ( \\cdot | z , s _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "to the directed policy in order to ensure a minimum level of", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 86, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 86, + 638, + 100, + 648 + ], + "score": 1.0, + "content": "103", + "type": "text" + }, + { + "bbox": [ + 105, + 636, + 506, + 649 + ], + "score": 1.0, + "content": "exploration and make the learning more robust. For the diffusing part, we rely on a simple random", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 87, + 648, + 411, + 659 + ], + "spans": [ + { + "bbox": [ + 87, + 650, + 100, + 658 + ], + "score": 1.0, + "content": "104", + "type": "text" + }, + { + "bbox": [ + 105, + 648, + 411, + 659 + ], + "score": 1.0, + "content": "walk policy (i.e., a stochastic policy with uniform distribution over actions).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 671, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 119, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 119, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "1In RL, stochasticity is injected at “train time” to boost exploration or improve robustness, while the policy", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 681, + 479, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 479, + 692 + ], + "score": 1.0, + "content": "executed at “test time” is deterministic. Here we refer to stochasticity introduced to better optimize (1).", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 690, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 119, + 690, + 506, + 704 + ], + "score": 1.0, + "content": "2Although [11, 9] employ rewards in the log domain, we find that using a reward that is a non-linear", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 180, + 713 + ], + "score": 1.0, + "content": "transformation into", + "type": "text" + }, + { + "bbox": [ + 181, + 702, + 200, + 712 + ], + "score": 0.34, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 701, + 505, + 713 + ], + "score": 1.0, + "content": "works better in practice, as also observed in [33, 5]. Furthermore, in practice we", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 711, + 443, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 134, + 722 + ], + "score": 1.0, + "content": "replace", + "type": "text" + }, + { + "bbox": [ + 135, + 712, + 159, + 722 + ], + "score": 0.91, + "content": "p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 711, + 322, + 722 + ], + "score": 1.0, + "content": "by the predictions of a learned discriminator", + "type": "text" + }, + { + "bbox": [ + 322, + 713, + 350, + 723 + ], + "score": 0.93, + "content": "q _ { \\phi } ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 711, + 443, + 722 + ], + "score": 1.0, + "content": "as explained in Sect. 3.4.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "index", + "bbox": [ + 93, + 72, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 92, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 92, + 76, + 100, + 83 + ], + "score": 1.0, + "content": "68", + "type": "text" + }, + { + "bbox": [ + 104, + 72, + 262, + 86 + ], + "score": 1.0, + "content": "episode from a designated initial state", + "type": "text" + }, + { + "bbox": [ + 262, + 73, + 291, + 84 + ], + "score": 0.93, + "content": "s _ { 0 } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 72, + 505, + 86 + ], + "score": 1.0, + "content": ". Upon termination of the chosen policy, the agent is", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 84, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 92, + 87, + 99, + 94 + ], + "score": 1.0, + "content": "69", + "type": "text" + }, + { + "bbox": [ + 105, + 84, + 158, + 96 + ], + "score": 1.0, + "content": "then reset to", + "type": "text" + }, + { + "bbox": [ + 158, + 85, + 168, + 95 + ], + "score": 0.84, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 84, + 505, + 96 + ], + "score": 1.0, + "content": ". This setting is particularly challenging from an exploration point of view since the", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 94, + 377, + 108 + ], + "spans": [ + { + "bbox": [ + 92, + 98, + 99, + 105 + ], + "score": 1.0, + "content": "70", + "type": "text" + }, + { + "bbox": [ + 105, + 94, + 377, + 108 + ], + "score": 1.0, + "content": "agent cannot rely on the initial distribution to cover the state space.", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 89, + 113, + 99, + 122 + ], + "score": 1.0, + "content": "71", + "type": "text" + }, + { + "bbox": [ + 105, + 110, + 470, + 124 + ], + "score": 1.0, + "content": "We recall the MI-based unsupervised skill discovery approach [see e.g., 11]. Denote by", + "type": "text" + }, + { + "bbox": [ + 471, + 111, + 479, + 121 + ], + "score": 0.82, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "some", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 89, + 124, + 100, + 133 + ], + "score": 1.0, + "content": "72", + "type": "text" + }, + { + "bbox": [ + 105, + 121, + 302, + 135 + ], + "score": 1.0, + "content": "(latent) variables on which the skills of length", + "type": "text" + }, + { + "bbox": [ + 302, + 122, + 311, + 132 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "are conditioned. There are three optimization", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 89, + 135, + 99, + 144 + ], + "score": 1.0, + "content": "73", + "type": "text" + }, + { + "bbox": [ + 105, + 132, + 150, + 145 + ], + "score": 1.0, + "content": "variables:", + "type": "text" + }, + { + "bbox": [ + 150, + 133, + 161, + 144 + ], + "score": 0.41, + "content": "( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 132, + 312, + 145 + ], + "score": 1.0, + "content": "the support of the skills denoted by", + "type": "text" + }, + { + "bbox": [ + 313, + 133, + 327, + 144 + ], + "score": 0.91, + "content": "| Z |", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 132, + 464, + 145 + ], + "score": 1.0, + "content": "(we consider it to be discrete so", + "type": "text" + }, + { + "bbox": [ + 464, + 133, + 478, + 144 + ], + "score": 0.9, + "content": "| Z |", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 142, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 89, + 146, + 100, + 155 + ], + "score": 1.0, + "content": "74", + "type": "text" + }, + { + "bbox": [ + 105, + 142, + 181, + 157 + ], + "score": 1.0, + "content": "number of skills),", + "type": "text" + }, + { + "bbox": [ + 182, + 144, + 195, + 155 + ], + "score": 0.55, + "content": "( i i )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 142, + 240, + 157 + ], + "score": 1.0, + "content": "the policy", + "type": "text" + }, + { + "bbox": [ + 240, + 144, + 260, + 155 + ], + "score": 0.93, + "content": "\\pi ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 142, + 337, + 157 + ], + "score": 1.0, + "content": "associated to skill", + "type": "text" + }, + { + "bbox": [ + 337, + 146, + 343, + 154 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 142, + 365, + 157 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 365, + 144, + 381, + 155 + ], + "score": 0.53, + "content": "( i i i )", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 142, + 456, + 157 + ], + "score": 1.0, + "content": "the sampling rule", + "type": "text" + }, + { + "bbox": [ + 456, + 146, + 463, + 155 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 142, + 486, + 157 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 486, + 144, + 505, + 156 + ], + "score": 0.89, + "content": "\\rho ( z )", + "type": "inline_equation" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 89, + 156, + 99, + 166 + ], + "score": 1.0, + "content": "75", + "type": "text" + }, + { + "bbox": [ + 105, + 154, + 250, + 167 + ], + "score": 1.0, + "content": "is the probability of sampling skill", + "type": "text" + }, + { + "bbox": [ + 251, + 156, + 257, + 164 + ], + "score": 0.8, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 154, + 462, + 167 + ], + "score": 1.0, + "content": "at the beginning of the episode). Let the variable", + "type": "text" + }, + { + "bbox": [ + 462, + 155, + 476, + 165 + ], + "score": 0.89, + "content": "S _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "be the", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 89, + 167, + 99, + 177 + ], + "score": 1.0, + "content": "76", + "type": "text" + }, + { + "bbox": [ + 105, + 164, + 302, + 178 + ], + "score": 1.0, + "content": "random (final) state induced by sampling a skill", + "type": "text" + }, + { + "bbox": [ + 302, + 169, + 308, + 175 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 164, + 331, + 178 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 332, + 167, + 338, + 177 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 164, + 485, + 178 + ], + "score": 1.0, + "content": "and executing the associated policy", + "type": "text" + }, + { + "bbox": [ + 485, + 166, + 505, + 177 + ], + "score": 0.91, + "content": "\\pi ( z )", + "type": "inline_equation" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 89, + 178, + 99, + 188 + ], + "score": 1.0, + "content": "77", + "type": "text" + }, + { + "bbox": [ + 105, + 176, + 129, + 190 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 130, + 178, + 140, + 188 + ], + "score": 0.83, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 176, + 272, + 190 + ], + "score": 1.0, + "content": "for an episode. We denote by", + "type": "text" + }, + { + "bbox": [ + 272, + 177, + 312, + 190 + ], + "score": 0.92, + "content": "p _ { \\pi ( z ) } ( s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "the distribution over (final) states induced by", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 187, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 89, + 190, + 100, + 199 + ], + "score": 1.0, + "content": "78", + "type": "text" + }, + { + "bbox": [ + 105, + 187, + 219, + 202 + ], + "score": 1.0, + "content": "executing the policy of skill", + "type": "text" + }, + { + "bbox": [ + 219, + 190, + 225, + 198 + ], + "score": 0.68, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 187, + 241, + 202 + ], + "score": 1.0, + "content": ", by", + "type": "text" + }, + { + "bbox": [ + 242, + 189, + 274, + 200 + ], + "score": 0.91, + "content": "p ( z | s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 187, + 347, + 202 + ], + "score": 1.0, + "content": "the probability of", + "type": "text" + }, + { + "bbox": [ + 347, + 190, + 353, + 198 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 187, + 472, + 202 + ], + "score": 1.0, + "content": "being the skill to induce state", + "type": "text" + }, + { + "bbox": [ + 473, + 190, + 484, + 199 + ], + "score": 0.81, + "content": "s _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 187, + 506, + 202 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 197, + 498, + 214 + ], + "spans": [ + { + "bbox": [ + 89, + 201, + 99, + 210 + ], + "score": 1.0, + "content": "79", + "type": "text" + }, + { + "bbox": [ + 104, + 197, + 119, + 214 + ], + "score": 1.0, + "content": "let", + "type": "text" + }, + { + "bbox": [ + 119, + 199, + 243, + 213 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\bar { p } ( s _ { T } ) \\bar { ( s _ { T } ) } = \\bar { \\sum _ { z \\in Z } \\bar { \\rho ( z ) } } p _ { \\pi ( z ) } ( s _ { T } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 197, + 385, + 214 + ], + "score": 1.0, + "content": ". Then maximizing the MI between", + "type": "text" + }, + { + "bbox": [ + 385, + 200, + 394, + 209 + ], + "score": 0.82, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 197, + 412, + 214 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 412, + 200, + 425, + 210 + ], + "score": 0.9, + "content": "S _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 197, + 498, + 214 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + } + ], + "index": 1, + "bbox_fs": [ + 92, + 72, + 505, + 108 + ] + }, + { + "type": "index", + "bbox": [ + 89, + 110, + 505, + 213 + ], + "lines": [], + "index": 7, + "bbox_fs": [ + 89, + 110, + 506, + 214 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 218, + 502, + 275 + ], + "lines": [ + { + "bbox": [ + 111, + 218, + 502, + 275 + ], + "spans": [ + { + "bbox": [ + 111, + 218, + 502, + 275 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } & { } & { \\displaystyle \\underset { | Z | , \\rho , \\pi } { \\operatorname* { m a x } } \\ : \\mathcal { Z } ( S _ { T } ; Z ) = \\mathcal { H } ( S _ { T } ) - \\mathcal { H } ( S _ { T } | Z ) = - { \\displaystyle \\sum _ { s _ { T } } } \\overline { { p } } ( s _ { T } ) \\log \\overline { { p } } ( s _ { T } ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p _ { \\pi ( z ) } ( s _ { T } ) \\right] } \\\\ & { } & { = \\mathcal { H } ( Z ) - \\mathcal { H } ( Z | S _ { T } ) = - { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\log \\rho ( z ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p ( z | s _ { T } ) \\right] , \\quad ( 1 ) } \\end{array}", + "type": "interline_equation", + "image_path": "12ce697ef699aff72c90ae32fd9cf68f47bc811b82b8a62f65f86a70a2e9cb19.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 111, + 218, + 502, + 237.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 111, + 237.0, + 502, + 256.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 111, + 256.0, + 502, + 275.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "index", + "bbox": [ + 93, + 281, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 90, + 281, + 504, + 295 + ], + "spans": [ + { + "bbox": [ + 90, + 281, + 208, + 295 + ], + "score": 1.0, + "content": "80 where in the expectations", + "type": "text" + }, + { + "bbox": [ + 208, + 281, + 272, + 294 + ], + "score": 0.92, + "content": "s _ { T } \\sim p _ { \\pi ( z ) } ( s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 281, + 452, + 295 + ], + "score": 1.0, + "content": ". As discussed in Sect. 1, learning the optimal", + "type": "text" + }, + { + "bbox": [ + 452, + 282, + 477, + 294 + ], + "score": 0.58, + "content": "| Z | , \\rho ,", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 281, + 497, + 295 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 497, + 285, + 504, + 291 + ], + "score": 0.69, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 291, + 284, + 306 + ], + "spans": [ + { + "bbox": [ + 89, + 291, + 284, + 306 + ], + "score": 1.0, + "content": "81 is a challenging problem [see e.g., 11, 9, 8].", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + } + ], + "index": 15.5, + "bbox_fs": [ + 89, + 281, + 504, + 306 + ] + }, + { + "type": "title", + "bbox": [ + 105, + 314, + 232, + 328 + ], + "lines": [ + { + "bbox": [ + 104, + 314, + 232, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 314, + 232, + 331 + ], + "score": 1.0, + "content": "3 Algorithm Structure", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 457, + 345 + ], + "score": 1.0, + "content": "UPSIDE is based on three main components: a) the skill learning corresponding to stage", + "type": "text" + }, + { + "bbox": [ + 457, + 334, + 466, + 343 + ], + "score": 0.79, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 333, + 484, + 345 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 484, + 334, + 493, + 343 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "Fig. 1 and described in Sect. 3.1, b) a constrained optimization problem used to optimize the number", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 167, + 367 + ], + "score": 1.0, + "content": "of skills (stage", + "type": "text" + }, + { + "bbox": [ + 168, + 356, + 177, + 365 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 354, + 394, + 367 + ], + "score": 1.0, + "content": "and Sect. 3.2) and c) a tree-building procedure (stage", + "type": "text" + }, + { + "bbox": [ + 394, + 356, + 404, + 365 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "and Sect. 3.3). Together,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 471, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 471, + 378 + ], + "score": 1.0, + "content": "these components allow UPSIDE to discover skills that combine coverage and directedness.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 333, + 506, + 378 + ] + }, + { + "type": "title", + "bbox": [ + 105, + 385, + 271, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 272, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 272, + 399 + ], + "score": 1.0, + "content": "3.1 Skill Structure and Optimization", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 101, + 399, + 506, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "score": 1.0, + "content": "As shown in e.g., [9, 30, 37], the level of stochasticity of each skill (e.g., induced via a regularization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 409, + 507, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 507, + 423 + ], + "score": 1.0, + "content": "on the entropy over the actions) plays a key role in trading off coverage and directedness. In fact,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 419, + 507, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 507, + 434 + ], + "score": 1.0, + "content": "while randomness promotes broader coverage, it may compromise the directedness of the skills.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 430, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 356, + 446 + ], + "score": 1.0, + "content": "In fact, a highly stochastic skill tends to induce a distribution", + "type": "text" + }, + { + "bbox": [ + 356, + 432, + 397, + 445 + ], + "score": 0.93, + "content": "p _ { \\pi ( z ) } ( s _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 430, + 506, + 446 + ], + "score": 1.0, + "content": "over final states with high", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 204, + 457 + ], + "score": 1.0, + "content": "entropy (thus decreasing", + "type": "text" + }, + { + "bbox": [ + 204, + 443, + 254, + 456 + ], + "score": 0.92, + "content": "- \\mathcal { H } ( S _ { T } | Z ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 443, + 506, + 457 + ], + "score": 1.0, + "content": ", which prevents the skill to be reusable in solving sparse-reward", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 455, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 506, + 467 + ], + "score": 1.0, + "content": "downstream tasks where the objective is to reliably reach a specific goal state of the environment.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "Determining how much stochasticity to inject to adequately balance both objectives and optimize (1)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 476, + 196, + 488 + ], + "spans": [ + { + "bbox": [ + 104, + 476, + 196, + 488 + ], + "score": 1.0, + "content": "is a difficult problem.1", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 398, + 507, + 488 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 492, + 355, + 504 + ], + "lines": [ + { + "bbox": [ + 87, + 491, + 357, + 507 + ], + "spans": [ + { + "bbox": [ + 87, + 491, + 357, + 507 + ], + "score": 1.0, + "content": "96 We propose to design skills with a decoupled policy structure:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 87, + 491, + 357, + 507 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 508, + 385, + 576 + ], + "lines": [ + { + "bbox": [ + 108, + 507, + 386, + 520 + ], + "spans": [ + { + "bbox": [ + 108, + 507, + 223, + 520 + ], + "score": 1.0, + "content": "• A directed part (of length", + "type": "text" + }, + { + "bbox": [ + 223, + 509, + 232, + 518 + ], + "score": 0.75, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 507, + 386, + 520 + ], + "score": 1.0, + "content": ") with low stochasticity and trained to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 519, + 387, + 532 + ], + "spans": [ + { + "bbox": [ + 116, + 519, + 387, + 532 + ], + "score": 1.0, + "content": "reach a specific region of the environment. It is responsible for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 530, + 266, + 542 + ], + "spans": [ + { + "bbox": [ + 116, + 531, + 175, + 542 + ], + "score": 1.0, + "content": "increasing the", + "type": "text" + }, + { + "bbox": [ + 176, + 530, + 217, + 542 + ], + "score": 0.93, + "content": "- { \\mathcal { H } } ( S | Z )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 531, + 266, + 542 + ], + "score": 1.0, + "content": "term in (1).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 109, + 541, + 386, + 555 + ], + "spans": [ + { + "bbox": [ + 109, + 541, + 227, + 555 + ], + "score": 1.0, + "content": "• A diffusing part (of length", + "type": "text" + }, + { + "bbox": [ + 228, + 542, + 238, + 552 + ], + "score": 0.75, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 541, + 386, + 555 + ], + "score": 1.0, + "content": ") with high stochasticity to promote", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 552, + 387, + 565 + ], + "spans": [ + { + "bbox": [ + 116, + 552, + 387, + 565 + ], + "score": 1.0, + "content": "local coverage of the states around the region reached by the directed", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 564, + 349, + 577 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 276, + 577 + ], + "score": 1.0, + "content": "part. It is responsible for increasing the", + "type": "text" + }, + { + "bbox": [ + 276, + 564, + 300, + 576 + ], + "score": 0.91, + "content": "\\mathcal { H } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 564, + 349, + 577 + ], + "score": 1.0, + "content": "term in (1).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.0, + "bbox_fs": [ + 108, + 507, + 387, + 577 + ] + }, + { + "type": "image", + "bbox": [ + 392, + 486, + 509, + 548 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 392, + 486, + 509, + 548 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 392, + 486, + 509, + 548 + ], + "spans": [ + { + "bbox": [ + 392, + 486, + 509, + 548 + ], + "score": 0.959, + "type": "image", + "image_path": "cd135dcbe8681b37eac698ab1aa41d695eeae5d2288d124ad1c5924bd7a5fec1.jpg" + } + ] + } + ], + "index": 34.0, + "virtual_lines": [ + { + "bbox": [ + 392, + 486, + 509, + 517.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 392, + 517.0, + 509, + 548.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 392, + 551, + 505, + 574 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 392, + 551, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 392, + 551, + 506, + 562 + ], + "score": 1.0, + "content": "Figure 2: Directed and diffus-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 391, + 563, + 468, + 573 + ], + "spans": [ + { + "bbox": [ + 391, + 563, + 468, + 573 + ], + "score": 1.0, + "content": "ing parts of the skill.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.0 + } + ], + "index": 37.0 + }, + { + "type": "index", + "bbox": [ + 89, + 579, + 506, + 659 + ], + "lines": [ + { + "bbox": [ + 89, + 578, + 507, + 592 + ], + "spans": [ + { + "bbox": [ + 89, + 582, + 100, + 591 + ], + "score": 1.0, + "content": "98", + "type": "text" + }, + { + "bbox": [ + 104, + 578, + 428, + 592 + ], + "score": 1.0, + "content": "Similar to prior work [e.g., 11, 9], the policy associated to the directed part of skill", + "type": "text" + }, + { + "bbox": [ + 429, + 581, + 435, + 589 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 578, + 507, + 592 + ], + "score": 1.0, + "content": "is trained to max-", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 590, + 504, + 604 + ], + "spans": [ + { + "bbox": [ + 89, + 592, + 100, + 602 + ], + "score": 1.0, + "content": "99", + "type": "text" + }, + { + "bbox": [ + 105, + 590, + 205, + 604 + ], + "score": 1.0, + "content": "imize an intrinsic reward", + "type": "text" + }, + { + "bbox": [ + 205, + 590, + 267, + 603 + ], + "score": 0.92, + "content": "r _ { z } ( s ) \\approx p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 590, + 300, + 604 + ], + "score": 1.0, + "content": ",2 where", + "type": "text" + }, + { + "bbox": [ + 300, + 590, + 326, + 603 + ], + "score": 0.93, + "content": "p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 590, + 497, + 604 + ], + "score": 1.0, + "content": "measures the “discriminability” of the skill", + "type": "text" + }, + { + "bbox": [ + 497, + 592, + 504, + 600 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 600, + 504, + 621 + ], + "spans": [ + { + "bbox": [ + 86, + 605, + 100, + 615 + ], + "score": 1.0, + "content": "100", + "type": "text" + }, + { + "bbox": [ + 102, + 600, + 168, + 621 + ], + "score": 1.0, + "content": "given the state", + "type": "text" + }, + { + "bbox": [ + 169, + 606, + 174, + 613 + ], + "score": 0.59, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 600, + 245, + 621 + ], + "score": 1.0, + "content": ". More formally,", + "type": "text" + }, + { + "bbox": [ + 245, + 604, + 265, + 615 + ], + "score": 0.9, + "content": "\\pi ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 600, + 407, + 621 + ], + "score": 1.0, + "content": "maximizes the cumulative reward", + "type": "text" + }, + { + "bbox": [ + 407, + 601, + 504, + 617 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathbb { E } _ { \\pi ( z ) } \\left[ \\sum _ { t = T + 1 } ^ { T + H } r _ { z } ( s _ { t } ) \\right] } \\end{array}", + "type": "inline_equation" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 86, + 616, + 100, + 626 + ], + "score": 1.0, + "content": "101", + "type": "text" + }, + { + "bbox": [ + 104, + 614, + 506, + 627 + ], + "score": 1.0, + "content": "over the states traversed by the policy during the diffusing part. In practice, we also add a small", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 86, + 627, + 100, + 637 + ], + "score": 1.0, + "content": "102", + "type": "text" + }, + { + "bbox": [ + 105, + 625, + 200, + 638 + ], + "score": 1.0, + "content": "entropy regularization", + "type": "text" + }, + { + "bbox": [ + 200, + 626, + 253, + 637 + ], + "score": 0.93, + "content": "\\mathcal { H } ( \\pi ( \\cdot | z , s _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "to the directed policy in order to ensure a minimum level of", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 86, + 638, + 100, + 648 + ], + "score": 1.0, + "content": "103", + "type": "text" + }, + { + "bbox": [ + 105, + 636, + 506, + 649 + ], + "score": 1.0, + "content": "exploration and make the learning more robust. For the diffusing part, we rely on a simple random", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 648, + 411, + 659 + ], + "spans": [ + { + "bbox": [ + 87, + 650, + 100, + 658 + ], + "score": 1.0, + "content": "104", + "type": "text" + }, + { + "bbox": [ + 105, + 648, + 411, + 659 + ], + "score": 1.0, + "content": "walk policy (i.e., a stochastic policy with uniform distribution over actions).", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 86, + 74, + 100, + 85 + ], + "score": 1.0, + "content": "105", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "Intuitively, the diffusing part defines a cluster of states that is used as a goal for the directed part.", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 86, + 86, + 99, + 95 + ], + "score": 1.0, + "content": "106", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 83, + 407, + 96 + ], + "score": 1.0, + "content": "This allows us to “ground” the latent variable representations of the skills", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 408, + 84, + 416, + 93 + ], + "score": 0.83, + "content": "Z", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 417, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "to specific regions of", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 86, + 96, + 100, + 107 + ], + "score": 1.0, + "content": "107", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "the environment (i.e., the clusters). As a result, maximizing the MI over such skills can be seen as", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 106, + 372, + 118 + ], + "spans": [ + { + "bbox": [ + 86, + 108, + 100, + 117 + ], + "score": 1.0, + "content": "108", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 106, + 372, + 118 + ], + "score": 1.0, + "content": "learning a set of “cluster-conditioned”, and thus directed, policies.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + } + ], + "index": 45, + "bbox_fs": [ + 86, + 578, + 507, + 659 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 87, + 72, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 86, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 86, + 74, + 100, + 85 + ], + "score": 1.0, + "content": "105", + "type": "text" + }, + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "Intuitively, the diffusing part defines a cluster of states that is used as a goal for the directed part.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 86, + 86, + 99, + 95 + ], + "score": 1.0, + "content": "106", + "type": "text" + }, + { + "bbox": [ + 106, + 83, + 407, + 96 + ], + "score": 1.0, + "content": "This allows us to “ground” the latent variable representations of the skills", + "type": "text" + }, + { + "bbox": [ + 408, + 84, + 416, + 93 + ], + "score": 0.83, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "to specific regions of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 86, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 86, + 96, + 100, + 107 + ], + "score": 1.0, + "content": "107", + "type": "text" + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "the environment (i.e., the clusters). As a result, maximizing the MI over such skills can be seen as", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 106, + 372, + 118 + ], + "spans": [ + { + "bbox": [ + 86, + 108, + 100, + 117 + ], + "score": 1.0, + "content": "108", + "type": "text" + }, + { + "bbox": [ + 106, + 106, + 372, + 118 + ], + "score": 1.0, + "content": "learning a set of “cluster-conditioned”, and thus directed, policies.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 105, + 127, + 271, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 272, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 272, + 142 + ], + "score": 1.0, + "content": "3.2 Skill Support and Sampling Rule", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 87, + 144, + 505, + 178 + ], + "lines": [ + { + "bbox": [ + 87, + 142, + 502, + 158 + ], + "spans": [ + { + "bbox": [ + 87, + 147, + 99, + 156 + ], + "score": 1.0, + "content": "110", + "type": "text" + }, + { + "bbox": [ + 102, + 142, + 375, + 158 + ], + "score": 1.0, + "content": "The MI objective (1) crucially depends on the number of skills", + "type": "text" + }, + { + "bbox": [ + 376, + 145, + 397, + 156 + ], + "score": 0.88, + "content": "( | Z | )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 142, + 483, + 158 + ], + "score": 1.0, + "content": "and the distribution", + "type": "text" + }, + { + "bbox": [ + 484, + 145, + 502, + 157 + ], + "score": 0.91, + "content": "\\rho ( z )", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 87, + 159, + 99, + 167 + ], + "score": 1.0, + "content": "111", + "type": "text" + }, + { + "bbox": [ + 105, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "Unfortunately, it is been shown [e.g., 8] that solving (1) is particularly challenging. In order to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 166, + 471, + 180 + ], + "spans": [ + { + "bbox": [ + 87, + 169, + 100, + 178 + ], + "score": 1.0, + "content": "112", + "type": "text" + }, + { + "bbox": [ + 104, + 166, + 471, + 180 + ], + "score": 1.0, + "content": "simplify the optimization and the associated learning problem, we modify (1) in two ways.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 86, + 183, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 86, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 86, + 185, + 100, + 195 + ], + "score": 1.0, + "content": "113", + "type": "text" + }, + { + "bbox": [ + 105, + 183, + 468, + 195 + ], + "score": 1.0, + "content": "First, coherently with the skill optimization detailed in Sect. 3.1, the random variable", + "type": "text" + }, + { + "bbox": [ + 469, + 184, + 477, + 194 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 86, + 196, + 100, + 205 + ], + "score": 1.0, + "content": "114", + "type": "text" + }, + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "conditional entropy is any state reached during the diffusing part of the skill and not just the terminal", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 204, + 504, + 218 + ], + "spans": [ + { + "bbox": [ + 86, + 206, + 100, + 217 + ], + "score": 1.0, + "content": "115", + "type": "text" + }, + { + "bbox": [ + 105, + 204, + 248, + 218 + ], + "score": 1.0, + "content": "state. More formally, we denote by", + "type": "text" + }, + { + "bbox": [ + 248, + 205, + 266, + 216 + ], + "score": 0.87, + "content": "S _ { \\mathrm { d i f f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 204, + 497, + 218 + ], + "score": 1.0, + "content": "the random variable and its distribution for a specific skill", + "type": "text" + }, + { + "bbox": [ + 497, + 207, + 504, + 215 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 214, + 507, + 234 + ], + "spans": [ + { + "bbox": [ + 86, + 219, + 100, + 230 + ], + "score": 1.0, + "content": "116", + "type": "text" + }, + { + "bbox": [ + 104, + 214, + 116, + 234 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 216, + 269, + 231 + ], + "score": 0.92, + "content": "p _ { \\pi ( z ) } ( s _ { \\mathrm { d i f f } } ) = 1 / \\dot { H } \\sum _ { t = T + 1 } ^ { T + H } p _ { \\pi ( z ) } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 214, + 507, + 234 + ], + "score": 1.0, + "content": ", i.e., the distribution over states obtained by averaging the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 85, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 85, + 231, + 100, + 242 + ], + "score": 1.0, + "content": "117", + "type": "text" + }, + { + "bbox": [ + 105, + 229, + 355, + 243 + ], + "score": 1.0, + "content": "distributions at any of the steps in the diffusing part. Similarly,", + "type": "text" + }, + { + "bbox": [ + 356, + 230, + 392, + 242 + ], + "score": 0.92, + "content": "p ( z | s _ { \\mathrm { d i f f } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "now denotes the probability", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 86, + 243, + 99, + 253 + ], + "score": 1.0, + "content": "118", + "type": "text" + }, + { + "bbox": [ + 105, + 240, + 117, + 254 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 243, + 123, + 251 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 240, + 224, + 254 + ], + "score": 1.0, + "content": "being the skill to traverse", + "type": "text" + }, + { + "bbox": [ + 224, + 243, + 241, + 253 + ], + "score": 0.87, + "content": "s _ { \\mathrm { d i f f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "during its diffusing part. As a result, training the skills to maximize", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 86, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 86, + 254, + 99, + 263 + ], + "score": 1.0, + "content": "119", + "type": "text" + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "MI naturally leads the diffusing parts to “push” the directed parts away so as to reach diverse regions", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 86, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 86, + 264, + 100, + 275 + ], + "score": 1.0, + "content": "120", + "type": "text" + }, + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "of the environment. The combination of “global” coverage of the directed parts and “local” coverage", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 86, + 272, + 493, + 286 + ], + "spans": [ + { + "bbox": [ + 86, + 276, + 99, + 285 + ], + "score": 1.0, + "content": "121", + "type": "text" + }, + { + "bbox": [ + 105, + 272, + 426, + 286 + ], + "score": 1.0, + "content": "of the diffusing part ensures that the whole environment is properly visited with", + "type": "text" + }, + { + "bbox": [ + 427, + 273, + 463, + 286 + ], + "score": 0.92, + "content": "| Z | \\ll S", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 272, + 493, + 286 + ], + "score": 1.0, + "content": "skills.3", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 86, + 290, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 86, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 86, + 292, + 100, + 301 + ], + "score": 1.0, + "content": "122", + "type": "text" + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "Second, we introduce an alternative problem that simplifies the optimization while preserving the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 85, + 300, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 85, + 303, + 100, + 313 + ], + "score": 1.0, + "content": "123", + "type": "text" + }, + { + "bbox": [ + 105, + 300, + 506, + 314 + ], + "score": 1.0, + "content": "coverage and directedness properties of MI. This is achieved by introducing a stronger requirement", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 86, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 86, + 315, + 100, + 323 + ], + "score": 1.0, + "content": "124", + "type": "text" + }, + { + "bbox": [ + 105, + 311, + 348, + 325 + ], + "score": 1.0, + "content": "on the discriminability. While the conditional entropy term", + "type": "text" + }, + { + "bbox": [ + 348, + 312, + 389, + 324 + ], + "score": 0.91, + "content": "- { \\mathcal { H } } ( Z | S )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "in (1) promotes the discrim-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 85, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 85, + 325, + 100, + 335 + ], + "score": 1.0, + "content": "125", + "type": "text" + }, + { + "bbox": [ + 104, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "inability of skills on average, we argue that a more suitable objective is to constrain each skill to", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 85, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 85, + 335, + 100, + 346 + ], + "score": 1.0, + "content": "126", + "type": "text" + }, + { + "bbox": [ + 106, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "achieve a minimum level of discriminability. First, we move from the average to the minimum over", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 86, + 344, + 314, + 358 + ], + "spans": [ + { + "bbox": [ + 86, + 347, + 100, + 357 + ], + "score": 1.0, + "content": "127", + "type": "text" + }, + { + "bbox": [ + 105, + 344, + 314, + 358 + ], + "score": 1.0, + "content": "skills by lower bounding the conditional entropy as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 362, + 454, + 390 + ], + "lines": [ + { + "bbox": [ + 155, + 362, + 454, + 390 + ], + "spans": [ + { + "bbox": [ + 155, + 362, + 454, + 390 + ], + "score": 0.91, + "content": "- \\mathcal { H } ( Z | S _ { \\mathrm { d i f f } } ) = \\sum _ { z \\in Z } \\rho ( z ) \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\operatorname* { m i n } _ { z \\in Z } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] ,", + "type": "interline_equation", + "image_path": "3157110c6c064eebeada077a520a63d063d2f20ec053138229dd1dff3acd2c56.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 155, + 362, + 454, + 371.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 155, + 371.3333333333333, + 454, + 380.66666666666663 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 155, + 380.66666666666663, + 454, + 389.99999999999994 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 87, + 396, + 427, + 408 + ], + "lines": [ + { + "bbox": [ + 85, + 395, + 428, + 410 + ], + "spans": [ + { + "bbox": [ + 85, + 395, + 318, + 410 + ], + "score": 1.0, + "content": "128 which leads to the following optimization (assuming", + "type": "text" + }, + { + "bbox": [ + 318, + 399, + 325, + 406 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 395, + 428, + 410 + ], + "score": 1.0, + "content": "is fixed for convenience)", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 414, + 398, + 438 + ], + "lines": [ + { + "bbox": [ + 212, + 414, + 398, + 438 + ], + "spans": [ + { + "bbox": [ + 212, + 414, + 398, + 438 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { | Z | = N , \\rho } \\bigg \\{ \\mathcal { H } ( Z ) + \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\bigg \\} ,", + "type": "interline_equation", + "image_path": "2f057bb1c2ed4a8288cb1135168369eb9b6e22ba2a4b80fd54634b61436bd511.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 212, + 414, + 398, + 438 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 86, + 445, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 86, + 446, + 503, + 458 + ], + "spans": [ + { + "bbox": [ + 86, + 448, + 100, + 457 + ], + "score": 1.0, + "content": "129", + "type": "text" + }, + { + "bbox": [ + 105, + 446, + 270, + 457 + ], + "score": 1.0, + "content": "where with an abuse of notation we use", + "type": "text" + }, + { + "bbox": [ + 271, + 446, + 304, + 458 + ], + "score": 0.92, + "content": "z \\in [ N ]", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 446, + 415, + 457 + ], + "score": 1.0, + "content": "to denote all skills in a set", + "type": "text" + }, + { + "bbox": [ + 416, + 447, + 424, + 456 + ], + "score": 0.83, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 446, + 492, + 457 + ], + "score": 1.0, + "content": "with cardinality", + "type": "text" + }, + { + "bbox": [ + 493, + 447, + 503, + 456 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 86, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 86, + 459, + 99, + 468 + ], + "score": 1.0, + "content": "130", + "type": "text" + }, + { + "bbox": [ + 106, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "Since (3) is a lower bound to MI, it tends to promote the same type of covering and directed skills.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 86, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 86, + 470, + 99, + 479 + ], + "score": 1.0, + "content": "131", + "type": "text" + }, + { + "bbox": [ + 105, + 468, + 469, + 481 + ], + "score": 1.0, + "content": "Furthermore, (2) no longer depends on the distribution over skills and the entropy term", + "type": "text" + }, + { + "bbox": [ + 469, + 468, + 493, + 480 + ], + "score": 0.91, + "content": "\\mathcal { H } ( Z )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 86, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 86, + 481, + 100, + 490 + ], + "score": 1.0, + "content": "132", + "type": "text" + }, + { + "bbox": [ + 105, + 478, + 194, + 492 + ], + "score": 1.0, + "content": "maximized by setting", + "type": "text" + }, + { + "bbox": [ + 195, + 481, + 201, + 491 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 478, + 330, + 492 + ], + "score": 1.0, + "content": "to the uniform distribution over", + "type": "text" + }, + { + "bbox": [ + 330, + 479, + 341, + 489 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 478, + 385, + 492 + ], + "score": 1.0, + "content": "skills (i.e.,", + "type": "text" + }, + { + "bbox": [ + 386, + 479, + 481, + 491 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\operatorname* { m a x } _ { \\rho } \\mathcal { H } ( Z ) \\bar { = } \\log ( N ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 478, + 506, + 492 + ], + "score": 1.0, + "content": ", thus", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 87, + 490, + 353, + 502 + ], + "spans": [ + { + "bbox": [ + 87, + 492, + 100, + 501 + ], + "score": 1.0, + "content": "133", + "type": "text" + }, + { + "bbox": [ + 105, + 490, + 339, + 502 + ], + "score": 1.0, + "content": "simplifying the optimization, which now only depends on", + "type": "text" + }, + { + "bbox": [ + 339, + 490, + 349, + 500 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 490, + 353, + 502 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 86, + 506, + 507, + 540 + ], + "lines": [ + { + "bbox": [ + 86, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 86, + 508, + 101, + 517 + ], + "score": 1.0, + "content": "134", + "type": "text" + }, + { + "bbox": [ + 104, + 505, + 282, + 519 + ], + "score": 1.0, + "content": "While optimizing (3) promotes a cardinality", + "type": "text" + }, + { + "bbox": [ + 282, + 507, + 292, + 516 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "such that all skills have good discriminability, a more", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 87, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 87, + 520, + 100, + 528 + ], + "score": 1.0, + "content": "135", + "type": "text" + }, + { + "bbox": [ + 103, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "convenient formulation is to explicitly set a minimum level of discriminability for all skills through", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 87, + 528, + 302, + 540 + ], + "spans": [ + { + "bbox": [ + 87, + 531, + 100, + 539 + ], + "score": 1.0, + "content": "136", + "type": "text" + }, + { + "bbox": [ + 105, + 528, + 302, + 540 + ], + "score": 1.0, + "content": "the following constrained optimization problem:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 546, + 425, + 566 + ], + "lines": [ + { + "bbox": [ + 187, + 546, + 425, + 566 + ], + "spans": [ + { + "bbox": [ + 187, + 546, + 425, + 566 + ], + "score": 0.9, + "content": "\\operatorname* { m a x } _ { N \\geq 1 } \\log ( N ) \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\log \\eta .", + "type": "interline_equation", + "image_path": "d4b44d35c1819e2e5cc983ddcde5fbbc82e0a5a49a11f7481ce5dd77a52ac2b5.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 187, + 546, + 425, + 566 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 86, + 572, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 86, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 86, + 575, + 100, + 585 + ], + "score": 1.0, + "content": "137", + "type": "text" + }, + { + "bbox": [ + 105, + 572, + 132, + 585 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 576, + 139, + 585 + ], + "score": 0.84, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 572, + 389, + 585 + ], + "score": 1.0, + "content": "is a parameter that defines the discriminability threshold. A skill", + "type": "text" + }, + { + "bbox": [ + 390, + 575, + 396, + 583 + ], + "score": 0.74, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 572, + 444, + 585 + ], + "score": 1.0, + "content": "is said to be", + "type": "text" + }, + { + "bbox": [ + 444, + 576, + 451, + 585 + ], + "score": 0.83, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "-consolidated", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 85, + 582, + 507, + 599 + ], + "spans": [ + { + "bbox": [ + 85, + 586, + 100, + 596 + ], + "score": 1.0, + "content": "138", + "type": "text" + }, + { + "bbox": [ + 103, + 582, + 277, + 599 + ], + "score": 1.0, + "content": "if it satisfies the constraint. Crucially, let", + "type": "text" + }, + { + "bbox": [ + 277, + 584, + 425, + 597 + ], + "score": 0.89, + "content": "\\begin{array} { r } { P _ { N } : = \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 582, + 507, + 599 + ], + "score": 1.0, + "content": ", then the sequence", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 85, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 85, + 597, + 100, + 608 + ], + "score": 1.0, + "content": "139", + "type": "text" + }, + { + "bbox": [ + 106, + 596, + 146, + 608 + ], + "score": 0.93, + "content": "( P _ { N } ) _ { N \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 596, + 238, + 609 + ], + "score": 1.0, + "content": "is non-increasing with", + "type": "text" + }, + { + "bbox": [ + 239, + 596, + 269, + 607 + ], + "score": 0.94, + "content": "P _ { 1 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "(i.e., the more skills the harder it is to meet the constraint).", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 85, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 85, + 609, + 100, + 619 + ], + "score": 1.0, + "content": "140", + "type": "text" + }, + { + "bbox": [ + 104, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "As a result, (4) can be optimized following a simple greedy strategy incrementally adding skills until", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 85, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 85, + 619, + 100, + 630 + ], + "score": 1.0, + "content": "141", + "type": "text" + }, + { + "bbox": [ + 106, + 618, + 255, + 630 + ], + "score": 1.0, + "content": "the constraint is violated. The optimal", + "type": "text" + }, + { + "bbox": [ + 256, + 618, + 266, + 628 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 618, + 406, + 630 + ], + "score": 1.0, + "content": "thus defines the effective number of", + "type": "text" + }, + { + "bbox": [ + 406, + 620, + 412, + 629 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "-consolidated skills and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 85, + 629, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 85, + 630, + 100, + 641 + ], + "score": 1.0, + "content": "142", + "type": "text" + }, + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "score": 1.0, + "content": "it corresponds to the largest number of skills that is guaranteed to display sufficient discriminability.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 85, + 639, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 85, + 641, + 100, + 652 + ], + "score": 1.0, + "content": "143", + "type": "text" + }, + { + "bbox": [ + 105, + 639, + 506, + 652 + ], + "score": 1.0, + "content": "Alternatively, we can interpret (4) as finding the largest number of clusters (i.e., the region reached", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 85, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 85, + 652, + 100, + 663 + ], + "score": 1.0, + "content": "144", + "type": "text" + }, + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "score": 1.0, + "content": "by the directed part of a skill and covered by its associated diffusing part) with a minimum level of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 85, + 660, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 85, + 663, + 100, + 673 + ], + "score": 1.0, + "content": "145", + "type": "text" + }, + { + "bbox": [ + 105, + 660, + 506, + 674 + ], + "score": 1.0, + "content": "inter-cluster distance. This effect is qualitatively illustrated in Fig. 1, where the states attained by the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 86, + 672, + 492, + 686 + ], + "spans": [ + { + "bbox": [ + 86, + 675, + 100, + 684 + ], + "score": 1.0, + "content": "146", + "type": "text" + }, + { + "bbox": [ + 105, + 672, + 492, + 686 + ], + "score": 1.0, + "content": "directed part of the skills attain different regions that are locally covered by their diffusing parts.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 691, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 690, + 507, + 704 + ], + "spans": [ + { + "bbox": [ + 118, + 690, + 272, + 704 + ], + "score": 1.0, + "content": "3Notice that (1) is maximized by setting", + "type": "text" + }, + { + "bbox": [ + 272, + 693, + 311, + 703 + ], + "score": 0.89, + "content": "| Z | = | S |", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 690, + 363, + 704 + ], + "score": 1.0, + "content": "(since maxY", + "type": "text" + }, + { + "bbox": [ + 363, + 691, + 486, + 703 + ], + "score": 0.87, + "content": "{ \\mathcal { T } } ( X , Y ) = { \\mathcal { T } } ( X , X ) = { \\mathcal { H } } ( X ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 690, + 507, + 704 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 505, + 713 + ], + "score": 1.0, + "content": "where each skills is a goal-conditioned policy reaching a different state. This implies having as many policies as", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 485, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 485, + 723 + ], + "score": 1.0, + "content": "states, which makes the learning particularly challenging as the complexity of the environment increases.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "index", + "bbox": [ + 87, + 72, + 505, + 117 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 86, + 72, + 505, + 118 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 105, + 127, + 271, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 272, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 272, + 142 + ], + "score": 1.0, + "content": "3.2 Skill Support and Sampling Rule", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "index", + "bbox": [ + 87, + 144, + 505, + 178 + ], + "lines": [ + { + "bbox": [ + 87, + 142, + 502, + 158 + ], + "spans": [ + { + "bbox": [ + 87, + 147, + 99, + 156 + ], + "score": 1.0, + "content": "110", + "type": "text" + }, + { + "bbox": [ + 102, + 142, + 375, + 158 + ], + "score": 1.0, + "content": "The MI objective (1) crucially depends on the number of skills", + "type": "text" + }, + { + "bbox": [ + 376, + 145, + 397, + 156 + ], + "score": 0.88, + "content": "( | Z | )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 142, + 483, + 158 + ], + "score": 1.0, + "content": "and the distribution", + "type": "text" + }, + { + "bbox": [ + 484, + 145, + 502, + 157 + ], + "score": 0.91, + "content": "\\rho ( z )", + "type": "inline_equation" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 87, + 159, + 99, + 167 + ], + "score": 1.0, + "content": "111", + "type": "text" + }, + { + "bbox": [ + 105, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "Unfortunately, it is been shown [e.g., 8] that solving (1) is particularly challenging. In order to", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 166, + 471, + 180 + ], + "spans": [ + { + "bbox": [ + 87, + 169, + 100, + 178 + ], + "score": 1.0, + "content": "112", + "type": "text" + }, + { + "bbox": [ + 104, + 166, + 471, + 180 + ], + "score": 1.0, + "content": "simplify the optimization and the associated learning problem, we modify (1) in two ways.", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 86, + 185, + 100, + 195 + ], + "score": 1.0, + "content": "113", + "type": "text" + }, + { + "bbox": [ + 105, + 183, + 468, + 195 + ], + "score": 1.0, + "content": "First, coherently with the skill optimization detailed in Sect. 3.1, the random variable", + "type": "text" + }, + { + "bbox": [ + 469, + 184, + 477, + 194 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 86, + 196, + 100, + 205 + ], + "score": 1.0, + "content": "114", + "type": "text" + }, + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "conditional entropy is any state reached during the diffusing part of the skill and not just the terminal", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 204, + 504, + 218 + ], + "spans": [ + { + "bbox": [ + 86, + 206, + 100, + 217 + ], + "score": 1.0, + "content": "115", + "type": "text" + }, + { + "bbox": [ + 105, + 204, + 248, + 218 + ], + "score": 1.0, + "content": "state. More formally, we denote by", + "type": "text" + }, + { + "bbox": [ + 248, + 205, + 266, + 216 + ], + "score": 0.87, + "content": "S _ { \\mathrm { d i f f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 204, + 497, + 218 + ], + "score": 1.0, + "content": "the random variable and its distribution for a specific skill", + "type": "text" + }, + { + "bbox": [ + 497, + 207, + 504, + 215 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 214, + 507, + 234 + ], + "spans": [ + { + "bbox": [ + 86, + 219, + 100, + 230 + ], + "score": 1.0, + "content": "116", + "type": "text" + }, + { + "bbox": [ + 104, + 214, + 116, + 234 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 216, + 269, + 231 + ], + "score": 0.92, + "content": "p _ { \\pi ( z ) } ( s _ { \\mathrm { d i f f } } ) = 1 / \\dot { H } \\sum _ { t = T + 1 } ^ { T + H } p _ { \\pi ( z ) } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 214, + 507, + 234 + ], + "score": 1.0, + "content": ", i.e., the distribution over states obtained by averaging the", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 85, + 231, + 100, + 242 + ], + "score": 1.0, + "content": "117", + "type": "text" + }, + { + "bbox": [ + 105, + 229, + 355, + 243 + ], + "score": 1.0, + "content": "distributions at any of the steps in the diffusing part. Similarly,", + "type": "text" + }, + { + "bbox": [ + 356, + 230, + 392, + 242 + ], + "score": 0.92, + "content": "p ( z | s _ { \\mathrm { d i f f } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "now denotes the probability", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 86, + 243, + 99, + 253 + ], + "score": 1.0, + "content": "118", + "type": "text" + }, + { + "bbox": [ + 105, + 240, + 117, + 254 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 243, + 123, + 251 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 240, + 224, + 254 + ], + "score": 1.0, + "content": "being the skill to traverse", + "type": "text" + }, + { + "bbox": [ + 224, + 243, + 241, + 253 + ], + "score": 0.87, + "content": "s _ { \\mathrm { d i f f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "during its diffusing part. As a result, training the skills to maximize", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 86, + 254, + 99, + 263 + ], + "score": 1.0, + "content": "119", + "type": "text" + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "MI naturally leads the diffusing parts to “push” the directed parts away so as to reach diverse regions", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 86, + 264, + 100, + 275 + ], + "score": 1.0, + "content": "120", + "type": "text" + }, + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "of the environment. The combination of “global” coverage of the directed parts and “local” coverage", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 272, + 493, + 286 + ], + "spans": [ + { + "bbox": [ + 86, + 276, + 99, + 285 + ], + "score": 1.0, + "content": "121", + "type": "text" + }, + { + "bbox": [ + 105, + 272, + 426, + 286 + ], + "score": 1.0, + "content": "of the diffusing part ensures that the whole environment is properly visited with", + "type": "text" + }, + { + "bbox": [ + 427, + 273, + 463, + 286 + ], + "score": 0.92, + "content": "| Z | \\ll S", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 272, + 493, + 286 + ], + "score": 1.0, + "content": "skills.3", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 86, + 292, + 100, + 301 + ], + "score": 1.0, + "content": "122", + "type": "text" + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "Second, we introduce an alternative problem that simplifies the optimization while preserving the", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 300, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 85, + 303, + 100, + 313 + ], + "score": 1.0, + "content": "123", + "type": "text" + }, + { + "bbox": [ + 105, + 300, + 506, + 314 + ], + "score": 1.0, + "content": "coverage and directedness properties of MI. This is achieved by introducing a stronger requirement", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 86, + 315, + 100, + 323 + ], + "score": 1.0, + "content": "124", + "type": "text" + }, + { + "bbox": [ + 105, + 311, + 348, + 325 + ], + "score": 1.0, + "content": "on the discriminability. While the conditional entropy term", + "type": "text" + }, + { + "bbox": [ + 348, + 312, + 389, + 324 + ], + "score": 0.91, + "content": "- { \\mathcal { H } } ( Z | S )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "in (1) promotes the discrim-", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 85, + 325, + 100, + 335 + ], + "score": 1.0, + "content": "125", + "type": "text" + }, + { + "bbox": [ + 104, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "inability of skills on average, we argue that a more suitable objective is to constrain each skill to", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 85, + 335, + 100, + 346 + ], + "score": 1.0, + "content": "126", + "type": "text" + }, + { + "bbox": [ + 106, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "achieve a minimum level of discriminability. First, we move from the average to the minimum over", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 344, + 314, + 358 + ], + "spans": [ + { + "bbox": [ + 86, + 347, + 100, + 357 + ], + "score": 1.0, + "content": "127", + "type": "text" + }, + { + "bbox": [ + 105, + 344, + 314, + 358 + ], + "score": 1.0, + "content": "skills by lower bounding the conditional entropy as", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + } + ], + "index": 6, + "bbox_fs": [ + 87, + 142, + 506, + 180 + ] + }, + { + "type": "index", + "bbox": [ + 86, + 183, + 505, + 286 + ], + "lines": [], + "index": 12, + "bbox_fs": [ + 85, + 183, + 507, + 286 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 86, + 290, + 505, + 357 + ], + "lines": [], + "index": 19.5, + "bbox_fs": [ + 85, + 290, + 506, + 358 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 362, + 454, + 390 + ], + "lines": [ + { + "bbox": [ + 155, + 362, + 454, + 390 + ], + "spans": [ + { + "bbox": [ + 155, + 362, + 454, + 390 + ], + "score": 0.91, + "content": "- \\mathcal { H } ( Z | S _ { \\mathrm { d i f f } } ) = \\sum _ { z \\in Z } \\rho ( z ) \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\operatorname* { m i n } _ { z \\in Z } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] ,", + "type": "interline_equation", + "image_path": "3157110c6c064eebeada077a520a63d063d2f20ec053138229dd1dff3acd2c56.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 155, + 362, + 454, + 371.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 155, + 371.3333333333333, + 454, + 380.66666666666663 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 155, + 380.66666666666663, + 454, + 389.99999999999994 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 87, + 396, + 427, + 408 + ], + "lines": [ + { + "bbox": [ + 85, + 395, + 428, + 410 + ], + "spans": [ + { + "bbox": [ + 85, + 395, + 318, + 410 + ], + "score": 1.0, + "content": "128 which leads to the following optimization (assuming", + "type": "text" + }, + { + "bbox": [ + 318, + 399, + 325, + 406 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 395, + 428, + 410 + ], + "score": 1.0, + "content": "is fixed for convenience)", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 85, + 395, + 428, + 410 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 414, + 398, + 438 + ], + "lines": [ + { + "bbox": [ + 212, + 414, + 398, + 438 + ], + "spans": [ + { + "bbox": [ + 212, + 414, + 398, + 438 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { | Z | = N , \\rho } \\bigg \\{ \\mathcal { H } ( Z ) + \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\bigg \\} ,", + "type": "interline_equation", + "image_path": "2f057bb1c2ed4a8288cb1135168369eb9b6e22ba2a4b80fd54634b61436bd511.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 212, + 414, + 398, + 438 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "index", + "bbox": [ + 86, + 445, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 86, + 446, + 503, + 458 + ], + "spans": [ + { + "bbox": [ + 86, + 448, + 100, + 457 + ], + "score": 1.0, + "content": "129", + "type": "text" + }, + { + "bbox": [ + 105, + 446, + 270, + 457 + ], + "score": 1.0, + "content": "where with an abuse of notation we use", + "type": "text" + }, + { + "bbox": [ + 271, + 446, + 304, + 458 + ], + "score": 0.92, + "content": "z \\in [ N ]", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 446, + 415, + 457 + ], + "score": 1.0, + "content": "to denote all skills in a set", + "type": "text" + }, + { + "bbox": [ + 416, + 447, + 424, + 456 + ], + "score": 0.83, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 446, + 492, + 457 + ], + "score": 1.0, + "content": "with cardinality", + "type": "text" + }, + { + "bbox": [ + 493, + 447, + 503, + 456 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 86, + 459, + 99, + 468 + ], + "score": 1.0, + "content": "130", + "type": "text" + }, + { + "bbox": [ + 106, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "Since (3) is a lower bound to MI, it tends to promote the same type of covering and directed skills.", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 86, + 470, + 99, + 479 + ], + "score": 1.0, + "content": "131", + "type": "text" + }, + { + "bbox": [ + 105, + 468, + 469, + 481 + ], + "score": 1.0, + "content": "Furthermore, (2) no longer depends on the distribution over skills and the entropy term", + "type": "text" + }, + { + "bbox": [ + 469, + 468, + 493, + 480 + ], + "score": 0.91, + "content": "\\mathcal { H } ( Z )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 86, + 481, + 100, + 490 + ], + "score": 1.0, + "content": "132", + "type": "text" + }, + { + "bbox": [ + 105, + 478, + 194, + 492 + ], + "score": 1.0, + "content": "maximized by setting", + "type": "text" + }, + { + "bbox": [ + 195, + 481, + 201, + 491 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 478, + 330, + 492 + ], + "score": 1.0, + "content": "to the uniform distribution over", + "type": "text" + }, + { + "bbox": [ + 330, + 479, + 341, + 489 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 478, + 385, + 492 + ], + "score": 1.0, + "content": "skills (i.e.,", + "type": "text" + }, + { + "bbox": [ + 386, + 479, + 481, + 491 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\operatorname* { m a x } _ { \\rho } \\mathcal { H } ( Z ) \\bar { = } \\log ( N ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 478, + 506, + 492 + ], + "score": 1.0, + "content": ", thus", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 490, + 353, + 502 + ], + "spans": [ + { + "bbox": [ + 87, + 492, + 100, + 501 + ], + "score": 1.0, + "content": "133", + "type": "text" + }, + { + "bbox": [ + 105, + 490, + 339, + 502 + ], + "score": 1.0, + "content": "simplifying the optimization, which now only depends on", + "type": "text" + }, + { + "bbox": [ + 339, + 490, + 349, + 500 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 490, + 353, + 502 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 86, + 508, + 101, + 517 + ], + "score": 1.0, + "content": "134", + "type": "text" + }, + { + "bbox": [ + 104, + 505, + 282, + 519 + ], + "score": 1.0, + "content": "While optimizing (3) promotes a cardinality", + "type": "text" + }, + { + "bbox": [ + 282, + 507, + 292, + 516 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "such that all skills have good discriminability, a more", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 87, + 520, + 100, + 528 + ], + "score": 1.0, + "content": "135", + "type": "text" + }, + { + "bbox": [ + 103, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "convenient formulation is to explicitly set a minimum level of discriminability for all skills through", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 528, + 302, + 540 + ], + "spans": [ + { + "bbox": [ + 87, + 531, + 100, + 539 + ], + "score": 1.0, + "content": "136", + "type": "text" + }, + { + "bbox": [ + 105, + 528, + 302, + 540 + ], + "score": 1.0, + "content": "the following constrained optimization problem:", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + } + ], + "index": 30, + "bbox_fs": [ + 86, + 446, + 506, + 502 + ] + }, + { + "type": "index", + "bbox": [ + 86, + 506, + 507, + 540 + ], + "lines": [], + "index": 34, + "bbox_fs": [ + 86, + 505, + 505, + 540 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 546, + 425, + 566 + ], + "lines": [ + { + "bbox": [ + 187, + 546, + 425, + 566 + ], + "spans": [ + { + "bbox": [ + 187, + 546, + 425, + 566 + ], + "score": 0.9, + "content": "\\operatorname* { m a x } _ { N \\geq 1 } \\log ( N ) \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\log \\eta .", + "type": "interline_equation", + "image_path": "d4b44d35c1819e2e5cc983ddcde5fbbc82e0a5a49a11f7481ce5dd77a52ac2b5.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 187, + 546, + 425, + 566 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "index", + "bbox": [ + 86, + 572, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 86, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 86, + 575, + 100, + 585 + ], + "score": 1.0, + "content": "137", + "type": "text" + }, + { + "bbox": [ + 105, + 572, + 132, + 585 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 576, + 139, + 585 + ], + "score": 0.84, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 572, + 389, + 585 + ], + "score": 1.0, + "content": "is a parameter that defines the discriminability threshold. A skill", + "type": "text" + }, + { + "bbox": [ + 390, + 575, + 396, + 583 + ], + "score": 0.74, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 572, + 444, + 585 + ], + "score": 1.0, + "content": "is said to be", + "type": "text" + }, + { + "bbox": [ + 444, + 576, + 451, + 585 + ], + "score": 0.83, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "-consolidated", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 582, + 507, + 599 + ], + "spans": [ + { + "bbox": [ + 85, + 586, + 100, + 596 + ], + "score": 1.0, + "content": "138", + "type": "text" + }, + { + "bbox": [ + 103, + 582, + 277, + 599 + ], + "score": 1.0, + "content": "if it satisfies the constraint. Crucially, let", + "type": "text" + }, + { + "bbox": [ + 277, + 584, + 425, + 597 + ], + "score": 0.89, + "content": "\\begin{array} { r } { P _ { N } : = \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 582, + 507, + 599 + ], + "score": 1.0, + "content": ", then the sequence", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 85, + 597, + 100, + 608 + ], + "score": 1.0, + "content": "139", + "type": "text" + }, + { + "bbox": [ + 106, + 596, + 146, + 608 + ], + "score": 0.93, + "content": "( P _ { N } ) _ { N \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 596, + 238, + 609 + ], + "score": 1.0, + "content": "is non-increasing with", + "type": "text" + }, + { + "bbox": [ + 239, + 596, + 269, + 607 + ], + "score": 0.94, + "content": "P _ { 1 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "(i.e., the more skills the harder it is to meet the constraint).", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 85, + 609, + 100, + 619 + ], + "score": 1.0, + "content": "140", + "type": "text" + }, + { + "bbox": [ + 104, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "As a result, (4) can be optimized following a simple greedy strategy incrementally adding skills until", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 85, + 619, + 100, + 630 + ], + "score": 1.0, + "content": "141", + "type": "text" + }, + { + "bbox": [ + 106, + 618, + 255, + 630 + ], + "score": 1.0, + "content": "the constraint is violated. The optimal", + "type": "text" + }, + { + "bbox": [ + 256, + 618, + 266, + 628 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 618, + 406, + 630 + ], + "score": 1.0, + "content": "thus defines the effective number of", + "type": "text" + }, + { + "bbox": [ + 406, + 620, + 412, + 629 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "-consolidated skills and", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 629, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 85, + 630, + 100, + 641 + ], + "score": 1.0, + "content": "142", + "type": "text" + }, + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "score": 1.0, + "content": "it corresponds to the largest number of skills that is guaranteed to display sufficient discriminability.", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 639, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 85, + 641, + 100, + 652 + ], + "score": 1.0, + "content": "143", + "type": "text" + }, + { + "bbox": [ + 105, + 639, + 506, + 652 + ], + "score": 1.0, + "content": "Alternatively, we can interpret (4) as finding the largest number of clusters (i.e., the region reached", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 85, + 652, + 100, + 663 + ], + "score": 1.0, + "content": "144", + "type": "text" + }, + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "score": 1.0, + "content": "by the directed part of a skill and covered by its associated diffusing part) with a minimum level of", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 660, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 85, + 663, + 100, + 673 + ], + "score": 1.0, + "content": "145", + "type": "text" + }, + { + "bbox": [ + 105, + 660, + 506, + 674 + ], + "score": 1.0, + "content": "inter-cluster distance. This effect is qualitatively illustrated in Fig. 1, where the states attained by the", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 672, + 492, + 686 + ], + "spans": [ + { + "bbox": [ + 86, + 675, + 100, + 684 + ], + "score": 1.0, + "content": "146", + "type": "text" + }, + { + "bbox": [ + 105, + 672, + 492, + 686 + ], + "score": 1.0, + "content": "directed part of the skills attain different regions that are locally covered by their diffusing parts.", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + } + ], + "index": 41.5, + "bbox_fs": [ + 85, + 572, + 507, + 686 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 75, + 197, + 87 + ], + "lines": [ + { + "bbox": [ + 106, + 74, + 198, + 88 + ], + "spans": [ + { + "bbox": [ + 106, + 74, + 198, + 88 + ], + "score": 1.0, + "content": "Algorithm 1: UPSIDE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 97, + 89, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 89, + 421, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 242, + 101 + ], + "score": 1.0, + "content": "Initialize: Discriminability threshold", + "type": "text" + }, + { + "bbox": [ + 243, + 90, + 280, + 101 + ], + "score": 0.91, + "content": "\\eta \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 89, + 345, + 101 + ], + "score": 1.0, + "content": ", branching factor", + "type": "text" + }, + { + "bbox": [ + 345, + 90, + 374, + 100 + ], + "score": 0.9, + "content": "N _ { 0 } \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 89, + 410, + 101 + ], + "score": 1.0, + "content": ", patience", + "type": "text" + }, + { + "bbox": [ + 410, + 91, + 419, + 99 + ], + "score": 0.74, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 89, + 421, + 101 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 98, + 436, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 164, + 111 + ], + "score": 1.0, + "content": "Initialize: Tree", + "type": "text" + }, + { + "bbox": [ + 164, + 100, + 173, + 109 + ], + "score": 0.81, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 98, + 396, + 111 + ], + "score": 1.0, + "content": "initialized as a root node indexed by 0, queue of parent nodes", + "type": "text" + }, + { + "bbox": [ + 396, + 99, + 433, + 110 + ], + "score": 0.91, + "content": "\\mathcal { W } = \\{ 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 98, + 436, + 111 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 108, + 255, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 130, + 121 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 131, + 109, + 159, + 120 + ], + "score": 0.77, + "content": "\\mathcal { W } \\neq \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 108, + 255, + 121 + ], + "score": 1.0, + "content": "do // tree expansion", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 119, + 447, + 131 + ], + "spans": [ + { + "bbox": [ + 121, + 119, + 200, + 131 + ], + "score": 1.0, + "content": "Dequeue a node/skill", + "type": "text" + }, + { + "bbox": [ + 200, + 120, + 230, + 129 + ], + "score": 0.89, + "content": "w \\in \\mathcal W", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 119, + 273, + 131 + ], + "score": 1.0, + "content": "and expand", + "type": "text" + }, + { + "bbox": [ + 273, + 120, + 282, + 129 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 119, + 291, + 131 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 291, + 122, + 299, + 129 + ], + "score": 0.78, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 119, + 357, + 131 + ], + "score": 1.0, + "content": "by adding a set", + "type": "text" + }, + { + "bbox": [ + 357, + 120, + 377, + 131 + ], + "score": 0.91, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 119, + 388, + 131 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 388, + 120, + 401, + 129 + ], + "score": 0.85, + "content": "N _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 119, + 447, + 131 + ], + "score": 1.0, + "content": "nodes/skills", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 129, + 264, + 141 + ], + "spans": [ + { + "bbox": [ + 99, + 132, + 104, + 138 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 120, + 129, + 207, + 141 + ], + "score": 1.0, + "content": "Create random policies", + "type": "text" + }, + { + "bbox": [ + 207, + 131, + 218, + 139 + ], + "score": 0.84, + "content": "\\pi _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 129, + 222, + 141 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 223, + 129, + 264, + 140 + ], + "score": 0.9, + "content": "\\forall z \\in \\mathcal { C } ( \\bar { \\boldsymbol { w } } )", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 139, + 278, + 150 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 104, + 149 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 121, + 139, + 206, + 150 + ], + "score": 1.0, + "content": "Initialize discriminator", + "type": "text" + }, + { + "bbox": [ + 206, + 141, + 216, + 150 + ], + "score": 0.87, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 139, + 235, + 150 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 236, + 140, + 249, + 150 + ], + "score": 0.88, + "content": "| \\tau |", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 139, + 278, + 150 + ], + "score": 1.0, + "content": "classes", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 149, + 284, + 160 + ], + "spans": [ + { + "bbox": [ + 99, + 153, + 104, + 159 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 121, + 149, + 163, + 160 + ], + "score": 1.0, + "content": "Continue", + "type": "text" + }, + { + "bbox": [ + 163, + 151, + 172, + 158 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 149, + 248, + 160 + ], + "score": 1.0, + "content": "true; Saturated", + "type": "text" + }, + { + "bbox": [ + 248, + 151, + 256, + 158 + ], + "score": 0.78, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 149, + 284, + 160 + ], + "score": 1.0, + "content": "false", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 159, + 196, + 169 + ], + "spans": [ + { + "bbox": [ + 99, + 162, + 104, + 169 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 121, + 159, + 196, + 169 + ], + "score": 1.0, + "content": "while Continue do", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 169, + 209, + 180 + ], + "spans": [ + { + "bbox": [ + 99, + 172, + 105, + 180 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 136, + 169, + 150, + 180 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 151, + 169, + 160, + 178 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 169, + 209, + 180 + ], + "score": 1.0, + "content": "iterations do", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 99, + 179, + 281, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 181, + 106, + 189 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 151, + 179, + 205, + 190 + ], + "score": 1.0, + "content": "Sample a skill", + "type": "text" + }, + { + "bbox": [ + 205, + 181, + 212, + 189 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 179, + 232, + 190 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 232, + 180, + 241, + 189 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 179, + 281, + 190 + ], + "score": 1.0, + "content": "at random", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 187, + 366, + 203 + ], + "spans": [ + { + "bbox": [ + 99, + 191, + 106, + 199 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 150, + 187, + 262, + 203 + ], + "score": 1.0, + "content": "Extract the sequence of nodes", + "type": "text" + }, + { + "bbox": [ + 262, + 191, + 302, + 200 + ], + "score": 0.89, + "content": "z _ { ( 1 ) } , \\dotsc , z", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 187, + 312, + 203 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 313, + 190, + 322, + 198 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 187, + 360, + 203 + ], + "score": 1.0, + "content": "leading to", + "type": "text" + }, + { + "bbox": [ + 361, + 192, + 366, + 198 + ], + "score": 0.7, + "content": "z", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 198, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 99, + 201, + 106, + 209 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 150, + 198, + 313, + 213 + ], + "score": 1.0, + "content": "Execute the composed (directed part) policy", + "type": "text" + }, + { + "bbox": [ + 314, + 200, + 370, + 212 + ], + "score": 0.91, + "content": "( \\pi _ { z _ { ( 1 ) } } , \\ldots , \\pi _ { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 198, + 481, + 213 + ], + "score": 1.0, + "content": "followed by the diffusing part", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 96, + 211, + 379, + 222 + ], + "spans": [ + { + "bbox": [ + 96, + 212, + 106, + 222 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 152, + 211, + 368, + 222 + ], + "score": 1.0, + "content": "Add states observed during the diffusion part to state buffer", + "type": "text" + }, + { + "bbox": [ + 368, + 212, + 379, + 221 + ], + "score": 0.86, + "content": "B _ { z }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 96, + 221, + 366, + 232 + ], + "spans": [ + { + "bbox": [ + 96, + 222, + 106, + 232 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 152, + 221, + 231, + 232 + ], + "score": 1.0, + "content": "Update discriminator", + "type": "text" + }, + { + "bbox": [ + 231, + 223, + 241, + 232 + ], + "score": 0.86, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 221, + 291, + 232 + ], + "score": 1.0, + "content": "with SGD on", + "type": "text" + }, + { + "bbox": [ + 291, + 222, + 303, + 231 + ], + "score": 0.88, + "content": "B _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 221, + 360, + 232 + ], + "score": 1.0, + "content": "to predict label", + "type": "text" + }, + { + "bbox": [ + 360, + 223, + 366, + 230 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 96, + 230, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 96, + 232, + 106, + 242 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 151, + 230, + 160, + 243 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 160, + 231, + 196, + 241 + ], + "score": 0.91, + "content": "z \\in \\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 230, + 217, + 243 + ], + "score": 1.0, + "content": "then", + "type": "text" + }, + { + "bbox": [ + 217, + 231, + 228, + 241 + ], + "score": 0.44, + "content": "/ /", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 230, + 469, + 243 + ], + "score": 1.0, + "content": "Update only new policies, other polices kept fixed", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 96, + 240, + 460, + 253 + ], + "spans": [ + { + "bbox": [ + 96, + 242, + 107, + 253 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 166, + 240, + 220, + 252 + ], + "score": 1.0, + "content": "Update policy", + "type": "text" + }, + { + "bbox": [ + 221, + 243, + 231, + 251 + ], + "score": 0.85, + "content": "\\pi _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 240, + 460, + 252 + ], + "score": 1.0, + "content": "using SAC to optimize the discriminator reward as in Sect. 3.1.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 96, + 251, + 468, + 266 + ], + "spans": [ + { + "bbox": [ + 96, + 254, + 105, + 264 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 136, + 251, + 263, + 266 + ], + "score": 1.0, + "content": "Compute the skill-discriminability", + "type": "text" + }, + { + "bbox": [ + 263, + 251, + 406, + 266 + ], + "score": 0.93, + "content": "\\begin{array} { r } { d ( z ) = \\widehat { q } _ { \\phi } ^ { \\scriptscriptstyle ( B ) } ( z ) = \\frac { 1 } { | \\mathscr { B } _ { z } | } \\sum _ { s \\in \\mathscr { B } _ { z } } q _ { \\phi } ( z | s ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 251, + 431, + 266 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 431, + 252, + 468, + 263 + ], + "score": 0.91, + "content": "z \\in \\mathcal { C } ( w )", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 96, + 263, + 318, + 278 + ], + "spans": [ + { + "bbox": [ + 96, + 266, + 105, + 276 + ], + "score": 1.0, + "content": "15", + "type": "text" + }, + { + "bbox": [ + 135, + 263, + 147, + 278 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 147, + 264, + 224, + 276 + ], + "score": 0.8, + "content": "\\begin{array} { r } { \\operatorname* { m i n } _ { z \\in \\mathcal { C } ( w ) } d ( z ) < \\eta } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 263, + 244, + 278 + ], + "score": 1.0, + "content": "then", + "type": "text" + }, + { + "bbox": [ + 244, + 265, + 255, + 275 + ], + "score": 0.44, + "content": "/ /", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 263, + 318, + 278 + ], + "score": 1.0, + "content": "Node removal", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 96, + 273, + 393, + 288 + ], + "spans": [ + { + "bbox": [ + 96, + 277, + 106, + 286 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 150, + 273, + 235, + 288 + ], + "score": 1.0, + "content": "Remove the node/skill", + "type": "text" + }, + { + "bbox": [ + 235, + 276, + 326, + 287 + ], + "score": 0.82, + "content": "z = \\arg \\operatorname* { m i n } _ { z \\in { \\mathcal { C } } ( w ) } d ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 273, + 347, + 288 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 347, + 275, + 367, + 286 + ], + "score": 0.9, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 273, + 383, + 288 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 275, + 393, + 285 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 95, + 284, + 239, + 297 + ], + "spans": [ + { + "bbox": [ + 95, + 286, + 106, + 297 + ], + "score": 1.0, + "content": "17", + "type": "text" + }, + { + "bbox": [ + 152, + 284, + 207, + 296 + ], + "score": 1.0, + "content": "Set Saturate", + "type": "text" + }, + { + "bbox": [ + 207, + 286, + 216, + 294 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 284, + 239, + 296 + ], + "score": 1.0, + "content": "true", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 96, + 295, + 244, + 307 + ], + "spans": [ + { + "bbox": [ + 96, + 297, + 105, + 306 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 136, + 295, + 244, + 307 + ], + "score": 1.0, + "content": "else if not Saturated then", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 95, + 304, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 95, + 307, + 106, + 316 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 151, + 304, + 249, + 317 + ], + "score": 1.0, + "content": "Add one new node/skill to", + "type": "text" + }, + { + "bbox": [ + 249, + 306, + 270, + 317 + ], + "score": 0.9, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 304, + 286, + 317 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 306, + 295, + 315 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 95, + 316, + 154, + 326 + ], + "spans": [ + { + "bbox": [ + 95, + 317, + 105, + 326 + ], + "score": 1.0, + "content": "20", + "type": "text" + }, + { + "bbox": [ + 137, + 316, + 154, + 326 + ], + "score": 1.0, + "content": "else", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 95, + 326, + 244, + 338 + ], + "spans": [ + { + "bbox": [ + 95, + 327, + 106, + 338 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 151, + 326, + 207, + 337 + ], + "score": 1.0, + "content": "Set Continue", + "type": "text" + }, + { + "bbox": [ + 208, + 327, + 216, + 335 + ], + "score": 0.58, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 326, + 244, + 337 + ], + "score": 1.0, + "content": "false", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 95, + 339, + 281, + 352 + ], + "spans": [ + { + "bbox": [ + 95, + 341, + 105, + 352 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 120, + 339, + 164, + 352 + ], + "score": 1.0, + "content": "Enqueue in", + "type": "text" + }, + { + "bbox": [ + 164, + 340, + 176, + 350 + ], + "score": 0.6, + "content": "\\mathcal { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 339, + 261, + 352 + ], + "score": 1.0, + "content": "the consolidated nodes", + "type": "text" + }, + { + "bbox": [ + 261, + 340, + 281, + 352 + ], + "score": 0.87, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + } + ], + "index": 25 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 90, + 378, + 288, + 390 + ], + "lines": [ + { + "bbox": [ + 88, + 377, + 289, + 394 + ], + "spans": [ + { + "bbox": [ + 88, + 377, + 289, + 394 + ], + "score": 1.0, + "content": "147 3.3 Composing Skills in a Tree Structure", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 105, + 395, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 504, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 493, + 408 + ], + "score": 1.0, + "content": "The MI optimization problem as well as our constrained variant (4) depend on the initial state", + "type": "text" + }, + { + "bbox": [ + 494, + 398, + 504, + 407 + ], + "score": 0.82, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "and on the length of each skill. Although these quantities are usually predefined and only appear", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "implicitly in the equations, they have a crucial impact on the obtained behavior. In fact, resetting after", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 426, + 503, + 442 + ], + "spans": [ + { + "bbox": [ + 104, + 426, + 409, + 442 + ], + "score": 1.0, + "content": "each skill execution unavoidably restricts the coverage to a radius of at most", + "type": "text" + }, + { + "bbox": [ + 410, + 428, + 439, + 439 + ], + "score": 0.91, + "content": "T + H", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 426, + 492, + 442 + ], + "score": 1.0, + "content": "steps around", + "type": "text" + }, + { + "bbox": [ + 493, + 430, + 503, + 439 + ], + "score": 0.83, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 437, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 437, + 200, + 453 + ], + "score": 1.0, + "content": "This may suggest to set", + "type": "text" + }, + { + "bbox": [ + 200, + 439, + 209, + 449 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 437, + 226, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 226, + 439, + 236, + 449 + ], + "score": 0.83, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 437, + 506, + 453 + ], + "score": 1.0, + "content": "to a large value. However, increasing the horizon makes the training", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 450, + 490, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 276, + 463 + ], + "score": 1.0, + "content": "of the skills more challenging, as learning", + "type": "text" + }, + { + "bbox": [ + 276, + 452, + 283, + 460 + ], + "score": 0.71, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 450, + 490, + 463 + ], + "score": 1.0, + "content": "would require solving a difficult RL problem itself.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 103, + 466, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "Instead, we propose to “extend” the length of the skills through composition. Indeed, the decoupled", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 421, + 489 + ], + "score": 1.0, + "content": "skill structure and the constraint in (4) entail that the directed part of each of the", + "type": "text" + }, + { + "bbox": [ + 421, + 479, + 428, + 489 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "-consolidated skills", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "reliably reach a specific (and distinct) region of the environment and it is thus re-usable and amenable", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "to composition. We propose to chain the directed part of the skills in order to reach further and further", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 476, + 523 + ], + "score": 1.0, + "content": "parts of the state space. Specifically, we build a growing tree, where the root is the initial state", + "type": "text" + }, + { + "bbox": [ + 477, + 511, + 487, + 521 + ], + "score": 0.84, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 510, + 506, + 523 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "edges represent the directed part of the skills, and the nodes represent the diffusing part of skills. As", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 197, + 545 + ], + "score": 1.0, + "content": "such, whenever a skill", + "type": "text" + }, + { + "bbox": [ + 197, + 533, + 204, + 542 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 531, + 506, + 545 + ], + "score": 1.0, + "content": "is selected, the directed part of all the policies associated to its predecessor", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 542, + 444, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 444, + 555 + ], + "score": 1.0, + "content": "skills in the tree are executed first (see Fig. 1 for an illustration of the tree structure).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "As a result, the agent naturally builds a curriculum on the episode lengths, which grow as the sequence", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 158, + 582 + ], + "score": 0.9, + "content": "( i T + H ) _ { i \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 569, + 505, + 583 + ], + "score": 1.0, + "content": ". As such, it does not require prior knowledge on an adequate horizon of the downstream", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 316, + 596 + ], + "score": 1.0, + "content": "goal-based task.4 Here this knowledge is replaced by", + "type": "text" + }, + { + "bbox": [ + 317, + 582, + 325, + 592 + ], + "score": 0.77, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 581, + 343, + 596 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 343, + 582, + 353, + 592 + ], + "score": 0.76, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 581, + 505, + 596 + ], + "score": 1.0, + "content": "which are more environment-agnostic", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "and task-agnostic quantities, as their choice rather has an impact on the size and shape of the learned", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 604, + 312, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 193, + 617 + ], + "score": 1.0, + "content": "tree (e.g., the smaller", + "type": "text" + }, + { + "bbox": [ + 193, + 604, + 202, + 614 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 604, + 219, + 617 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 604, + 230, + 614 + ], + "score": 0.82, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 604, + 312, + 617 + ], + "score": 1.0, + "content": "the bigger the tree).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 108, + 626, + 228, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 228, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 228, + 639 + ], + "score": 1.0, + "content": "3.4 The UPSIDE Algorithm", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 105, + 642, + 503, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "We are now ready to introduce UPSIDE, which provides a specific implementation of the components", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "described before (see Fig. 1 for a qualitative illustration and Algorithm 1 for the detailed pseudo-code).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + }, + { + "type": "text", + "bbox": [ + 97, + 669, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 95, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 95, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "0 We perform standard approximations to make the constraint in (4) easier to estimate. We approximate", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 203, + 694 + ], + "score": 1.0, + "content": "the unknown posterior", + "type": "text" + }, + { + "bbox": [ + 203, + 681, + 229, + 692 + ], + "score": 0.92, + "content": "p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 679, + 352, + 694 + ], + "score": 1.0, + "content": "with a learned discriminator", + "type": "text" + }, + { + "bbox": [ + 353, + 681, + 383, + 693 + ], + "score": 0.93, + "content": "q _ { \\phi } ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 679, + 455, + 694 + ], + "score": 1.0, + "content": "with parameters", + "type": "text" + }, + { + "bbox": [ + 455, + 681, + 462, + 691 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 679, + 506, + 694 + ], + "score": 1.0, + "content": ". We also", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 702, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 118, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "4See e.g., the discussion in [26] on the “importance of properly choosing the training horizon in accordance", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 347, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 347, + 723 + ], + "score": 1.0, + "content": "with the downstream-task horizon the policy will eventually face.”", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 75, + 197, + 87 + ], + "lines": [ + { + "bbox": [ + 106, + 74, + 198, + 88 + ], + "spans": [ + { + "bbox": [ + 106, + 74, + 198, + 88 + ], + "score": 1.0, + "content": "Algorithm 1: UPSIDE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "index", + "bbox": [ + 97, + 89, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 89, + 421, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 242, + 101 + ], + "score": 1.0, + "content": "Initialize: Discriminability threshold", + "type": "text" + }, + { + "bbox": [ + 243, + 90, + 280, + 101 + ], + "score": 0.91, + "content": "\\eta \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 89, + 345, + 101 + ], + "score": 1.0, + "content": ", branching factor", + "type": "text" + }, + { + "bbox": [ + 345, + 90, + 374, + 100 + ], + "score": 0.9, + "content": "N _ { 0 } \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 89, + 410, + 101 + ], + "score": 1.0, + "content": ", patience", + "type": "text" + }, + { + "bbox": [ + 410, + 91, + 419, + 99 + ], + "score": 0.74, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 89, + 421, + 101 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 98, + 436, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 164, + 111 + ], + "score": 1.0, + "content": "Initialize: Tree", + "type": "text" + }, + { + "bbox": [ + 164, + 100, + 173, + 109 + ], + "score": 0.81, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 98, + 396, + 111 + ], + "score": 1.0, + "content": "initialized as a root node indexed by 0, queue of parent nodes", + "type": "text" + }, + { + "bbox": [ + 396, + 99, + 433, + 110 + ], + "score": 0.91, + "content": "\\mathcal { W } = \\{ 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 98, + 436, + 111 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 108, + 255, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 130, + 121 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 131, + 109, + 159, + 120 + ], + "score": 0.77, + "content": "\\mathcal { W } \\neq \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 108, + 255, + 121 + ], + "score": 1.0, + "content": "do // tree expansion", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 121, + 119, + 447, + 131 + ], + "spans": [ + { + "bbox": [ + 121, + 119, + 200, + 131 + ], + "score": 1.0, + "content": "Dequeue a node/skill", + "type": "text" + }, + { + "bbox": [ + 200, + 120, + 230, + 129 + ], + "score": 0.89, + "content": "w \\in \\mathcal W", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 119, + 273, + 131 + ], + "score": 1.0, + "content": "and expand", + "type": "text" + }, + { + "bbox": [ + 273, + 120, + 282, + 129 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 119, + 291, + 131 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 291, + 122, + 299, + 129 + ], + "score": 0.78, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 119, + 357, + 131 + ], + "score": 1.0, + "content": "by adding a set", + "type": "text" + }, + { + "bbox": [ + 357, + 120, + 377, + 131 + ], + "score": 0.91, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 119, + 388, + 131 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 388, + 120, + 401, + 129 + ], + "score": 0.85, + "content": "N _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 119, + 447, + 131 + ], + "score": 1.0, + "content": "nodes/skills", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 129, + 264, + 141 + ], + "spans": [ + { + "bbox": [ + 99, + 132, + 104, + 138 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 120, + 129, + 207, + 141 + ], + "score": 1.0, + "content": "Create random policies", + "type": "text" + }, + { + "bbox": [ + 207, + 131, + 218, + 139 + ], + "score": 0.84, + "content": "\\pi _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 129, + 222, + 141 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 223, + 129, + 264, + 140 + ], + "score": 0.9, + "content": "\\forall z \\in \\mathcal { C } ( \\bar { \\boldsymbol { w } } )", + "type": "inline_equation" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 139, + 278, + 150 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 104, + 149 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 121, + 139, + 206, + 150 + ], + "score": 1.0, + "content": "Initialize discriminator", + "type": "text" + }, + { + "bbox": [ + 206, + 141, + 216, + 150 + ], + "score": 0.87, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 139, + 235, + 150 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 236, + 140, + 249, + 150 + ], + "score": 0.88, + "content": "| \\tau |", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 139, + 278, + 150 + ], + "score": 1.0, + "content": "classes", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 149, + 284, + 160 + ], + "spans": [ + { + "bbox": [ + 99, + 153, + 104, + 159 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 121, + 149, + 163, + 160 + ], + "score": 1.0, + "content": "Continue", + "type": "text" + }, + { + "bbox": [ + 163, + 151, + 172, + 158 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 149, + 248, + 160 + ], + "score": 1.0, + "content": "true; Saturated", + "type": "text" + }, + { + "bbox": [ + 248, + 151, + 256, + 158 + ], + "score": 0.78, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 149, + 284, + 160 + ], + "score": 1.0, + "content": "false", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 159, + 196, + 169 + ], + "spans": [ + { + "bbox": [ + 99, + 162, + 104, + 169 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 121, + 159, + 196, + 169 + ], + "score": 1.0, + "content": "while Continue do", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 169, + 209, + 180 + ], + "spans": [ + { + "bbox": [ + 99, + 172, + 105, + 180 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 136, + 169, + 150, + 180 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 151, + 169, + 160, + 178 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 169, + 209, + 180 + ], + "score": 1.0, + "content": "iterations do", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 179, + 281, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 181, + 106, + 189 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 151, + 179, + 205, + 190 + ], + "score": 1.0, + "content": "Sample a skill", + "type": "text" + }, + { + "bbox": [ + 205, + 181, + 212, + 189 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 179, + 232, + 190 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 232, + 180, + 241, + 189 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 179, + 281, + 190 + ], + "score": 1.0, + "content": "at random", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 187, + 366, + 203 + ], + "spans": [ + { + "bbox": [ + 99, + 191, + 106, + 199 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 150, + 187, + 262, + 203 + ], + "score": 1.0, + "content": "Extract the sequence of nodes", + "type": "text" + }, + { + "bbox": [ + 262, + 191, + 302, + 200 + ], + "score": 0.89, + "content": "z _ { ( 1 ) } , \\dotsc , z", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 187, + 312, + 203 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 313, + 190, + 322, + 198 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 187, + 360, + 203 + ], + "score": 1.0, + "content": "leading to", + "type": "text" + }, + { + "bbox": [ + 361, + 192, + 366, + 198 + ], + "score": 0.7, + "content": "z", + "type": "inline_equation" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 198, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 99, + 201, + 106, + 209 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 150, + 198, + 313, + 213 + ], + "score": 1.0, + "content": "Execute the composed (directed part) policy", + "type": "text" + }, + { + "bbox": [ + 314, + 200, + 370, + 212 + ], + "score": 0.91, + "content": "( \\pi _ { z _ { ( 1 ) } } , \\ldots , \\pi _ { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 198, + 481, + 213 + ], + "score": 1.0, + "content": "followed by the diffusing part", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 211, + 379, + 222 + ], + "spans": [ + { + "bbox": [ + 96, + 212, + 106, + 222 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 152, + 211, + 368, + 222 + ], + "score": 1.0, + "content": "Add states observed during the diffusion part to state buffer", + "type": "text" + }, + { + "bbox": [ + 368, + 212, + 379, + 221 + ], + "score": 0.86, + "content": "B _ { z }", + "type": "inline_equation" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 221, + 366, + 232 + ], + "spans": [ + { + "bbox": [ + 96, + 222, + 106, + 232 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 152, + 221, + 231, + 232 + ], + "score": 1.0, + "content": "Update discriminator", + "type": "text" + }, + { + "bbox": [ + 231, + 223, + 241, + 232 + ], + "score": 0.86, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 221, + 291, + 232 + ], + "score": 1.0, + "content": "with SGD on", + "type": "text" + }, + { + "bbox": [ + 291, + 222, + 303, + 231 + ], + "score": 0.88, + "content": "B _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 221, + 360, + 232 + ], + "score": 1.0, + "content": "to predict label", + "type": "text" + }, + { + "bbox": [ + 360, + 223, + 366, + 230 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 230, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 96, + 232, + 106, + 242 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 151, + 230, + 160, + 243 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 160, + 231, + 196, + 241 + ], + "score": 0.91, + "content": "z \\in \\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 230, + 217, + 243 + ], + "score": 1.0, + "content": "then", + "type": "text" + }, + { + "bbox": [ + 217, + 231, + 228, + 241 + ], + "score": 0.44, + "content": "/ /", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 230, + 469, + 243 + ], + "score": 1.0, + "content": "Update only new policies, other polices kept fixed", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 240, + 460, + 253 + ], + "spans": [ + { + "bbox": [ + 96, + 242, + 107, + 253 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 166, + 240, + 220, + 252 + ], + "score": 1.0, + "content": "Update policy", + "type": "text" + }, + { + "bbox": [ + 221, + 243, + 231, + 251 + ], + "score": 0.85, + "content": "\\pi _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 240, + 460, + 252 + ], + "score": 1.0, + "content": "using SAC to optimize the discriminator reward as in Sect. 3.1.", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 251, + 468, + 266 + ], + "spans": [ + { + "bbox": [ + 96, + 254, + 105, + 264 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 136, + 251, + 263, + 266 + ], + "score": 1.0, + "content": "Compute the skill-discriminability", + "type": "text" + }, + { + "bbox": [ + 263, + 251, + 406, + 266 + ], + "score": 0.93, + "content": "\\begin{array} { r } { d ( z ) = \\widehat { q } _ { \\phi } ^ { \\scriptscriptstyle ( B ) } ( z ) = \\frac { 1 } { | \\mathscr { B } _ { z } | } \\sum _ { s \\in \\mathscr { B } _ { z } } q _ { \\phi } ( z | s ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 251, + 431, + 266 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 431, + 252, + 468, + 263 + ], + "score": 0.91, + "content": "z \\in \\mathcal { C } ( w )", + "type": "inline_equation" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 263, + 318, + 278 + ], + "spans": [ + { + "bbox": [ + 96, + 266, + 105, + 276 + ], + "score": 1.0, + "content": "15", + "type": "text" + }, + { + "bbox": [ + 135, + 263, + 147, + 278 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 147, + 264, + 224, + 276 + ], + "score": 0.8, + "content": "\\begin{array} { r } { \\operatorname* { m i n } _ { z \\in \\mathcal { C } ( w ) } d ( z ) < \\eta } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 263, + 244, + 278 + ], + "score": 1.0, + "content": "then", + "type": "text" + }, + { + "bbox": [ + 244, + 265, + 255, + 275 + ], + "score": 0.44, + "content": "/ /", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 263, + 318, + 278 + ], + "score": 1.0, + "content": "Node removal", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 273, + 393, + 288 + ], + "spans": [ + { + "bbox": [ + 96, + 277, + 106, + 286 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 150, + 273, + 235, + 288 + ], + "score": 1.0, + "content": "Remove the node/skill", + "type": "text" + }, + { + "bbox": [ + 235, + 276, + 326, + 287 + ], + "score": 0.82, + "content": "z = \\arg \\operatorname* { m i n } _ { z \\in { \\mathcal { C } } ( w ) } d ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 273, + 347, + 288 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 347, + 275, + 367, + 286 + ], + "score": 0.9, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 273, + 383, + 288 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 275, + 393, + 285 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 284, + 239, + 297 + ], + "spans": [ + { + "bbox": [ + 95, + 286, + 106, + 297 + ], + "score": 1.0, + "content": "17", + "type": "text" + }, + { + "bbox": [ + 152, + 284, + 207, + 296 + ], + "score": 1.0, + "content": "Set Saturate", + "type": "text" + }, + { + "bbox": [ + 207, + 286, + 216, + 294 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 284, + 239, + 296 + ], + "score": 1.0, + "content": "true", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 295, + 244, + 307 + ], + "spans": [ + { + "bbox": [ + 96, + 297, + 105, + 306 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 136, + 295, + 244, + 307 + ], + "score": 1.0, + "content": "else if not Saturated then", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 304, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 95, + 307, + 106, + 316 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 151, + 304, + 249, + 317 + ], + "score": 1.0, + "content": "Add one new node/skill to", + "type": "text" + }, + { + "bbox": [ + 249, + 306, + 270, + 317 + ], + "score": 0.9, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 304, + 286, + 317 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 306, + 295, + 315 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 316, + 154, + 326 + ], + "spans": [ + { + "bbox": [ + 95, + 317, + 105, + 326 + ], + "score": 1.0, + "content": "20", + "type": "text" + }, + { + "bbox": [ + 137, + 316, + 154, + 326 + ], + "score": 1.0, + "content": "else", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 326, + 244, + 338 + ], + "spans": [ + { + "bbox": [ + 95, + 327, + 106, + 338 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 151, + 326, + 207, + 337 + ], + "score": 1.0, + "content": "Set Continue", + "type": "text" + }, + { + "bbox": [ + 208, + 327, + 216, + 335 + ], + "score": 0.58, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 326, + 244, + 337 + ], + "score": 1.0, + "content": "false", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 339, + 281, + 352 + ], + "spans": [ + { + "bbox": [ + 95, + 341, + 105, + 352 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 120, + 339, + 164, + 352 + ], + "score": 1.0, + "content": "Enqueue in", + "type": "text" + }, + { + "bbox": [ + 164, + 340, + 176, + 350 + ], + "score": 0.6, + "content": "\\mathcal { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 339, + 261, + 352 + ], + "score": 1.0, + "content": "the consolidated nodes", + "type": "text" + }, + { + "bbox": [ + 261, + 340, + 281, + 352 + ], + "score": 0.87, + "content": "\\mathcal { C } ( w )", + "type": "inline_equation" + } + ], + "index": 25, + "is_list_start_line": true + } + ], + "index": 13, + "bbox_fs": [ + 95, + 89, + 481, + 352 + ] + }, + { + "type": "title", + "bbox": [ + 90, + 378, + 288, + 390 + ], + "lines": [ + { + "bbox": [ + 88, + 377, + 289, + 394 + ], + "spans": [ + { + "bbox": [ + 88, + 377, + 289, + 394 + ], + "score": 1.0, + "content": "147 3.3 Composing Skills in a Tree Structure", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 105, + 395, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 504, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 493, + 408 + ], + "score": 1.0, + "content": "The MI optimization problem as well as our constrained variant (4) depend on the initial state", + "type": "text" + }, + { + "bbox": [ + 494, + 398, + 504, + 407 + ], + "score": 0.82, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "and on the length of each skill. Although these quantities are usually predefined and only appear", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "implicitly in the equations, they have a crucial impact on the obtained behavior. In fact, resetting after", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 426, + 503, + 442 + ], + "spans": [ + { + "bbox": [ + 104, + 426, + 409, + 442 + ], + "score": 1.0, + "content": "each skill execution unavoidably restricts the coverage to a radius of at most", + "type": "text" + }, + { + "bbox": [ + 410, + 428, + 439, + 439 + ], + "score": 0.91, + "content": "T + H", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 426, + 492, + 442 + ], + "score": 1.0, + "content": "steps around", + "type": "text" + }, + { + "bbox": [ + 493, + 430, + 503, + 439 + ], + "score": 0.83, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 437, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 437, + 200, + 453 + ], + "score": 1.0, + "content": "This may suggest to set", + "type": "text" + }, + { + "bbox": [ + 200, + 439, + 209, + 449 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 437, + 226, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 226, + 439, + 236, + 449 + ], + "score": 0.83, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 437, + 506, + 453 + ], + "score": 1.0, + "content": "to a large value. However, increasing the horizon makes the training", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 450, + 490, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 276, + 463 + ], + "score": 1.0, + "content": "of the skills more challenging, as learning", + "type": "text" + }, + { + "bbox": [ + 276, + 452, + 283, + 460 + ], + "score": 0.71, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 450, + 490, + 463 + ], + "score": 1.0, + "content": "would require solving a difficult RL problem itself.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 394, + 506, + 463 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 466, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "Instead, we propose to “extend” the length of the skills through composition. Indeed, the decoupled", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 421, + 489 + ], + "score": 1.0, + "content": "skill structure and the constraint in (4) entail that the directed part of each of the", + "type": "text" + }, + { + "bbox": [ + 421, + 479, + 428, + 489 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "-consolidated skills", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "reliably reach a specific (and distinct) region of the environment and it is thus re-usable and amenable", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "to composition. We propose to chain the directed part of the skills in order to reach further and further", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 476, + 523 + ], + "score": 1.0, + "content": "parts of the state space. Specifically, we build a growing tree, where the root is the initial state", + "type": "text" + }, + { + "bbox": [ + 477, + 511, + 487, + 521 + ], + "score": 0.84, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 510, + 506, + 523 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "edges represent the directed part of the skills, and the nodes represent the diffusing part of skills. As", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 197, + 545 + ], + "score": 1.0, + "content": "such, whenever a skill", + "type": "text" + }, + { + "bbox": [ + 197, + 533, + 204, + 542 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 531, + 506, + 545 + ], + "score": 1.0, + "content": "is selected, the directed part of all the policies associated to its predecessor", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 542, + 444, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 444, + 555 + ], + "score": 1.0, + "content": "skills in the tree are executed first (see Fig. 1 for an illustration of the tree structure).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 466, + 506, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "As a result, the agent naturally builds a curriculum on the episode lengths, which grow as the sequence", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 158, + 582 + ], + "score": 0.9, + "content": "( i T + H ) _ { i \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 569, + 505, + 583 + ], + "score": 1.0, + "content": ". As such, it does not require prior knowledge on an adequate horizon of the downstream", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 316, + 596 + ], + "score": 1.0, + "content": "goal-based task.4 Here this knowledge is replaced by", + "type": "text" + }, + { + "bbox": [ + 317, + 582, + 325, + 592 + ], + "score": 0.77, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 581, + 343, + 596 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 343, + 582, + 353, + 592 + ], + "score": 0.76, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 581, + 505, + 596 + ], + "score": 1.0, + "content": "which are more environment-agnostic", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "and task-agnostic quantities, as their choice rather has an impact on the size and shape of the learned", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 604, + 312, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 193, + 617 + ], + "score": 1.0, + "content": "tree (e.g., the smaller", + "type": "text" + }, + { + "bbox": [ + 193, + 604, + 202, + 614 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 604, + 219, + 617 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 604, + 230, + 614 + ], + "score": 0.82, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 604, + 312, + 617 + ], + "score": 1.0, + "content": "the bigger the tree).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 559, + 505, + 617 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 626, + 228, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 228, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 228, + 639 + ], + "score": 1.0, + "content": "3.4 The UPSIDE Algorithm", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 105, + 642, + 503, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "We are now ready to introduce UPSIDE, which provides a specific implementation of the components", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "described before (see Fig. 1 for a qualitative illustration and Algorithm 1 for the detailed pseudo-code).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 106, + 641, + 505, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 97, + 669, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 95, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 95, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "0 We perform standard approximations to make the constraint in (4) easier to estimate. We approximate", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 203, + 694 + ], + "score": 1.0, + "content": "the unknown posterior", + "type": "text" + }, + { + "bbox": [ + 203, + 681, + 229, + 692 + ], + "score": 0.92, + "content": "p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 679, + 352, + 694 + ], + "score": 1.0, + "content": "with a learned discriminator", + "type": "text" + }, + { + "bbox": [ + 353, + 681, + 383, + 693 + ], + "score": 0.93, + "content": "q _ { \\phi } ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 679, + 455, + 694 + ], + "score": 1.0, + "content": "with parameters", + "type": "text" + }, + { + "bbox": [ + 455, + 681, + 462, + 691 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 679, + 506, + 694 + ], + "score": 1.0, + "content": ". We also", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5, + "bbox_fs": [ + 95, + 668, + 506, + 694 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 86, + 72, + 505, + 118 + ], + "lines": [ + { + "bbox": [ + 86, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 86, + 75, + 100, + 84 + ], + "score": 1.0, + "content": "172", + "type": "text" + }, + { + "bbox": [ + 105, + 72, + 417, + 86 + ], + "score": 1.0, + "content": "remove the logarithm from the constraint to have an estimation range of", + "type": "text" + }, + { + "bbox": [ + 417, + 73, + 438, + 84 + ], + "score": 0.7, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "and thus lower", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 82, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 86, + 86, + 100, + 95 + ], + "score": 1.0, + "content": "173", + "type": "text" + }, + { + "bbox": [ + 104, + 82, + 306, + 99 + ], + "score": 1.0, + "content": "variance2. Finally, we replace the expectation over", + "type": "text" + }, + { + "bbox": [ + 307, + 86, + 312, + 93 + ], + "score": 0.59, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 82, + 420, + 99 + ], + "score": 1.0, + "content": "with an empirical estimate", + "type": "text" + }, + { + "bbox": [ + 420, + 84, + 448, + 97 + ], + "score": 0.92, + "content": "\\widehat { q } _ { \\phi } ^ { ( B ) } ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 82, + 506, + 99 + ], + "score": 1.0, + "content": "averaging the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 86, + 95, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 86, + 98, + 99, + 108 + ], + "score": 1.0, + "content": "174", + "type": "text" + }, + { + "bbox": [ + 104, + 95, + 295, + 109 + ], + "score": 1.0, + "content": "value of the discriminator evaluated on the last", + "type": "text" + }, + { + "bbox": [ + 295, + 96, + 304, + 106 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 95, + 506, + 109 + ], + "score": 1.0, + "content": "states observed while executing the diffusing part", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 108, + 319, + 119 + ], + "spans": [ + { + "bbox": [ + 86, + 108, + 100, + 118 + ], + "score": 1.0, + "content": "175", + "type": "text" + }, + { + "bbox": [ + 106, + 108, + 117, + 119 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 109, + 123, + 116 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 108, + 319, + 119 + ], + "score": 1.0, + "content": ". Integrating these approximations in (4) leads to", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 124, + 385, + 145 + ], + "lines": [ + { + "bbox": [ + 226, + 124, + 385, + 145 + ], + "spans": [ + { + "bbox": [ + 226, + 124, + 385, + 145 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { N \\geq 1 , \\pi } N \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\widehat { q } _ { \\phi } ^ { ( B ) } ( z ) \\geq \\eta .", + "type": "interline_equation", + "image_path": "3cf11c6f247fdf3cb31c0bd3fbdd8676fa016c31b28947636c19d28d2e51cbd2.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 226, + 124, + 385, + 145 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 86, + 151, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 86, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 86, + 154, + 100, + 164 + ], + "score": 1.0, + "content": "176", + "type": "text" + }, + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "As discussed in Sect. 3.2, this problem can be conveniently optimized using a greedy strategy. We", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 164, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 86, + 165, + 100, + 175 + ], + "score": 1.0, + "content": "177", + "type": "text" + }, + { + "bbox": [ + 105, + 164, + 505, + 175 + ], + "score": 1.0, + "content": "then integrate the optimization of (5) into an adaptive tree expansion strategy: (Generating new", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 172, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 86, + 176, + 99, + 185 + ], + "score": 1.0, + "content": "178", + "type": "text" + }, + { + "bbox": [ + 104, + 172, + 438, + 187 + ], + "score": 1.0, + "content": "skills) Given a tree structure as described in Sect. 3.3, we expand the tree at a leaf", + "type": "text" + }, + { + "bbox": [ + 438, + 175, + 447, + 184 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 172, + 490, + 187 + ], + "score": 1.0, + "content": "by adding", + "type": "text" + }, + { + "bbox": [ + 491, + 175, + 504, + 185 + ], + "score": 0.82, + "content": "N _ { 0 }", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 184, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 86, + 187, + 100, + 197 + ], + "score": 1.0, + "content": "179", + "type": "text" + }, + { + "bbox": [ + 105, + 184, + 506, + 197 + ], + "score": 1.0, + "content": "new nodes/skills following a breadth-first-search approach (lines 1, 2). Then (Skill Learning) the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 196, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 86, + 198, + 100, + 207 + ], + "score": 1.0, + "content": "180", + "type": "text" + }, + { + "bbox": [ + 105, + 196, + 505, + 208 + ], + "score": 1.0, + "content": "new skills are optimized by: i) sampling random skills in the tree to update the discriminator (lines", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 207, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 86, + 208, + 99, + 219 + ], + "score": 1.0, + "content": "181", + "type": "text" + }, + { + "bbox": [ + 105, + 207, + 506, + 219 + ], + "score": 1.0, + "content": "7-11), and ii) by updating the policies to optimize the discriminability reward (Sect. 3.1) computed", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 86, + 219, + 100, + 230 + ], + "score": 1.0, + "content": "182", + "type": "text" + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "using the discriminator (lines 13). To speed-up convergence, we only update the policies that have be", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 228, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 86, + 230, + 100, + 241 + ], + "score": 1.0, + "content": "183", + "type": "text" + }, + { + "bbox": [ + 105, + 228, + 506, + 241 + ], + "score": 1.0, + "content": "added to the tree structure, keeping all the previous policies fixed (line 12). Note that in the update of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 238, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 86, + 241, + 100, + 250 + ], + "score": 1.0, + "content": "184", + "type": "text" + }, + { + "bbox": [ + 104, + 238, + 506, + 254 + ], + "score": 1.0, + "content": "the discriminator we leverage the states observed in previous phases of the algorithm by maintaining", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 86, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 86, + 252, + 100, + 262 + ], + "score": 1.0, + "content": "185", + "type": "text" + }, + { + "bbox": [ + 104, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "a (small) replay buffer of states for each skill. (Node Consolidation) After a patience period (line 6),", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 86, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 86, + 263, + 100, + 272 + ], + "score": 1.0, + "content": "186", + "type": "text" + }, + { + "bbox": [ + 105, + 261, + 167, + 274 + ], + "score": 1.0, + "content": "if all skills are", + "type": "text" + }, + { + "bbox": [ + 167, + 263, + 174, + 273 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 261, + 401, + 274 + ], + "score": 1.0, + "content": "-consolidated, we tentatively add more skills to the leaf", + "type": "text" + }, + { + "bbox": [ + 401, + 263, + 410, + 271 + ], + "score": 0.67, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "(line 18). On the other", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 86, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 86, + 274, + 100, + 284 + ], + "score": 1.0, + "content": "187", + "type": "text" + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "hand, if any skill does not meet the discriminability threshold, we remove it and consolidate the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 86, + 282, + 386, + 296 + ], + "spans": [ + { + "bbox": [ + 86, + 285, + 100, + 294 + ], + "score": 1.0, + "content": "188", + "type": "text" + }, + { + "bbox": [ + 105, + 282, + 386, + 296 + ], + "score": 1.0, + "content": "remaining skills into the tree (lines 16, 17) and we repeat the process.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 86, + 299, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 86, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 86, + 302, + 99, + 311 + ], + "score": 1.0, + "content": "189", + "type": "text" + }, + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "Model selection. A core aspect of any RL algorithm is model selection, i.e., finding the best", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 86, + 310, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 86, + 312, + 100, + 322 + ], + "score": 1.0, + "content": "190", + "type": "text" + }, + { + "bbox": [ + 106, + 310, + 506, + 323 + ], + "score": 1.0, + "content": "configuration of hyperparameters. In URL with no prior knowledge of the downstream task(s), it", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 86, + 321, + 507, + 334 + ], + "spans": [ + { + "bbox": [ + 86, + 323, + 99, + 333 + ], + "score": 1.0, + "content": "191", + "type": "text" + }, + { + "bbox": [ + 105, + 321, + 507, + 334 + ], + "score": 1.0, + "content": "is non-trivial to devise an adequate criterion for model selection and this aspect is rarely addressed,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 86, + 332, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 86, + 334, + 100, + 344 + ], + "score": 1.0, + "content": "192", + "type": "text" + }, + { + "bbox": [ + 106, + 332, + 506, + 345 + ], + "score": 1.0, + "content": "despite being crucial in practice. For instance, while the coverage of the state space may be a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 86, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 86, + 345, + 100, + 355 + ], + "score": 1.0, + "content": "193", + "type": "text" + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "good proxy for the performance of a URL algorithm [see e.g., 8], it may be difficult to measure in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 86, + 355, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 86, + 357, + 99, + 365 + ], + "score": 1.0, + "content": "194", + "type": "text" + }, + { + "bbox": [ + 106, + 355, + 505, + 366 + ], + "score": 1.0, + "content": "continuous problems. Interestingly, our optimization problem directly provides a single, task-agnostic", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 86, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 86, + 367, + 100, + 377 + ], + "score": 1.0, + "content": "195", + "type": "text" + }, + { + "bbox": [ + 106, + 365, + 420, + 378 + ], + "score": 1.0, + "content": "and environment-agnostic criterion for model selection, which is the number", + "type": "text" + }, + { + "bbox": [ + 420, + 365, + 430, + 375 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 365, + 442, + 378 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 443, + 366, + 449, + 376 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "-consolidated", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 86, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 86, + 379, + 99, + 387 + ], + "score": 1.0, + "content": "196", + "type": "text" + }, + { + "bbox": [ + 106, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "skills discovered by the agent. Indeed in all of our experiments we simply select the model (i.e., set", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 86, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 86, + 388, + 100, + 398 + ], + "score": 1.0, + "content": "197", + "type": "text" + }, + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "score": 1.0, + "content": "of hyperparameters) that maximizes", + "type": "text" + }, + { + "bbox": [ + 253, + 387, + 263, + 397 + ], + "score": 0.71, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 387, + 506, + 399 + ], + "score": 1.0, + "content": ". This is a significant advantage w.r.t. existing methods, such", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 86, + 397, + 444, + 410 + ], + "spans": [ + { + "bbox": [ + 86, + 400, + 99, + 408 + ], + "score": 1.0, + "content": "198", + "type": "text" + }, + { + "bbox": [ + 104, + 397, + 444, + 410 + ], + "score": 1.0, + "content": "as VIC and DIAYN, for which no principled approach to model selection is provided.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 90, + 419, + 194, + 432 + ], + "lines": [ + { + "bbox": [ + 86, + 417, + 196, + 433 + ], + "spans": [ + { + "bbox": [ + 86, + 417, + 196, + 433 + ], + "score": 1.0, + "content": "199 4 Related work", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 102, + 444, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "score": 1.0, + "content": "Unsupervised Reinforcement Learning methods can be broadly decomposed according to the way", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "score": 1.0, + "content": "they summarize the experience accumulated during the unsupervised phase into reusable knowledge", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "to solve downstream tasks. This includes both off-policy model-free [e.g., 27] and model-based", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "score": 1.0, + "content": "[e.g., 29] methods that seek to populate a representative replay buffer and build accurate value or", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 488, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 507, + 501 + ], + "score": 1.0, + "content": "model estimates, that are used to solve a given downstream task in a zero- or few-shot manner.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "The accumulated experience during train time can also be compressed into a low-dimensional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "representation for value functions as well as policies and to improve exploration [e.g., 36]. An", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "alternative line of work focuses on the discovery of a set of skills in an unsupervised manner. Our", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 532, + 427, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 427, + 545 + ], + "score": 1.0, + "content": "approach falls in this category, on which we now focus our related work review.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 85, + 547, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 86, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 86, + 549, + 99, + 558 + ], + "score": 1.0, + "content": "209", + "type": "text" + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "Skill discovery based on MI maximization was first proposed in VIC [11], where only the final", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 86, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 86, + 560, + 99, + 569 + ], + "score": 1.0, + "content": "210", + "type": "text" + }, + { + "bbox": [ + 104, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "states of each trajectory are considered in the reverse form of (1) and where both the skills and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 86, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 86, + 571, + 99, + 580 + ], + "score": 1.0, + "content": "211", + "type": "text" + }, + { + "bbox": [ + 105, + 568, + 392, + 582 + ], + "score": 1.0, + "content": "their sampling rules are simultaneously learned (with a fixed support", + "type": "text" + }, + { + "bbox": [ + 393, + 569, + 407, + 581 + ], + "score": 0.89, + "content": "| Z |", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 568, + 506, + 582 + ], + "score": 1.0, + "content": ", i.e., a fixed number of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 86, + 577, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 86, + 582, + 99, + 591 + ], + "score": 1.0, + "content": "212", + "type": "text" + }, + { + "bbox": [ + 104, + 577, + 506, + 594 + ], + "score": 1.0, + "content": "skills). DIAYN [9] fixes the sampling rule to be uniform, and weighs the skills with an action-entropy", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 86, + 591, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 86, + 593, + 100, + 602 + ], + "score": 1.0, + "content": "213", + "type": "text" + }, + { + "bbox": [ + 106, + 591, + 506, + 603 + ], + "score": 1.0, + "content": "coefficient (i.e., it additionally minimizes the MI between actions and skills given the state), so as", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 86, + 602, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 86, + 603, + 100, + 614 + ], + "score": 1.0, + "content": "214", + "type": "text" + }, + { + "bbox": [ + 105, + 602, + 506, + 614 + ], + "score": 1.0, + "content": "to push the skills away from each other and enhance coverage. DADS [30] learns skills that are not", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 86, + 612, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 86, + 614, + 100, + 624 + ], + "score": 1.0, + "content": "215", + "type": "text" + }, + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "score": 1.0, + "content": "only diverse but also predictable by learned dynamics models, by using a generative model over", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 85, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 85, + 625, + 100, + 636 + ], + "score": 1.0, + "content": "216", + "type": "text" + }, + { + "bbox": [ + 105, + 623, + 429, + 636 + ], + "score": 1.0, + "content": "observations (rather than over skills) and optimizing a forward form of MI, namely", + "type": "text" + }, + { + "bbox": [ + 429, + 623, + 469, + 636 + ], + "score": 0.93, + "content": "\\mathcal { T } ( s ^ { \\prime } ; z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "between", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 86, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 86, + 636, + 100, + 646 + ], + "score": 1.0, + "content": "217", + "type": "text" + }, + { + "bbox": [ + 105, + 634, + 160, + 647 + ], + "score": 1.0, + "content": "the next state", + "type": "text" + }, + { + "bbox": [ + 160, + 635, + 169, + 644 + ], + "score": 0.85, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 634, + 235, + 647 + ], + "score": 1.0, + "content": "and current skill", + "type": "text" + }, + { + "bbox": [ + 235, + 636, + 242, + 644 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 634, + 463, + 647 + ], + "score": 1.0, + "content": "(with continuous latent) conditioned on the current state", + "type": "text" + }, + { + "bbox": [ + 463, + 637, + 469, + 644 + ], + "score": 0.68, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 634, + 505, + 647 + ], + "score": 1.0, + "content": ". EDL [8]", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 86, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 86, + 646, + 100, + 658 + ], + "score": 1.0, + "content": "218", + "type": "text" + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "shows that existing skill discovery approaches can provide insufficient coverage, and instead proposes", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 86, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 86, + 658, + 99, + 667 + ], + "score": 1.0, + "content": "219", + "type": "text" + }, + { + "bbox": [ + 106, + 656, + 273, + 668 + ], + "score": 1.0, + "content": "to rely on a fixed distribution over states", + "type": "text" + }, + { + "bbox": [ + 273, + 656, + 291, + 668 + ], + "score": 0.91, + "content": "p ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "which is either provided by an oracle or learned. In", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 86, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 86, + 669, + 99, + 678 + ], + "score": 1.0, + "content": "220", + "type": "text" + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "SMM [19], the MI formalism is used to learn a policy for which the state marginal distribution matches", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 86, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 86, + 680, + 99, + 690 + ], + "score": 1.0, + "content": "221", + "type": "text" + }, + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "score": 1.0, + "content": "a given target state distribution (e.g., uniform), which can be seen as a more scalable way of tackling", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 86, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 86, + 690, + 100, + 701 + ], + "score": 1.0, + "content": "222", + "type": "text" + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "the problem of maximum entropy over the state space [15], and as a way to encourage skills to go", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 86, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 86, + 702, + 100, + 712 + ], + "score": 1.0, + "content": "223", + "type": "text" + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "through unknown state regions. Other MI-based skill discovery methods include [10, 14, 24, 5, 34],", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 86, + 710, + 413, + 724 + ], + "spans": [ + { + "bbox": [ + 86, + 713, + 99, + 722 + ], + "score": 1.0, + "content": "224", + "type": "text" + }, + { + "bbox": [ + 104, + 710, + 413, + 724 + ], + "score": 1.0, + "content": "as well as [35, 20] which investigate skill discovery in non-episodic settings.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 45.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "index", + "bbox": [ + 86, + 72, + 505, + 118 + ], + "lines": [ + { + "bbox": [ + 86, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 86, + 75, + 100, + 84 + ], + "score": 1.0, + "content": "172", + "type": "text" + }, + { + "bbox": [ + 105, + 72, + 417, + 86 + ], + "score": 1.0, + "content": "remove the logarithm from the constraint to have an estimation range of", + "type": "text" + }, + { + "bbox": [ + 417, + 73, + 438, + 84 + ], + "score": 0.7, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "and thus lower", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 82, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 86, + 86, + 100, + 95 + ], + "score": 1.0, + "content": "173", + "type": "text" + }, + { + "bbox": [ + 104, + 82, + 306, + 99 + ], + "score": 1.0, + "content": "variance2. Finally, we replace the expectation over", + "type": "text" + }, + { + "bbox": [ + 307, + 86, + 312, + 93 + ], + "score": 0.59, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 82, + 420, + 99 + ], + "score": 1.0, + "content": "with an empirical estimate", + "type": "text" + }, + { + "bbox": [ + 420, + 84, + 448, + 97 + ], + "score": 0.92, + "content": "\\widehat { q } _ { \\phi } ^ { ( B ) } ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 82, + 506, + 99 + ], + "score": 1.0, + "content": "averaging the", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 95, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 86, + 98, + 99, + 108 + ], + "score": 1.0, + "content": "174", + "type": "text" + }, + { + "bbox": [ + 104, + 95, + 295, + 109 + ], + "score": 1.0, + "content": "value of the discriminator evaluated on the last", + "type": "text" + }, + { + "bbox": [ + 295, + 96, + 304, + 106 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 95, + 506, + 109 + ], + "score": 1.0, + "content": "states observed while executing the diffusing part", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 108, + 319, + 119 + ], + "spans": [ + { + "bbox": [ + 86, + 108, + 100, + 118 + ], + "score": 1.0, + "content": "175", + "type": "text" + }, + { + "bbox": [ + 106, + 108, + 117, + 119 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 109, + 123, + 116 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 108, + 319, + 119 + ], + "score": 1.0, + "content": ". Integrating these approximations in (4) leads to", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + } + ], + "index": 1.5, + "bbox_fs": [ + 86, + 72, + 506, + 119 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 124, + 385, + 145 + ], + "lines": [ + { + "bbox": [ + 226, + 124, + 385, + 145 + ], + "spans": [ + { + "bbox": [ + 226, + 124, + 385, + 145 + ], + "score": 0.93, + "content": "\\operatorname* { m a x } _ { N \\geq 1 , \\pi } N \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\widehat { q } _ { \\phi } ^ { ( B ) } ( z ) \\geq \\eta .", + "type": "interline_equation", + "image_path": "3cf11c6f247fdf3cb31c0bd3fbdd8676fa016c31b28947636c19d28d2e51cbd2.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 226, + 124, + 385, + 145 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "index", + "bbox": [ + 86, + 151, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 86, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 86, + 154, + 100, + 164 + ], + "score": 1.0, + "content": "176", + "type": "text" + }, + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "As discussed in Sect. 3.2, this problem can be conveniently optimized using a greedy strategy. We", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 164, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 86, + 165, + 100, + 175 + ], + "score": 1.0, + "content": "177", + "type": "text" + }, + { + "bbox": [ + 105, + 164, + 505, + 175 + ], + "score": 1.0, + "content": "then integrate the optimization of (5) into an adaptive tree expansion strategy: (Generating new", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 172, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 86, + 176, + 99, + 185 + ], + "score": 1.0, + "content": "178", + "type": "text" + }, + { + "bbox": [ + 104, + 172, + 438, + 187 + ], + "score": 1.0, + "content": "skills) Given a tree structure as described in Sect. 3.3, we expand the tree at a leaf", + "type": "text" + }, + { + "bbox": [ + 438, + 175, + 447, + 184 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 172, + 490, + 187 + ], + "score": 1.0, + "content": "by adding", + "type": "text" + }, + { + "bbox": [ + 491, + 175, + 504, + 185 + ], + "score": 0.82, + "content": "N _ { 0 }", + "type": "inline_equation" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 184, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 86, + 187, + 100, + 197 + ], + "score": 1.0, + "content": "179", + "type": "text" + }, + { + "bbox": [ + 105, + 184, + 506, + 197 + ], + "score": 1.0, + "content": "new nodes/skills following a breadth-first-search approach (lines 1, 2). Then (Skill Learning) the", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 196, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 86, + 198, + 100, + 207 + ], + "score": 1.0, + "content": "180", + "type": "text" + }, + { + "bbox": [ + 105, + 196, + 505, + 208 + ], + "score": 1.0, + "content": "new skills are optimized by: i) sampling random skills in the tree to update the discriminator (lines", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 207, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 86, + 208, + 99, + 219 + ], + "score": 1.0, + "content": "181", + "type": "text" + }, + { + "bbox": [ + 105, + 207, + 506, + 219 + ], + "score": 1.0, + "content": "7-11), and ii) by updating the policies to optimize the discriminability reward (Sect. 3.1) computed", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 86, + 219, + 100, + 230 + ], + "score": 1.0, + "content": "182", + "type": "text" + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "using the discriminator (lines 13). To speed-up convergence, we only update the policies that have be", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 228, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 86, + 230, + 100, + 241 + ], + "score": 1.0, + "content": "183", + "type": "text" + }, + { + "bbox": [ + 105, + 228, + 506, + 241 + ], + "score": 1.0, + "content": "added to the tree structure, keeping all the previous policies fixed (line 12). Note that in the update of", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 238, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 86, + 241, + 100, + 250 + ], + "score": 1.0, + "content": "184", + "type": "text" + }, + { + "bbox": [ + 104, + 238, + 506, + 254 + ], + "score": 1.0, + "content": "the discriminator we leverage the states observed in previous phases of the algorithm by maintaining", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 86, + 252, + 100, + 262 + ], + "score": 1.0, + "content": "185", + "type": "text" + }, + { + "bbox": [ + 104, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "a (small) replay buffer of states for each skill. (Node Consolidation) After a patience period (line 6),", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 86, + 263, + 100, + 272 + ], + "score": 1.0, + "content": "186", + "type": "text" + }, + { + "bbox": [ + 105, + 261, + 167, + 274 + ], + "score": 1.0, + "content": "if all skills are", + "type": "text" + }, + { + "bbox": [ + 167, + 263, + 174, + 273 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 261, + 401, + 274 + ], + "score": 1.0, + "content": "-consolidated, we tentatively add more skills to the leaf", + "type": "text" + }, + { + "bbox": [ + 401, + 263, + 410, + 271 + ], + "score": 0.67, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "(line 18). On the other", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 86, + 274, + 100, + 284 + ], + "score": 1.0, + "content": "187", + "type": "text" + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "hand, if any skill does not meet the discriminability threshold, we remove it and consolidate the", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 282, + 386, + 296 + ], + "spans": [ + { + "bbox": [ + 86, + 285, + 100, + 294 + ], + "score": 1.0, + "content": "188", + "type": "text" + }, + { + "bbox": [ + 105, + 282, + 386, + 296 + ], + "score": 1.0, + "content": "remaining skills into the tree (lines 16, 17) and we repeat the process.", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 86, + 302, + 99, + 311 + ], + "score": 1.0, + "content": "189", + "type": "text" + }, + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "Model selection. A core aspect of any RL algorithm is model selection, i.e., finding the best", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 310, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 86, + 312, + 100, + 322 + ], + "score": 1.0, + "content": "190", + "type": "text" + }, + { + "bbox": [ + 106, + 310, + 506, + 323 + ], + "score": 1.0, + "content": "configuration of hyperparameters. In URL with no prior knowledge of the downstream task(s), it", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 321, + 507, + 334 + ], + "spans": [ + { + "bbox": [ + 86, + 323, + 99, + 333 + ], + "score": 1.0, + "content": "191", + "type": "text" + }, + { + "bbox": [ + 105, + 321, + 507, + 334 + ], + "score": 1.0, + "content": "is non-trivial to devise an adequate criterion for model selection and this aspect is rarely addressed,", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 332, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 86, + 334, + 100, + 344 + ], + "score": 1.0, + "content": "192", + "type": "text" + }, + { + "bbox": [ + 106, + 332, + 506, + 345 + ], + "score": 1.0, + "content": "despite being crucial in practice. For instance, while the coverage of the state space may be a", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 86, + 345, + 100, + 355 + ], + "score": 1.0, + "content": "193", + "type": "text" + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "good proxy for the performance of a URL algorithm [see e.g., 8], it may be difficult to measure in", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 355, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 86, + 357, + 99, + 365 + ], + "score": 1.0, + "content": "194", + "type": "text" + }, + { + "bbox": [ + 106, + 355, + 505, + 366 + ], + "score": 1.0, + "content": "continuous problems. Interestingly, our optimization problem directly provides a single, task-agnostic", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 86, + 367, + 100, + 377 + ], + "score": 1.0, + "content": "195", + "type": "text" + }, + { + "bbox": [ + 106, + 365, + 420, + 378 + ], + "score": 1.0, + "content": "and environment-agnostic criterion for model selection, which is the number", + "type": "text" + }, + { + "bbox": [ + 420, + 365, + 430, + 375 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 365, + 442, + 378 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 443, + 366, + 449, + 376 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "-consolidated", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 86, + 379, + 99, + 387 + ], + "score": 1.0, + "content": "196", + "type": "text" + }, + { + "bbox": [ + 106, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "skills discovered by the agent. Indeed in all of our experiments we simply select the model (i.e., set", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 86, + 388, + 100, + 398 + ], + "score": 1.0, + "content": "197", + "type": "text" + }, + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "score": 1.0, + "content": "of hyperparameters) that maximizes", + "type": "text" + }, + { + "bbox": [ + 253, + 387, + 263, + 397 + ], + "score": 0.71, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 387, + 506, + 399 + ], + "score": 1.0, + "content": ". This is a significant advantage w.r.t. existing methods, such", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 397, + 444, + 410 + ], + "spans": [ + { + "bbox": [ + 86, + 400, + 99, + 408 + ], + "score": 1.0, + "content": "198", + "type": "text" + }, + { + "bbox": [ + 104, + 397, + 444, + 410 + ], + "score": 1.0, + "content": "as VIC and DIAYN, for which no principled approach to model selection is provided.", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + } + ], + "index": 11, + "bbox_fs": [ + 86, + 151, + 506, + 296 + ] + }, + { + "type": "index", + "bbox": [ + 86, + 299, + 505, + 409 + ], + "lines": [], + "index": 22.5, + "bbox_fs": [ + 86, + 299, + 507, + 410 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 90, + 419, + 194, + 432 + ], + "lines": [ + { + "bbox": [ + 86, + 417, + 196, + 433 + ], + "spans": [ + { + "bbox": [ + 86, + 417, + 196, + 433 + ], + "score": 1.0, + "content": "199 4 Related work", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 102, + 444, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "score": 1.0, + "content": "Unsupervised Reinforcement Learning methods can be broadly decomposed according to the way", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "score": 1.0, + "content": "they summarize the experience accumulated during the unsupervised phase into reusable knowledge", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "to solve downstream tasks. This includes both off-policy model-free [e.g., 27] and model-based", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "score": 1.0, + "content": "[e.g., 29] methods that seek to populate a representative replay buffer and build accurate value or", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 488, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 507, + 501 + ], + "score": 1.0, + "content": "model estimates, that are used to solve a given downstream task in a zero- or few-shot manner.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 512 + ], + "score": 1.0, + "content": "The accumulated experience during train time can also be compressed into a low-dimensional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "representation for value functions as well as policies and to improve exploration [e.g., 36]. An", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "alternative line of work focuses on the discovery of a set of skills in an unsupervised manner. Our", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 532, + 427, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 427, + 545 + ], + "score": 1.0, + "content": "approach falls in this category, on which we now focus our related work review.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 443, + 507, + 545 + ] + }, + { + "type": "index", + "bbox": [ + 85, + 547, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 86, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 86, + 549, + 99, + 558 + ], + "score": 1.0, + "content": "209", + "type": "text" + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "Skill discovery based on MI maximization was first proposed in VIC [11], where only the final", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 86, + 560, + 99, + 569 + ], + "score": 1.0, + "content": "210", + "type": "text" + }, + { + "bbox": [ + 104, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "states of each trajectory are considered in the reverse form of (1) and where both the skills and", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 86, + 571, + 99, + 580 + ], + "score": 1.0, + "content": "211", + "type": "text" + }, + { + "bbox": [ + 105, + 568, + 392, + 582 + ], + "score": 1.0, + "content": "their sampling rules are simultaneously learned (with a fixed support", + "type": "text" + }, + { + "bbox": [ + 393, + 569, + 407, + 581 + ], + "score": 0.89, + "content": "| Z |", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 568, + 506, + 582 + ], + "score": 1.0, + "content": ", i.e., a fixed number of", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 577, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 86, + 582, + 99, + 591 + ], + "score": 1.0, + "content": "212", + "type": "text" + }, + { + "bbox": [ + 104, + 577, + 506, + 594 + ], + "score": 1.0, + "content": "skills). DIAYN [9] fixes the sampling rule to be uniform, and weighs the skills with an action-entropy", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 591, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 86, + 593, + 100, + 602 + ], + "score": 1.0, + "content": "213", + "type": "text" + }, + { + "bbox": [ + 106, + 591, + 506, + 603 + ], + "score": 1.0, + "content": "coefficient (i.e., it additionally minimizes the MI between actions and skills given the state), so as", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 602, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 86, + 603, + 100, + 614 + ], + "score": 1.0, + "content": "214", + "type": "text" + }, + { + "bbox": [ + 105, + 602, + 506, + 614 + ], + "score": 1.0, + "content": "to push the skills away from each other and enhance coverage. DADS [30] learns skills that are not", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 612, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 86, + 614, + 100, + 624 + ], + "score": 1.0, + "content": "215", + "type": "text" + }, + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "score": 1.0, + "content": "only diverse but also predictable by learned dynamics models, by using a generative model over", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 85, + 625, + 100, + 636 + ], + "score": 1.0, + "content": "216", + "type": "text" + }, + { + "bbox": [ + 105, + 623, + 429, + 636 + ], + "score": 1.0, + "content": "observations (rather than over skills) and optimizing a forward form of MI, namely", + "type": "text" + }, + { + "bbox": [ + 429, + 623, + 469, + 636 + ], + "score": 0.93, + "content": "\\mathcal { T } ( s ^ { \\prime } ; z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "between", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 86, + 636, + 100, + 646 + ], + "score": 1.0, + "content": "217", + "type": "text" + }, + { + "bbox": [ + 105, + 634, + 160, + 647 + ], + "score": 1.0, + "content": "the next state", + "type": "text" + }, + { + "bbox": [ + 160, + 635, + 169, + 644 + ], + "score": 0.85, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 634, + 235, + 647 + ], + "score": 1.0, + "content": "and current skill", + "type": "text" + }, + { + "bbox": [ + 235, + 636, + 242, + 644 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 634, + 463, + 647 + ], + "score": 1.0, + "content": "(with continuous latent) conditioned on the current state", + "type": "text" + }, + { + "bbox": [ + 463, + 637, + 469, + 644 + ], + "score": 0.68, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 634, + 505, + 647 + ], + "score": 1.0, + "content": ". EDL [8]", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 86, + 646, + 100, + 658 + ], + "score": 1.0, + "content": "218", + "type": "text" + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "shows that existing skill discovery approaches can provide insufficient coverage, and instead proposes", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 86, + 658, + 99, + 667 + ], + "score": 1.0, + "content": "219", + "type": "text" + }, + { + "bbox": [ + 106, + 656, + 273, + 668 + ], + "score": 1.0, + "content": "to rely on a fixed distribution over states", + "type": "text" + }, + { + "bbox": [ + 273, + 656, + 291, + 668 + ], + "score": 0.91, + "content": "p ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "which is either provided by an oracle or learned. In", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 86, + 669, + 99, + 678 + ], + "score": 1.0, + "content": "220", + "type": "text" + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "SMM [19], the MI formalism is used to learn a policy for which the state marginal distribution matches", + "type": "text" + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 678, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 86, + 680, + 99, + 690 + ], + "score": 1.0, + "content": "221", + "type": "text" + }, + { + "bbox": [ + 105, + 678, + 506, + 691 + ], + "score": 1.0, + "content": "a given target state distribution (e.g., uniform), which can be seen as a more scalable way of tackling", + "type": "text" + } + ], + "index": 50, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 86, + 690, + 100, + 701 + ], + "score": 1.0, + "content": "222", + "type": "text" + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "the problem of maximum entropy over the state space [15], and as a way to encourage skills to go", + "type": "text" + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 86, + 702, + 100, + 712 + ], + "score": 1.0, + "content": "223", + "type": "text" + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "through unknown state regions. Other MI-based skill discovery methods include [10, 14, 24, 5, 34],", + "type": "text" + } + ], + "index": 52, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 710, + 413, + 724 + ], + "spans": [ + { + "bbox": [ + 86, + 713, + 99, + 722 + ], + "score": 1.0, + "content": "224", + "type": "text" + }, + { + "bbox": [ + 104, + 710, + 413, + 724 + ], + "score": 1.0, + "content": "as well as [35, 20] which investigate skill discovery in non-episodic settings.", + "type": "text" + } + ], + "index": 53, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 86, + 343, + 99, + 352 + ], + "score": 1.0, + "content": "225", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "Our approach shares a similar motivation to prior MI-based works of targeting skills that are both", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 86, + 354, + 99, + 364 + ], + "score": 1.0, + "content": "226", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "directed and state-covering. In particular, the decoupled structure introduced in Sect. 3.1 can be seen", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 86, + 365, + 99, + 374 + ], + "score": 1.0, + "content": "227", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "as a more suitable way to achieve the objective of improving the coverage of VIC as done in DIAYN", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 374, + 354, + 387 + ], + "spans": [ + { + "bbox": [ + 86, + 376, + 99, + 385 + ], + "score": 1.0, + "content": "228", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 374, + 354, + 387 + ], + "score": 1.0, + "content": "and SMM, without compromising the directedness of the skills.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 387, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 87, + 391, + 99, + 400 + ], + "score": 1.0, + "content": "229", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 387, + 505, + 402 + ], + "score": 1.0, + "content": "While most skill discovery approaches consider a fixed number of skills, a curriculum with increasing", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 87, + 401, + 99, + 411 + ], + "score": 1.0, + "content": "230", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "number of skills is studied in [1, 3]. Our discriminability constraint is what enables skills to be", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 87, + 412, + 99, + 422 + ], + "score": 1.0, + "content": "231", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "composed along a tree structure, which allows increases or decreases the support of available skills", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 421, + 279, + 434 + ], + "spans": [ + { + "bbox": [ + 87, + 424, + 99, + 433 + ], + "score": 1.0, + "content": "232", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 421, + 279, + 434 + ], + "score": 1.0, + "content": "depending on the region of the state space.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + } + ], + "index": 45.5, + "bbox_fs": [ + 85, + 546, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 69, + 494, + 266 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 69, + 494, + 266 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 69, + 494, + 266 + ], + "spans": [ + { + "bbox": [ + 125, + 69, + 494, + 266 + ], + "score": 0.976, + "type": "image", + "image_path": "e9c4d365719fb312ceafdef45970c4f57a1b44e9db7d7229c05847b9c4b09596.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 69, + 494, + 134.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 134.66666666666669, + 494, + 200.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 200.33333333333337, + 494, + 266.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 276, + 505, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "Figure 3: UPSIDE, DIAYN-curriculum and SMM-10 skills learned in a bottleneck maze (Top) and a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "U-maze (Bottom). For both DIAYN and SMM we report the stochastic execution of the learned skills and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "for UPSIDE we report the deterministic directed parts (that are composed) followed by the (stochastic)", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 308, + 382, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 382, + 322 + ], + "score": 1.0, + "content": "diffusing part, which is the same protocol used to evaluate coverage.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 87, + 340, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 86, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 86, + 343, + 99, + 352 + ], + "score": 1.0, + "content": "225", + "type": "text" + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "Our approach shares a similar motivation to prior MI-based works of targeting skills that are both", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 86, + 354, + 99, + 364 + ], + "score": 1.0, + "content": "226", + "type": "text" + }, + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "directed and state-covering. In particular, the decoupled structure introduced in Sect. 3.1 can be seen", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 86, + 365, + 99, + 374 + ], + "score": 1.0, + "content": "227", + "type": "text" + }, + { + "bbox": [ + 104, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "as a more suitable way to achieve the objective of improving the coverage of VIC as done in DIAYN", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 374, + 354, + 387 + ], + "spans": [ + { + "bbox": [ + 86, + 376, + 99, + 385 + ], + "score": 1.0, + "content": "228", + "type": "text" + }, + { + "bbox": [ + 105, + 374, + 354, + 387 + ], + "score": 1.0, + "content": "and SMM, without compromising the directedness of the skills.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 90, + 388, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 87, + 387, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 87, + 391, + 99, + 400 + ], + "score": 1.0, + "content": "229", + "type": "text" + }, + { + "bbox": [ + 105, + 387, + 505, + 402 + ], + "score": 1.0, + "content": "While most skill discovery approaches consider a fixed number of skills, a curriculum with increasing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 87, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 87, + 401, + 99, + 411 + ], + "score": 1.0, + "content": "230", + "type": "text" + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "number of skills is studied in [1, 3]. Our discriminability constraint is what enables skills to be", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 87, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 87, + 412, + 99, + 422 + ], + "score": 1.0, + "content": "231", + "type": "text" + }, + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "composed along a tree structure, which allows increases or decreases the support of available skills", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 87, + 421, + 279, + 434 + ], + "spans": [ + { + "bbox": [ + 87, + 424, + 99, + 433 + ], + "score": 1.0, + "content": "232", + "type": "text" + }, + { + "bbox": [ + 105, + 421, + 279, + 434 + ], + "score": 1.0, + "content": "depending on the region of the state space.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 98, + 436, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "Recently, [37] proposed a hierarchical RL method that discovers abstract and task-agnostic skills", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "while jointly learning a higher-level policy which is trained to maximize environment reward. Our", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 412, + 471 + ], + "score": 1.0, + "content": "approach builds on a similar promise of composing skills instead of resetting to", + "type": "text" + }, + { + "bbox": [ + 413, + 460, + 423, + 469 + ], + "score": 0.85, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "after each execution,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "score": 1.0, + "content": "yet we articulate the composition differently, by exploiting the direct-then-diffuse structure to ground", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 480, + 334, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 334, + 492 + ], + "score": 1.0, + "content": "learned skills to the state space instead of being abstract.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 86, + 495, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 86, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 86, + 497, + 99, + 507 + ], + "score": 1.0, + "content": "238", + "type": "text" + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "In addition, approaches such as DISCERN [33] and Skew-Fit [27] learn a goal-conditioned policy in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 85, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 85, + 507, + 100, + 518 + ], + "score": 1.0, + "content": "239", + "type": "text" + }, + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "an unsupervised way with an MI objective. As explained in [8, Sect. 5], this can be interpreted as a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 85, + 516, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 85, + 518, + 100, + 529 + ], + "score": 1.0, + "content": "240", + "type": "text" + }, + { + "bbox": [ + 104, + 516, + 254, + 529 + ], + "score": 1.0, + "content": "skill discovery approach with latent", + "type": "text" + }, + { + "bbox": [ + 254, + 517, + 284, + 527 + ], + "score": 0.9, + "content": "Z = S", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 516, + 507, + 529 + ], + "score": 1.0, + "content": ", i.e., where each goal state can define a different skill.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 86, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 86, + 530, + 99, + 540 + ], + "score": 1.0, + "content": "241", + "type": "text" + }, + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Conditioning on either goal states or abstract latent skills forms two extremes of the spectrum of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 85, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 85, + 540, + 100, + 551 + ], + "score": 1.0, + "content": "242", + "type": "text" + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "unsupervised RL. We target an intermediate approach, seeking to benefit from the groundedness of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 85, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 85, + 551, + 100, + 562 + ], + "score": 1.0, + "content": "243", + "type": "text" + }, + { + "bbox": [ + 105, + 549, + 164, + 562 + ], + "score": 1.0, + "content": "the latent skill", + "type": "text" + }, + { + "bbox": [ + 165, + 550, + 173, + 560 + ], + "score": 0.74, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 549, + 230, + 562 + ], + "score": 1.0, + "content": "and the states", + "type": "text" + }, + { + "bbox": [ + 230, + 550, + 238, + 560 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "(and thus amenability to composition) of goal-conditioned RL, and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 85, + 561, + 379, + 573 + ], + "spans": [ + { + "bbox": [ + 85, + 562, + 100, + 572 + ], + "score": 1.0, + "content": "244", + "type": "text" + }, + { + "bbox": [ + 105, + 561, + 379, + 573 + ], + "score": 1.0, + "content": "from the reduced search space and sampling ease of skill-based RL.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 101, + 575, + 504, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "An alternative approach to skill discovery builds on “spectral” properties of the dynamics of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "environment. This includes eigenoptions [21, 22] and covering options [17, 18], as well as the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "algorithm of [4] that builds a discrete graph representation which learns and composes spectral skills.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 105, + 620, + 191, + 633 + ], + "lines": [ + { + "bbox": [ + 104, + 618, + 193, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 618, + 193, + 638 + ], + "score": 1.0, + "content": "5 Experiments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "score": 1.0, + "content": "In this section, we investigate the following questions: i) Can the adaptive tree structure of UPSIDE in-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 650, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 664 + ], + "score": 1.0, + "content": "crementally cover an unknown environment while preserving directedness of the skills? ii) Following", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 661, + 484, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 484, + 676 + ], + "score": 1.0, + "content": "the unsupervised phase, how can UPSIDE be leveraged to solve goal-based downstream tasks?", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 103, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "We report results on: a) Navigation problems in continuous mazes, where actions represent the desired", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 136, + 700 + ], + "score": 1.0, + "content": "shift in", + "type": "text" + }, + { + "bbox": [ + 137, + 690, + 144, + 699 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 688, + 161, + 700 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 690, + 168, + 700 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "coordinates; b) A difficult instance of CartPole, where the cart starts with zero speed", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "and the pole is oriented downside; c) The Reacher [32] problem using the MuJoCo implementation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 711, + 425, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 384, + 723 + ], + "score": 1.0, + "content": "in Gym [7]. In all environments, the per-dimension action space is in", + "type": "text" + }, + { + "bbox": [ + 384, + 711, + 420, + 723 + ], + "score": 0.92, + "content": "[ - 1 ; + 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 711, + 425, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 69, + 494, + 266 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 69, + 494, + 266 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 69, + 494, + 266 + ], + "spans": [ + { + "bbox": [ + 125, + 69, + 494, + 266 + ], + "score": 0.976, + "type": "image", + "image_path": "e9c4d365719fb312ceafdef45970c4f57a1b44e9db7d7229c05847b9c4b09596.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 69, + 494, + 134.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 134.66666666666669, + 494, + 200.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 200.33333333333337, + 494, + 266.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 276, + 505, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "Figure 3: UPSIDE, DIAYN-curriculum and SMM-10 skills learned in a bottleneck maze (Top) and a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "U-maze (Bottom). For both DIAYN and SMM we report the stochastic execution of the learned skills and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "for UPSIDE we report the deterministic directed parts (that are composed) followed by the (stochastic)", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 308, + 382, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 382, + 322 + ], + "score": 1.0, + "content": "diffusing part, which is the same protocol used to evaluate coverage.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "index", + "bbox": [ + 87, + 340, + 505, + 385 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 86, + 340, + 505, + 387 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 90, + 388, + 505, + 433 + ], + "lines": [], + "index": 12.5, + "bbox_fs": [ + 87, + 387, + 505, + 434 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 98, + 436, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "Recently, [37] proposed a hierarchical RL method that discovers abstract and task-agnostic skills", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "while jointly learning a higher-level policy which is trained to maximize environment reward. Our", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 412, + 471 + ], + "score": 1.0, + "content": "approach builds on a similar promise of composing skills instead of resetting to", + "type": "text" + }, + { + "bbox": [ + 413, + 460, + 423, + 469 + ], + "score": 0.85, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "after each execution,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 483 + ], + "score": 1.0, + "content": "yet we articulate the composition differently, by exploiting the direct-then-diffuse structure to ground", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 480, + 334, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 334, + 492 + ], + "score": 1.0, + "content": "learned skills to the state space instead of being abstract.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 437, + 506, + 492 + ] + }, + { + "type": "index", + "bbox": [ + 86, + 495, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 86, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 86, + 497, + 99, + 507 + ], + "score": 1.0, + "content": "238", + "type": "text" + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "In addition, approaches such as DISCERN [33] and Skew-Fit [27] learn a goal-conditioned policy in", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 85, + 507, + 100, + 518 + ], + "score": 1.0, + "content": "239", + "type": "text" + }, + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "an unsupervised way with an MI objective. As explained in [8, Sect. 5], this can be interpreted as a", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 516, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 85, + 518, + 100, + 529 + ], + "score": 1.0, + "content": "240", + "type": "text" + }, + { + "bbox": [ + 104, + 516, + 254, + 529 + ], + "score": 1.0, + "content": "skill discovery approach with latent", + "type": "text" + }, + { + "bbox": [ + 254, + 517, + 284, + 527 + ], + "score": 0.9, + "content": "Z = S", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 516, + 507, + 529 + ], + "score": 1.0, + "content": ", i.e., where each goal state can define a different skill.", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 86, + 530, + 99, + 540 + ], + "score": 1.0, + "content": "241", + "type": "text" + }, + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Conditioning on either goal states or abstract latent skills forms two extremes of the spectrum of", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 85, + 540, + 100, + 551 + ], + "score": 1.0, + "content": "242", + "type": "text" + }, + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "unsupervised RL. We target an intermediate approach, seeking to benefit from the groundedness of", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 85, + 551, + 100, + 562 + ], + "score": 1.0, + "content": "243", + "type": "text" + }, + { + "bbox": [ + 105, + 549, + 164, + 562 + ], + "score": 1.0, + "content": "the latent skill", + "type": "text" + }, + { + "bbox": [ + 165, + 550, + 173, + 560 + ], + "score": 0.74, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 549, + 230, + 562 + ], + "score": 1.0, + "content": "and the states", + "type": "text" + }, + { + "bbox": [ + 230, + 550, + 238, + 560 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "(and thus amenability to composition) of goal-conditioned RL, and", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 561, + 379, + 573 + ], + "spans": [ + { + "bbox": [ + 85, + 562, + 100, + 572 + ], + "score": 1.0, + "content": "244", + "type": "text" + }, + { + "bbox": [ + 105, + 561, + 379, + 573 + ], + "score": 1.0, + "content": "from the reduced search space and sampling ease of skill-based RL.", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + } + ], + "index": 23, + "bbox_fs": [ + 85, + 495, + 507, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 101, + 575, + 504, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "An alternative approach to skill discovery builds on “spectral” properties of the dynamics of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "environment. This includes eigenoptions [21, 22] and covering options [17, 18], as well as the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "algorithm of [4] that builds a discrete graph representation which learns and composes spectral skills.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 575, + 505, + 609 + ] + }, + { + "type": "title", + "bbox": [ + 105, + 620, + 191, + 633 + ], + "lines": [ + { + "bbox": [ + 104, + 618, + 193, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 618, + 193, + 638 + ], + "score": 1.0, + "content": "5 Experiments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "score": 1.0, + "content": "In this section, we investigate the following questions: i) Can the adaptive tree structure of UPSIDE in-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 650, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 664 + ], + "score": 1.0, + "content": "crementally cover an unknown environment while preserving directedness of the skills? ii) Following", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 661, + 484, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 484, + 676 + ], + "score": 1.0, + "content": "the unsupervised phase, how can UPSIDE be leveraged to solve goal-based downstream tasks?", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 639, + 507, + 676 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "We report results on: a) Navigation problems in continuous mazes, where actions represent the desired", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 136, + 700 + ], + "score": 1.0, + "content": "shift in", + "type": "text" + }, + { + "bbox": [ + 137, + 690, + 144, + 699 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 688, + 161, + 700 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 690, + 168, + 700 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "coordinates; b) A difficult instance of CartPole, where the cart starts with zero speed", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "and the pole is oriented downside; c) The Reacher [32] problem using the MuJoCo implementation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 711, + 425, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 384, + 723 + ], + "score": 1.0, + "content": "in Gym [7]. In all environments, the per-dimension action space is in", + "type": "text" + }, + { + "bbox": [ + 384, + 711, + 420, + 723 + ], + "score": 0.92, + "content": "[ - 1 ; + 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 711, + 425, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 678, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 70, + 481, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 70, + 481, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 70, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 481, + 189 + ], + "score": 0.973, + "type": "image", + "image_path": "f67a926a679d5569359c50d0e257b1aa3baa0708a7b5b489b2d719b79dee97b3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 70, + 481, + 109.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 109.66666666666666, + 481, + 149.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 149.33333333333331, + 481, + 188.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 187, + 194, + 422, + 206 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 194, + 423, + 207 + ], + "spans": [ + { + "bbox": [ + 188, + 194, + 423, + 207 + ], + "score": 1.0, + "content": "Figure 4: Normalized coverage in U-maze and bottleneck.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 86, + 225, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 86, + 225, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 86, + 227, + 99, + 236 + ], + "score": 1.0, + "content": "256", + "type": "text" + }, + { + "bbox": [ + 106, + 226, + 322, + 236 + ], + "score": 1.0, + "content": "We compare to different baselines. DIAYN-K, where", + "type": "text" + }, + { + "bbox": [ + 322, + 225, + 333, + 235 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 226, + 505, + 236 + ], + "score": 1.0, + "content": "is a fixed number of skills, is the original", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 85, + 237, + 100, + 248 + ], + "score": 1.0, + "content": "257", + "type": "text" + }, + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "algorithm proposed in [9]. DIAYN-Curriculum is a variant where the number of skills is automatically", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 246, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 86, + 249, + 99, + 258 + ], + "score": 1.0, + "content": "258", + "type": "text" + }, + { + "bbox": [ + 105, + 246, + 506, + 260 + ], + "score": 1.0, + "content": "tuned following the same procedure as in UPSIDE ensuring a good discriminability. We also compare", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 86, + 260, + 99, + 269 + ], + "score": 1.0, + "content": "259", + "type": "text" + }, + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "to SMM [19], which is similar to DIAYN, but it includes an exploration bonus encouraging the policies", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 85, + 269, + 100, + 281 + ], + "score": 1.0, + "content": "260", + "type": "text" + }, + { + "bbox": [ + 104, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "to visit rarely encountered states. In our implementation, the exploration bonus is obtained by", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 279, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 86, + 282, + 99, + 291 + ], + "score": 1.0, + "content": "261", + "type": "text" + }, + { + "bbox": [ + 105, + 279, + 506, + 294 + ], + "score": 1.0, + "content": "maintaining a multinomial distribution over “buckets of states” obtained by discretization, resulting", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 86, + 293, + 99, + 302 + ], + "score": 1.0, + "content": "262", + "type": "text" + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "in an computation-efficient and stable implementation that is more stable than the original VAE-based", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 302, + 453, + 314 + ], + "spans": [ + { + "bbox": [ + 86, + 304, + 99, + 312 + ], + "score": 1.0, + "content": "263", + "type": "text" + }, + { + "bbox": [ + 105, + 302, + 453, + 314 + ], + "score": 1.0, + "content": "method. UPSIDE and all baselines are implemented with Soft-Actor Critic (SAC) [13].", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 88, + 317, + 505, + 406 + ], + "lines": [ + { + "bbox": [ + 86, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 86, + 320, + 100, + 330 + ], + "score": 1.0, + "content": "264", + "type": "text" + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "Unsupervised Phase. We run all methods until convergence. We then do model selection according", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 330, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 86, + 331, + 100, + 340 + ], + "score": 1.0, + "content": "265", + "type": "text" + }, + { + "bbox": [ + 106, + 330, + 505, + 340 + ], + "score": 1.0, + "content": "to the criterion of either the final number of skills for UPSIDE and DIAYN-curriculum and the final", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 85, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 85, + 341, + 100, + 352 + ], + "score": 1.0, + "content": "266", + "type": "text" + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "average discriminability for DIAYN-K and SMM. To compute the coverage, we perform rollouts by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 85, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 85, + 352, + 100, + 363 + ], + "score": 1.0, + "content": "267", + "type": "text" + }, + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "first sampling a skill uniformly at random and executing its associated policy until termination. We", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 85, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 85, + 363, + 100, + 374 + ], + "score": 1.0, + "content": "268", + "type": "text" + }, + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "discretize states into buckets (50 interval per dimension for mazes and 10 for control environments)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 86, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 86, + 375, + 99, + 384 + ], + "score": 1.0, + "content": "269", + "type": "text" + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "and report the proportion of buckets reached by each method as a function of the total number of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 85, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 85, + 385, + 100, + 396 + ], + "score": 1.0, + "content": "270", + "type": "text" + }, + { + "bbox": [ + 105, + 384, + 505, + 395 + ], + "score": 1.0, + "content": "steps executed in the environment over multiple rollouts. Since only a small portion of the discretized", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 86, + 394, + 449, + 407 + ], + "spans": [ + { + "bbox": [ + 86, + 397, + 99, + 406 + ], + "score": 1.0, + "content": "271", + "type": "text" + }, + { + "bbox": [ + 106, + 394, + 449, + 407 + ], + "score": 1.0, + "content": "states can be reached, we normalize the coverage such that the best method obtains 1.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 87, + 410, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 86, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 86, + 412, + 99, + 422 + ], + "score": 1.0, + "content": "272", + "type": "text" + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "We consider two topologies of mazes with size (height and width) 50 such that exploration is non-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 86, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 86, + 424, + 99, + 433 + ], + "score": 1.0, + "content": "273", + "type": "text" + }, + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "trivial (i.e., a random policy is only able to cover a small part of the state space): a U-shaped maze", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 86, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 86, + 434, + 99, + 443 + ], + "score": 1.0, + "content": "274", + "type": "text" + }, + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "and a Bottleneck maze (which is a harder version of the one in [8, Fig. 1] which is only of size 10", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 85, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 85, + 444, + 100, + 455 + ], + "score": 1.0, + "content": "275", + "type": "text" + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "for the same action space). In Fig. 3 we show that UPSIDE succeeds in covering the near-entirety", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 86, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 86, + 457, + 99, + 466 + ], + "score": 1.0, + "content": "276", + "type": "text" + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "of the state space by creating a tree of directed skills. Moreover, UPSIDE created directed skills", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 86, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 86, + 467, + 99, + 477 + ], + "score": 1.0, + "content": "277", + "type": "text" + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "with a low entropy, while the two baselines tend to create skills that are more stochastic. This is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 86, + 475, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 86, + 478, + 99, + 488 + ], + "score": 1.0, + "content": "278", + "type": "text" + }, + { + "bbox": [ + 105, + 475, + 505, + 489 + ], + "score": 1.0, + "content": "particularly evident for SMM, due to the state-entropy exploration bonus, that while it encourages", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 86, + 487, + 283, + 499 + ], + "spans": [ + { + "bbox": [ + 86, + 489, + 99, + 498 + ], + "score": 1.0, + "content": "279", + "type": "text" + }, + { + "bbox": [ + 105, + 487, + 283, + 499 + ], + "score": 1.0, + "content": "broader coverage makes skills less directed.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 105, + 503, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "score": 1.0, + "content": "In Fig. 4 we report the coverage on the Bottleneck maze and U-Maze. For UPSIDE, executing a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "skill corresponds to executing the directed part of all the “parent” skills in the tree and concluding", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "with the diffusion part of the skill. SMM achieves better coverage than DIAYN thanks to the increased", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "level of stochasticity (diffusion) of its skills. UPSIDE outperforms both by reaching regions of the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 547, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 452, + 559 + ], + "score": 1.0, + "content": "environment that are not be achieved by other methods. Here, we plot UPSIDE with", + "type": "text" + }, + { + "bbox": [ + 453, + 547, + 487, + 557 + ], + "score": 0.91, + "content": "T = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 547, + 506, + 559 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 558, + 488, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 140, + 569 + ], + "score": 0.88, + "content": "H = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 558, + 488, + 570 + ], + "score": 1.0, + "content": ", but we found UPSIDE to be robust to these parameters as shown in the supplementary.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 101, + 574, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 417, + 587 + ], + "score": 1.0, + "content": "Results are similar in the CartPole problem (see Fig. 5) where UPSIDE (with", + "type": "text" + }, + { + "bbox": [ + 417, + 574, + 450, + 585 + ], + "score": 0.89, + "content": "T = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 573, + 467, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 468, + 574, + 501, + 585 + ], + "score": 0.87, + "content": "H = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 573, + 506, + 587 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "obtains better coverage than baselines. On the other hand, in Reacher (see Fig. 5), DIAYN-50", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "outperforms UPSIDE in terms of coverage. This can be explained by the fact that, in this environment,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "highly stochastic skills provide a good coverage. Nonetheless, this comes at the cost of very low", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "discriminability (rightmost plot), which suggests DIAYN-50 skills have poor directedness. On the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "other hand, UPSIDE (and DIAYN-curriculum) achieves much larger discriminability by removing", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 319, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 319, + 652 + ], + "score": 1.0, + "content": "redundant skills and favoring more directed policies.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 656, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Downstream Tasks. Following the unsupervised phase, UPSIDE has learned a tree of skills. We", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "now investigate how these skills are used to tackle a downstream task. In that setting, we propose to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "use skill-based approaches (i.e UPSIDE, DIAYN and SMM) in the following way: a) (exploration) first", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "we sample rollouts over the different skills. b) We then select the best skill based on the maximum", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "cumulative reward collected and c) we fine-tune this skill to maximize the reward. We report results", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "on mazes (additional results are provided in the supplementary). We consider a sparse positive reward", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 70, + 481, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 70, + 481, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 70, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 481, + 189 + ], + "score": 0.973, + "type": "image", + "image_path": "f67a926a679d5569359c50d0e257b1aa3baa0708a7b5b489b2d719b79dee97b3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 70, + 481, + 109.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 109.66666666666666, + 481, + 149.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 149.33333333333331, + 481, + 188.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 187, + 194, + 422, + 206 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 194, + 423, + 207 + ], + "spans": [ + { + "bbox": [ + 188, + 194, + 423, + 207 + ], + "score": 1.0, + "content": "Figure 4: Normalized coverage in U-maze and bottleneck.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "index", + "bbox": [ + 86, + 225, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 86, + 225, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 86, + 227, + 99, + 236 + ], + "score": 1.0, + "content": "256", + "type": "text" + }, + { + "bbox": [ + 106, + 226, + 322, + 236 + ], + "score": 1.0, + "content": "We compare to different baselines. DIAYN-K, where", + "type": "text" + }, + { + "bbox": [ + 322, + 225, + 333, + 235 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 226, + 505, + 236 + ], + "score": 1.0, + "content": "is a fixed number of skills, is the original", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 85, + 237, + 100, + 248 + ], + "score": 1.0, + "content": "257", + "type": "text" + }, + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "algorithm proposed in [9]. DIAYN-Curriculum is a variant where the number of skills is automatically", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 246, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 86, + 249, + 99, + 258 + ], + "score": 1.0, + "content": "258", + "type": "text" + }, + { + "bbox": [ + 105, + 246, + 506, + 260 + ], + "score": 1.0, + "content": "tuned following the same procedure as in UPSIDE ensuring a good discriminability. We also compare", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 86, + 260, + 99, + 269 + ], + "score": 1.0, + "content": "259", + "type": "text" + }, + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "to SMM [19], which is similar to DIAYN, but it includes an exploration bonus encouraging the policies", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 85, + 269, + 100, + 281 + ], + "score": 1.0, + "content": "260", + "type": "text" + }, + { + "bbox": [ + 104, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "to visit rarely encountered states. In our implementation, the exploration bonus is obtained by", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 279, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 86, + 282, + 99, + 291 + ], + "score": 1.0, + "content": "261", + "type": "text" + }, + { + "bbox": [ + 105, + 279, + 506, + 294 + ], + "score": 1.0, + "content": "maintaining a multinomial distribution over “buckets of states” obtained by discretization, resulting", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 86, + 293, + 99, + 302 + ], + "score": 1.0, + "content": "262", + "type": "text" + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "in an computation-efficient and stable implementation that is more stable than the original VAE-based", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 302, + 453, + 314 + ], + "spans": [ + { + "bbox": [ + 86, + 304, + 99, + 312 + ], + "score": 1.0, + "content": "263", + "type": "text" + }, + { + "bbox": [ + 105, + 302, + 453, + 314 + ], + "score": 1.0, + "content": "method. UPSIDE and all baselines are implemented with Soft-Actor Critic (SAC) [13].", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 86, + 320, + 100, + 330 + ], + "score": 1.0, + "content": "264", + "type": "text" + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "Unsupervised Phase. We run all methods until convergence. We then do model selection according", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 330, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 86, + 331, + 100, + 340 + ], + "score": 1.0, + "content": "265", + "type": "text" + }, + { + "bbox": [ + 106, + 330, + 505, + 340 + ], + "score": 1.0, + "content": "to the criterion of either the final number of skills for UPSIDE and DIAYN-curriculum and the final", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 85, + 341, + 100, + 352 + ], + "score": 1.0, + "content": "266", + "type": "text" + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "average discriminability for DIAYN-K and SMM. To compute the coverage, we perform rollouts by", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 85, + 352, + 100, + 363 + ], + "score": 1.0, + "content": "267", + "type": "text" + }, + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "first sampling a skill uniformly at random and executing its associated policy until termination. We", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 85, + 363, + 100, + 374 + ], + "score": 1.0, + "content": "268", + "type": "text" + }, + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "discretize states into buckets (50 interval per dimension for mazes and 10 for control environments)", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 86, + 375, + 99, + 384 + ], + "score": 1.0, + "content": "269", + "type": "text" + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "and report the proportion of buckets reached by each method as a function of the total number of", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 85, + 385, + 100, + 396 + ], + "score": 1.0, + "content": "270", + "type": "text" + }, + { + "bbox": [ + 105, + 384, + 505, + 395 + ], + "score": 1.0, + "content": "steps executed in the environment over multiple rollouts. Since only a small portion of the discretized", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 394, + 449, + 407 + ], + "spans": [ + { + "bbox": [ + 86, + 397, + 99, + 406 + ], + "score": 1.0, + "content": "271", + "type": "text" + }, + { + "bbox": [ + 106, + 394, + 449, + 407 + ], + "score": 1.0, + "content": "states can be reached, we normalize the coverage such that the best method obtains 1.", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 86, + 412, + 99, + 422 + ], + "score": 1.0, + "content": "272", + "type": "text" + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "We consider two topologies of mazes with size (height and width) 50 such that exploration is non-", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 86, + 424, + 99, + 433 + ], + "score": 1.0, + "content": "273", + "type": "text" + }, + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "trivial (i.e., a random policy is only able to cover a small part of the state space): a U-shaped maze", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 86, + 434, + 99, + 443 + ], + "score": 1.0, + "content": "274", + "type": "text" + }, + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "and a Bottleneck maze (which is a harder version of the one in [8, Fig. 1] which is only of size 10", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 85, + 444, + 100, + 455 + ], + "score": 1.0, + "content": "275", + "type": "text" + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "for the same action space). In Fig. 3 we show that UPSIDE succeeds in covering the near-entirety", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 86, + 457, + 99, + 466 + ], + "score": 1.0, + "content": "276", + "type": "text" + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "of the state space by creating a tree of directed skills. Moreover, UPSIDE created directed skills", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 86, + 467, + 99, + 477 + ], + "score": 1.0, + "content": "277", + "type": "text" + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "with a low entropy, while the two baselines tend to create skills that are more stochastic. This is", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 475, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 86, + 478, + 99, + 488 + ], + "score": 1.0, + "content": "278", + "type": "text" + }, + { + "bbox": [ + 105, + 475, + 505, + 489 + ], + "score": 1.0, + "content": "particularly evident for SMM, due to the state-entropy exploration bonus, that while it encourages", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 487, + 283, + 499 + ], + "spans": [ + { + "bbox": [ + 86, + 489, + 99, + 498 + ], + "score": 1.0, + "content": "279", + "type": "text" + }, + { + "bbox": [ + 105, + 487, + 283, + 499 + ], + "score": 1.0, + "content": "broader coverage makes skills less directed.", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + } + ], + "index": 7.5, + "bbox_fs": [ + 85, + 225, + 506, + 314 + ] + }, + { + "type": "index", + "bbox": [ + 88, + 317, + 505, + 406 + ], + "lines": [], + "index": 15.5, + "bbox_fs": [ + 85, + 317, + 506, + 407 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 87, + 410, + 505, + 498 + ], + "lines": [], + "index": 23.5, + "bbox_fs": [ + 85, + 410, + 506, + 499 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 105, + 503, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "score": 1.0, + "content": "In Fig. 4 we report the coverage on the Bottleneck maze and U-Maze. For UPSIDE, executing a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "skill corresponds to executing the directed part of all the “parent” skills in the tree and concluding", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "with the diffusion part of the skill. SMM achieves better coverage than DIAYN thanks to the increased", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "level of stochasticity (diffusion) of its skills. UPSIDE outperforms both by reaching regions of the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 547, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 452, + 559 + ], + "score": 1.0, + "content": "environment that are not be achieved by other methods. Here, we plot UPSIDE with", + "type": "text" + }, + { + "bbox": [ + 453, + 547, + 487, + 557 + ], + "score": 0.91, + "content": "T = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 547, + 506, + 559 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 558, + 488, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 140, + 569 + ], + "score": 0.88, + "content": "H = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 558, + 488, + 570 + ], + "score": 1.0, + "content": ", but we found UPSIDE to be robust to these parameters as shown in the supplementary.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 502, + 506, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 101, + 574, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 417, + 587 + ], + "score": 1.0, + "content": "Results are similar in the CartPole problem (see Fig. 5) where UPSIDE (with", + "type": "text" + }, + { + "bbox": [ + 417, + 574, + 450, + 585 + ], + "score": 0.89, + "content": "T = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 573, + 467, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 468, + 574, + 501, + 585 + ], + "score": 0.87, + "content": "H = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 573, + 506, + 587 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "obtains better coverage than baselines. On the other hand, in Reacher (see Fig. 5), DIAYN-50", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "outperforms UPSIDE in terms of coverage. This can be explained by the fact that, in this environment,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "highly stochastic skills provide a good coverage. Nonetheless, this comes at the cost of very low", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "discriminability (rightmost plot), which suggests DIAYN-50 skills have poor directedness. On the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "other hand, UPSIDE (and DIAYN-curriculum) achieves much larger discriminability by removing", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 319, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 319, + 652 + ], + "score": 1.0, + "content": "redundant skills and favoring more directed policies.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 573, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 656, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Downstream Tasks. Following the unsupervised phase, UPSIDE has learned a tree of skills. We", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "now investigate how these skills are used to tackle a downstream task. In that setting, we propose to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "use skill-based approaches (i.e UPSIDE, DIAYN and SMM) in the following way: a) (exploration) first", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "we sample rollouts over the different skills. b) We then select the best skill based on the maximum", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "cumulative reward collected and c) we fine-tune this skill to maximize the reward. We report results", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "on mazes (additional results are provided in the supplementary). We consider a sparse positive reward", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 656, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 501, + 173 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 70, + 501, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 70, + 501, + 173 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 501, + 173 + ], + "score": 0.969, + "type": "image", + "image_path": "c0fef9fb524cc8a68d444c1ca195519b14d66a4947d40fb5df6c83c782ac91ca.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 70, + 501, + 104.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 104.33333333333334, + 501, + 138.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 138.66666666666669, + 501, + 173.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 176, + 506, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 174, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 189 + ], + "score": 1.0, + "content": "Figure 5: Normalized coverage in Cartpole (Left) and Reacher (Middle). (Right) Average discrim-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 186, + 301, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 301, + 200 + ], + "score": 1.0, + "content": "inability of the skills during training in Reacher.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 86, + 211, + 505, + 300 + ], + "lines": [ + { + "bbox": [ + 86, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 86, + 213, + 99, + 223 + ], + "score": 1.0, + "content": "299", + "type": "text" + }, + { + "bbox": [ + 105, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "when reaching a particular defined goal.5 We consider goals at different distances from the initial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 86, + 225, + 99, + 234 + ], + "score": 1.0, + "content": "300", + "type": "text" + }, + { + "bbox": [ + 105, + 223, + 127, + 235 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 127, + 224, + 137, + 234 + ], + "score": 0.83, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 223, + 505, + 235 + ], + "score": 1.0, + "content": ", the further, the harder. Fig. 6 shows the learning curves obtained when fine-tuning the best", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 84, + 235, + 100, + 245 + ], + "score": 1.0, + "content": "301", + "type": "text" + }, + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "skill for the different models and compare to a classical SAC algorithm where a single policy is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 85, + 245, + 100, + 256 + ], + "score": 1.0, + "content": "302", + "type": "text" + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "learned from scratch. DIAYN/SMM means we use the best state-covering policies between DIAYN and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 86, + 257, + 99, + 267 + ], + "score": 1.0, + "content": "303", + "type": "text" + }, + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "SMM. For the “close” goal setting, both UPSIDE and DIAYN/SMM are able to learn to reach this goal", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 86, + 268, + 99, + 277 + ], + "score": 1.0, + "content": "304", + "type": "text" + }, + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "efficiently while SAC solves the task only for some of the training runs. Note that we do not show", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 85, + 278, + 100, + 289 + ], + "score": 1.0, + "content": "305", + "type": "text" + }, + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "DIAYN performance since it is lower than the SMM one. For the “far” goal setting, only UPSIDE learns", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 289, + 361, + 301 + ], + "spans": [ + { + "bbox": [ + 86, + 290, + 99, + 300 + ], + "score": 1.0, + "content": "306", + "type": "text" + }, + { + "bbox": [ + 106, + 289, + 361, + 301 + ], + "score": 1.0, + "content": "to reach this goal. Obtained trajectories are illustrated in Fig. 6.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 110, + 309, + 495, + 528 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 309, + 495, + 528 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 309, + 495, + 528 + ], + "spans": [ + { + "bbox": [ + 110, + 309, + 495, + 528 + ], + "score": 0.945, + "type": "image", + "image_path": "5c8600e6316f678077840de51c561851a9e26dcd890e6d78ce9628180fe6febb.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 110, + 309, + 495, + 382.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 110, + 382.0, + 495, + 455.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 110, + 455.0, + 495, + 528.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 531, + 386, + 564 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 530, + 386, + 542 + ], + "spans": [ + { + "bbox": [ + 145, + 530, + 386, + 542 + ], + "score": 1.0, + "content": "(Left): Learning curves for “short” distance and (Middle)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 540, + 387, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 387, + 554 + ], + "score": 1.0, + "content": "Figure 6: “medium” distance goals. (Right): Learned policies after", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 145, + 552, + 366, + 565 + ], + "spans": [ + { + "bbox": [ + 145, + 552, + 366, + 565 + ], + "score": 1.0, + "content": "fine-tuning (Top) U-maze. (Bottom): Bottleneck maze.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 99, + 578, + 183, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 185, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 185, + 594 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "We introduced UPSIDE, a novel algorithm for unsupervised skill discovery designed to trade off", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "between coverage and directedness and develop a tree of skills that can be used to both perform", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "efficient exploration of the environment and learn effective goal-directed policies. Natural venues for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "future investigation are: 1) The diffusing part of each skill could be explicitly trained to maximize", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "score": 1.0, + "content": "local coverage; 2) UPSIDE assumes a good representation of the state is provided as input, it would", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "be interesting to pair UPSIDE with effective representation learning techniques to tackle problems", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "with high-dimensional input (e.g., image-based RL); 3) While UPSIDE is grounded on the solid", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "principle of MI maximization, a more thorough theoretical investigation is needed to explicitly link", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 682, + 433, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 433, + 695 + ], + "score": 1.0, + "content": "the optimization problem and its approximations to the downstream performance.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 701, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 118, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "5Notice that if the goal was known, the learned discriminator could be directly used to identify the most", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 209, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 209, + 723 + ], + "score": 1.0, + "content": "promising skill to fine-tune.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 85, + 588, + 99, + 694 + ], + "lines": [ + { + "bbox": [ + 85, + 597, + 100, + 607 + ], + "spans": [ + { + "bbox": [ + 85, + 597, + 100, + 607 + ], + "score": 1.0, + "content": "308", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 607, + 100, + 618 + ], + "spans": [ + { + "bbox": [ + 85, + 607, + 100, + 618 + ], + "score": 1.0, + "content": "309", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 618, + 100, + 630 + ], + "spans": [ + { + "bbox": [ + 85, + 618, + 100, + 630 + ], + "score": 1.0, + "content": "310", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 630, + 100, + 641 + ], + "spans": [ + { + "bbox": [ + 85, + 630, + 100, + 641 + ], + "score": 1.0, + "content": "311", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 640, + 100, + 652 + ], + "spans": [ + { + "bbox": [ + 85, + 640, + 100, + 652 + ], + "score": 1.0, + "content": "312", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 651, + 100, + 663 + ], + "spans": [ + { + "bbox": [ + 85, + 651, + 100, + 663 + ], + "score": 1.0, + "content": "313", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 663, + 100, + 673 + ], + "spans": [ + { + "bbox": [ + 85, + 663, + 100, + 673 + ], + "score": 1.0, + "content": "314", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 673, + 100, + 684 + ], + "spans": [ + { + "bbox": [ + 85, + 673, + 100, + 684 + ], + "score": 1.0, + "content": "315", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 684, + 100, + 695 + ], + "spans": [ + { + "bbox": [ + 85, + 684, + 100, + 695 + ], + "score": 1.0, + "content": "316", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 501, + 173 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 70, + 501, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 70, + 501, + 173 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 501, + 173 + ], + "score": 0.969, + "type": "image", + "image_path": "c0fef9fb524cc8a68d444c1ca195519b14d66a4947d40fb5df6c83c782ac91ca.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 70, + 501, + 104.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 104.33333333333334, + 501, + 138.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 138.66666666666669, + 501, + 173.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 176, + 506, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 174, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 189 + ], + "score": 1.0, + "content": "Figure 5: Normalized coverage in Cartpole (Left) and Reacher (Middle). (Right) Average discrim-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 186, + 301, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 301, + 200 + ], + "score": 1.0, + "content": "inability of the skills during training in Reacher.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "index", + "bbox": [ + 86, + 211, + 505, + 300 + ], + "lines": [ + { + "bbox": [ + 86, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 86, + 213, + 99, + 223 + ], + "score": 1.0, + "content": "299", + "type": "text" + }, + { + "bbox": [ + 105, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "when reaching a particular defined goal.5 We consider goals at different distances from the initial", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 86, + 225, + 99, + 234 + ], + "score": 1.0, + "content": "300", + "type": "text" + }, + { + "bbox": [ + 105, + 223, + 127, + 235 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 127, + 224, + 137, + 234 + ], + "score": 0.83, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 223, + 505, + 235 + ], + "score": 1.0, + "content": ", the further, the harder. Fig. 6 shows the learning curves obtained when fine-tuning the best", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 84, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 84, + 235, + 100, + 245 + ], + "score": 1.0, + "content": "301", + "type": "text" + }, + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "skill for the different models and compare to a classical SAC algorithm where a single policy is", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 85, + 245, + 100, + 256 + ], + "score": 1.0, + "content": "302", + "type": "text" + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "learned from scratch. DIAYN/SMM means we use the best state-covering policies between DIAYN and", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 86, + 257, + 99, + 267 + ], + "score": 1.0, + "content": "303", + "type": "text" + }, + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "SMM. For the “close” goal setting, both UPSIDE and DIAYN/SMM are able to learn to reach this goal", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 86, + 268, + 99, + 277 + ], + "score": 1.0, + "content": "304", + "type": "text" + }, + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "efficiently while SAC solves the task only for some of the training runs. Note that we do not show", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 85, + 278, + 100, + 289 + ], + "score": 1.0, + "content": "305", + "type": "text" + }, + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "DIAYN performance since it is lower than the SMM one. For the “far” goal setting, only UPSIDE learns", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 289, + 361, + 301 + ], + "spans": [ + { + "bbox": [ + 86, + 290, + 99, + 300 + ], + "score": 1.0, + "content": "306", + "type": "text" + }, + { + "bbox": [ + 106, + 289, + 361, + 301 + ], + "score": 1.0, + "content": "to reach this goal. Obtained trajectories are illustrated in Fig. 6.", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + } + ], + "index": 8.5, + "bbox_fs": [ + 84, + 210, + 506, + 301 + ] + }, + { + "type": "image", + "bbox": [ + 110, + 309, + 495, + 528 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 309, + 495, + 528 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 309, + 495, + 528 + ], + "spans": [ + { + "bbox": [ + 110, + 309, + 495, + 528 + ], + "score": 0.945, + "type": "image", + "image_path": "5c8600e6316f678077840de51c561851a9e26dcd890e6d78ce9628180fe6febb.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 110, + 309, + 495, + 382.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 110, + 382.0, + 495, + 455.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 110, + 455.0, + 495, + 528.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 531, + 386, + 564 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 530, + 386, + 542 + ], + "spans": [ + { + "bbox": [ + 145, + 530, + 386, + 542 + ], + "score": 1.0, + "content": "(Left): Learning curves for “short” distance and (Middle)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 540, + 387, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 387, + 554 + ], + "score": 1.0, + "content": "Figure 6: “medium” distance goals. (Right): Learned policies after", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 145, + 552, + 366, + 565 + ], + "spans": [ + { + "bbox": [ + 145, + 552, + 366, + 565 + ], + "score": 1.0, + "content": "fine-tuning (Top) U-maze. (Bottom): Bottleneck maze.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 99, + 578, + 183, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 185, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 185, + 594 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "We introduced UPSIDE, a novel algorithm for unsupervised skill discovery designed to trade off", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "between coverage and directedness and develop a tree of skills that can be used to both perform", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "efficient exploration of the environment and learn effective goal-directed policies. Natural venues for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "future investigation are: 1) The diffusing part of each skill could be explicitly trained to maximize", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "score": 1.0, + "content": "local coverage; 2) UPSIDE assumes a good representation of the state is provided as input, it would", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "be interesting to pair UPSIDE with effective representation learning techniques to tackle problems", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "with high-dimensional input (e.g., image-based RL); 3) While UPSIDE is grounded on the solid", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "principle of MI maximization, a more thorough theoretical investigation is needed to explicitly link", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 682, + 433, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 433, + 695 + ], + "score": 1.0, + "content": "the optimization problem and its approximations to the downstream performance.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 104, + 595, + 506, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 93, + 68, + 507, + 729 + ], + "lines": [ + { + "bbox": [ + 98, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 98, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 90, + 507, + 104 + ], + "spans": [ + { + "bbox": [ + 110, + 90, + 507, + 104 + ], + "score": 1.0, + "content": "[1] J. Achiam, H. Edwards, D. Amodei, and P. Abbeel. Variational option discovery algorithms.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 101, + 291, + 115 + ], + "spans": [ + { + "bbox": [ + 127, + 101, + 291, + 115 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1807.10299, 2018.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 120, + 507, + 135 + ], + "spans": [ + { + "bbox": [ + 110, + 120, + 507, + 135 + ], + "score": 1.0, + "content": "[2] M. Andrychowicz, D. Crow, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 132, + 423, + 145 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 423, + 145 + ], + "score": 1.0, + "content": "P. Abbeel, and W. Zaremba. Hindsight experience replay. In NIPS, 2017.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 150, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 109, + 150, + 506, + 165 + ], + "score": 1.0, + "content": "[3] A. Aubret, L. Matignon, and S. Hassas. Elsim: End-to-end learning of reusable skills through", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 163, + 373, + 174 + ], + "spans": [ + { + "bbox": [ + 127, + 163, + 373, + 174 + ], + "score": 1.0, + "content": "intrinsic motivation. arXiv preprint arXiv:2006.12903, 2020.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 110, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "[4] A. Bagaria, J. Crowley, J. W. N. Lim, and G. Konidaris. Skill discovery for exploration and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 192, + 291, + 206 + ], + "spans": [ + { + "bbox": [ + 125, + 192, + 291, + 206 + ], + "score": 1.0, + "content": "planning using deep skill graphs. 2021.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 211, + 507, + 225 + ], + "spans": [ + { + "bbox": [ + 110, + 211, + 507, + 225 + ], + "score": 1.0, + "content": "[5] K. Baumli, D. Warde-Farley, S. Hansen, and V. Mnih. Relative variational intrinsic control.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 222, + 291, + 235 + ], + "spans": [ + { + "bbox": [ + 126, + 222, + 291, + 235 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2012.07827, 2020.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 110, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "[6] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 127, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 263, + 155, + 277 + ], + "spans": [ + { + "bbox": [ + 127, + 263, + 155, + 277 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 110, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 110, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "[7] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 294, + 208, + 306 + ], + "spans": [ + { + "bbox": [ + 127, + 294, + 208, + 306 + ], + "score": 1.0, + "content": "Openai gym, 2016.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 109, + 312, + 507, + 328 + ], + "spans": [ + { + "bbox": [ + 109, + 312, + 507, + 328 + ], + "score": 1.0, + "content": "[8] V. Campos, A. Trott, C. Xiong, R. Socher, X. Giro-i Nieto, and J. Torres. Explore, discover and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 323, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 126, + 323, + 506, + 338 + ], + "score": 1.0, + "content": "learn: Unsupervised discovery of state-covering skills. In International Conference on Machine", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 335, + 195, + 348 + ], + "spans": [ + { + "bbox": [ + 127, + 335, + 195, + 348 + ], + "score": 1.0, + "content": "Learning, 2020.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 110, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 110, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "[9] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 365, + 497, + 379 + ], + "spans": [ + { + "bbox": [ + 127, + 365, + 497, + 379 + ], + "score": 1.0, + "content": "without a reward function. In International Conference on Learning Representations, 2019.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "[10] C. Florensa, Y. Duan, and P. Abbeel. Stochastic neural networks for hierarchical reinforcement", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 395, + 329, + 408 + ], + "spans": [ + { + "bbox": [ + 126, + 395, + 329, + 408 + ], + "score": 1.0, + "content": "learning. arXiv preprint arXiv:1704.03012, 2017.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "[11] K. Gregor, D. J. Rezende, and D. Wierstra. Variational intrinsic control. arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 425, + 232, + 439 + ], + "spans": [ + { + "bbox": [ + 127, + 425, + 232, + 439 + ], + "score": 1.0, + "content": "arXiv:1611.07507, 2016.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 446, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 104, + 446, + 506, + 457 + ], + "score": 1.0, + "content": "[12] S. Gu, E. Holly, T. Lillicrap, and S. Levine. Deep reinforcement learning for robotic manipula-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 127, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "tion with asynchronous off-policy updates. In 2017 IEEE international conference on robotics", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 466, + 354, + 480 + ], + "spans": [ + { + "bbox": [ + 127, + 466, + 354, + 480 + ], + "score": 1.0, + "content": "and automation (ICRA), pages 3389–3396. IEEE, 2017.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 92, + 484, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 92, + 484, + 506, + 501 + ], + "score": 1.0, + "content": "45 [13] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine. Soft actor-critic: Off-policy maximum entropy", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 496, + 460, + 510 + ], + "spans": [ + { + "bbox": [ + 127, + 496, + 460, + 510 + ], + "score": 1.0, + "content": "deep reinforcement learning with a stochastic actor. CoRR, abs/1801.01290, 2018.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 91, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 91, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "47 [14] S. Hansen, W. Dabney, A. Barreto, D. Warde-Farley, T. Van de Wiele, and V. Mnih. Fast task", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 524, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 125, + 524, + 506, + 542 + ], + "score": 1.0, + "content": "inference with variational intrinsic successor features. In International Conference on Learning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 538, + 221, + 550 + ], + "spans": [ + { + "bbox": [ + 127, + 538, + 221, + 550 + ], + "score": 1.0, + "content": "Representations, 2019.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 92, + 554, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 92, + 554, + 506, + 573 + ], + "score": 1.0, + "content": "50 [15] E. Hazan, S. Kakade, K. Singh, and A. Van Soest. Provably efficient maximum entropy", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 568, + 484, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 568, + 484, + 582 + ], + "score": 1.0, + "content": "exploration. In International Conference on Machine Learning, pages 2681–2691, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 104, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "[16] R. Houthooft, X. Chen, Y. Duan, J. Schulman, F. De Turck, and P. Abbeel. Vime: variational", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 127, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "information maximizing exploration. In Proceedings of the 30th International Conference on", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 609, + 391, + 623 + ], + "spans": [ + { + "bbox": [ + 127, + 609, + 391, + 623 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pages 1117–1125, 2016.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "[17] Y. Jinnai, J. W. Park, D. Abel, and G. Konidaris. Discovering options for exploration by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 639, + 507, + 652 + ], + "spans": [ + { + "bbox": [ + 126, + 639, + 507, + 652 + ], + "score": 1.0, + "content": "minimizing cover time. In International Conference on Machine Learning, pages 3130–3139.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 649, + 185, + 664 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 185, + 664 + ], + "score": 1.0, + "content": "PMLR, 2019.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 668, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 668, + 506, + 684 + ], + "score": 1.0, + "content": "[18] Y. Jinnai, J. W. Park, M. C. Machado, and G. Konidaris. Exploration in reinforcement learning", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 681, + 502, + 694 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 502, + 694 + ], + "score": 1.0, + "content": "with deep covering options. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 103, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "[19] L. Lee, B. Eysenbach, E. Parisotto, E. Xing, S. Levine, and R. Salakhutdinov. Efficient", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 125, + 709, + 455, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 709, + 455, + 725 + ], + "score": 1.0, + "content": "exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 312, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 85, + 76, + 103, + 734 + ], + "lines": [ + { + "bbox": [ + 83, + 73, + 101, + 87 + ], + "spans": [ + { + "bbox": [ + 83, + 73, + 101, + 87 + ], + "score": 1.0, + "content": "317", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 91, + 101, + 103 + ], + "spans": [ + { + "bbox": [ + 84, + 91, + 101, + 103 + ], + "score": 1.0, + "content": "318", + "type": "text" + } + ] + }, + { + "bbox": [ + 83, + 102, + 101, + 116 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 101, + 116 + ], + "score": 1.0, + "content": "319", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 122, + 101, + 133 + ], + "spans": [ + { + "bbox": [ + 85, + 122, + 101, + 133 + ], + "score": 1.0, + "content": "320", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 133, + 101, + 149 + ], + "spans": [ + { + "bbox": [ + 84, + 133, + 101, + 149 + ], + "score": 1.0, + "content": "321", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 152, + 101, + 164 + ], + "spans": [ + { + "bbox": [ + 85, + 152, + 101, + 164 + ], + "score": 1.0, + "content": "322", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 162, + 101, + 177 + ], + "spans": [ + { + "bbox": [ + 84, + 162, + 101, + 177 + ], + "score": 1.0, + "content": "323", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 182, + 101, + 194 + ], + "spans": [ + { + "bbox": [ + 85, + 182, + 101, + 194 + ], + "score": 1.0, + "content": "324", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 192, + 101, + 209 + ], + "spans": [ + { + "bbox": [ + 84, + 192, + 101, + 209 + ], + "score": 1.0, + "content": "325", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 213, + 101, + 225 + ], + "spans": [ + { + "bbox": [ + 85, + 213, + 101, + 225 + ], + "score": 1.0, + "content": "326", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 223, + 101, + 239 + ], + "spans": [ + { + "bbox": [ + 84, + 223, + 101, + 239 + ], + "score": 1.0, + "content": "327", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 243, + 101, + 255 + ], + "spans": [ + { + "bbox": [ + 85, + 243, + 101, + 255 + ], + "score": 1.0, + "content": "328", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 253, + 101, + 266 + ], + "spans": [ + { + "bbox": [ + 85, + 253, + 101, + 266 + ], + "score": 1.0, + "content": "329", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 264, + 101, + 279 + ], + "spans": [ + { + "bbox": [ + 84, + 264, + 101, + 279 + ], + "score": 1.0, + "content": "330", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 284, + 100, + 296 + ], + "spans": [ + { + "bbox": [ + 85, + 284, + 100, + 296 + ], + "score": 1.0, + "content": "331", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 294, + 101, + 311 + ], + "spans": [ + { + "bbox": [ + 84, + 294, + 101, + 311 + ], + "score": 1.0, + "content": "332", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 313, + 101, + 328 + ], + "spans": [ + { + "bbox": [ + 84, + 313, + 101, + 328 + ], + "score": 1.0, + "content": "333", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 325, + 101, + 337 + ], + "spans": [ + { + "bbox": [ + 85, + 325, + 101, + 337 + ], + "score": 1.0, + "content": "334", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 336, + 101, + 351 + ], + "spans": [ + { + "bbox": [ + 84, + 336, + 101, + 351 + ], + "score": 1.0, + "content": "335", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 355, + 101, + 367 + ], + "spans": [ + { + "bbox": [ + 85, + 355, + 101, + 367 + ], + "score": 1.0, + "content": "336", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 365, + 101, + 380 + ], + "spans": [ + { + "bbox": [ + 84, + 365, + 101, + 380 + ], + "score": 1.0, + "content": "337", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 384, + 101, + 398 + ], + "spans": [ + { + "bbox": [ + 84, + 384, + 101, + 398 + ], + "score": 1.0, + "content": "338", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 396, + 101, + 412 + ], + "spans": [ + { + "bbox": [ + 84, + 396, + 101, + 412 + ], + "score": 1.0, + "content": "339", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 415, + 101, + 428 + ], + "spans": [ + { + "bbox": [ + 85, + 415, + 101, + 428 + ], + "score": 1.0, + "content": "340", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 427, + 101, + 441 + ], + "spans": [ + { + "bbox": [ + 85, + 427, + 101, + 441 + ], + "score": 1.0, + "content": "341", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 446, + 101, + 458 + ], + "spans": [ + { + "bbox": [ + 85, + 446, + 101, + 458 + ], + "score": 1.0, + "content": "342", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 457, + 101, + 469 + ], + "spans": [ + { + "bbox": [ + 85, + 457, + 101, + 469 + ], + "score": 1.0, + "content": "343", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 467, + 101, + 484 + ], + "spans": [ + { + "bbox": [ + 84, + 467, + 101, + 484 + ], + "score": 1.0, + "content": "344", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 487, + 101, + 499 + ], + "spans": [ + { + "bbox": [ + 84, + 487, + 101, + 499 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 497, + 101, + 513 + ], + "spans": [ + { + "bbox": [ + 84, + 497, + 101, + 513 + ], + "score": 1.0, + "content": "346", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 517, + 101, + 529 + ], + "spans": [ + { + "bbox": [ + 85, + 517, + 101, + 529 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + }, + { + "bbox": [ + 83, + 527, + 101, + 541 + ], + "spans": [ + { + "bbox": [ + 83, + 527, + 101, + 541 + ], + "score": 1.0, + "content": "348", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 540, + 101, + 552 + ], + "spans": [ + { + "bbox": [ + 85, + 540, + 101, + 552 + ], + "score": 1.0, + "content": "349", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 558, + 101, + 570 + ], + "spans": [ + { + "bbox": [ + 85, + 558, + 101, + 570 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 569, + 101, + 585 + ], + "spans": [ + { + "bbox": [ + 84, + 569, + 101, + 585 + ], + "score": 1.0, + "content": "351", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 588, + 101, + 600 + ], + "spans": [ + { + "bbox": [ + 84, + 588, + 101, + 600 + ], + "score": 1.0, + "content": "352", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 600, + 101, + 612 + ], + "spans": [ + { + "bbox": [ + 84, + 600, + 101, + 612 + ], + "score": 1.0, + "content": "353", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 611, + 101, + 625 + ], + "spans": [ + { + "bbox": [ + 85, + 611, + 101, + 625 + ], + "score": 1.0, + "content": "354", + "type": "text" + } + ] + }, + { + "bbox": [ + 83, + 628, + 101, + 643 + ], + "spans": [ + { + "bbox": [ + 83, + 628, + 101, + 643 + ], + "score": 1.0, + "content": "355", + "type": "text" + } + ] + }, + { + "bbox": [ + 83, + 640, + 101, + 654 + ], + "spans": [ + { + "bbox": [ + 83, + 640, + 101, + 654 + ], + "score": 1.0, + "content": "356", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 651, + 101, + 667 + ], + "spans": [ + { + "bbox": [ + 84, + 651, + 101, + 667 + ], + "score": 1.0, + "content": "357", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 671, + 101, + 683 + ], + "spans": [ + { + "bbox": [ + 85, + 671, + 101, + 683 + ], + "score": 1.0, + "content": "358", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 681, + 101, + 695 + ], + "spans": [ + { + "bbox": [ + 84, + 681, + 101, + 695 + ], + "score": 1.0, + "content": "359", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 701, + 101, + 713 + ], + "spans": [ + { + "bbox": [ + 84, + 701, + 101, + 713 + ], + "score": 1.0, + "content": "360", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 712, + 100, + 725 + ], + "spans": [ + { + "bbox": [ + 85, + 712, + 100, + 725 + ], + "score": 1.0, + "content": "361", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 93, + 68, + 507, + 729 + ], + "lines": [ + { + "bbox": [ + 98, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 98, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 90, + 507, + 104 + ], + "spans": [ + { + "bbox": [ + 110, + 90, + 507, + 104 + ], + "score": 1.0, + "content": "[1] J. Achiam, H. Edwards, D. Amodei, and P. Abbeel. Variational option discovery algorithms.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 101, + 291, + 115 + ], + "spans": [ + { + "bbox": [ + 127, + 101, + 291, + 115 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1807.10299, 2018.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 120, + 507, + 135 + ], + "spans": [ + { + "bbox": [ + 110, + 120, + 507, + 135 + ], + "score": 1.0, + "content": "[2] M. Andrychowicz, D. Crow, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 132, + 423, + 145 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 423, + 145 + ], + "score": 1.0, + "content": "P. Abbeel, and W. Zaremba. Hindsight experience replay. In NIPS, 2017.", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 150, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 109, + 150, + 506, + 165 + ], + "score": 1.0, + "content": "[3] A. Aubret, L. Matignon, and S. Hassas. Elsim: End-to-end learning of reusable skills through", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 163, + 373, + 174 + ], + "spans": [ + { + "bbox": [ + 127, + 163, + 373, + 174 + ], + "score": 1.0, + "content": "intrinsic motivation. arXiv preprint arXiv:2006.12903, 2020.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 110, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "[4] A. Bagaria, J. Crowley, J. W. N. Lim, and G. Konidaris. Skill discovery for exploration and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 192, + 291, + 206 + ], + "spans": [ + { + "bbox": [ + 125, + 192, + 291, + 206 + ], + "score": 1.0, + "content": "planning using deep skill graphs. 2021.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 211, + 507, + 225 + ], + "spans": [ + { + "bbox": [ + 110, + 211, + 507, + 225 + ], + "score": 1.0, + "content": "[5] K. Baumli, D. Warde-Farley, S. Hansen, and V. Mnih. Relative variational intrinsic control.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 222, + 291, + 235 + ], + "spans": [ + { + "bbox": [ + 126, + 222, + 291, + 235 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2012.07827, 2020.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 110, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "[6] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 127, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 263, + 155, + 277 + ], + "spans": [ + { + "bbox": [ + 127, + 263, + 155, + 277 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 110, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "[7] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 294, + 208, + 306 + ], + "spans": [ + { + "bbox": [ + 127, + 294, + 208, + 306 + ], + "score": 1.0, + "content": "Openai gym, 2016.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 312, + 507, + 328 + ], + "spans": [ + { + "bbox": [ + 109, + 312, + 507, + 328 + ], + "score": 1.0, + "content": "[8] V. Campos, A. Trott, C. Xiong, R. Socher, X. Giro-i Nieto, and J. Torres. Explore, discover and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 323, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 126, + 323, + 506, + 338 + ], + "score": 1.0, + "content": "learn: Unsupervised discovery of state-covering skills. In International Conference on Machine", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 335, + 195, + 348 + ], + "spans": [ + { + "bbox": [ + 127, + 335, + 195, + 348 + ], + "score": 1.0, + "content": "Learning, 2020.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 110, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "[9] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 365, + 497, + 379 + ], + "spans": [ + { + "bbox": [ + 127, + 365, + 497, + 379 + ], + "score": 1.0, + "content": "without a reward function. In International Conference on Learning Representations, 2019.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "[10] C. Florensa, Y. Duan, and P. Abbeel. Stochastic neural networks for hierarchical reinforcement", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 395, + 329, + 408 + ], + "spans": [ + { + "bbox": [ + 126, + 395, + 329, + 408 + ], + "score": 1.0, + "content": "learning. arXiv preprint arXiv:1704.03012, 2017.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 415, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 505, + 427 + ], + "score": 1.0, + "content": "[11] K. Gregor, D. J. Rezende, and D. Wierstra. Variational intrinsic control. arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 425, + 232, + 439 + ], + "spans": [ + { + "bbox": [ + 127, + 425, + 232, + 439 + ], + "score": 1.0, + "content": "arXiv:1611.07507, 2016.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 446, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 104, + 446, + 506, + 457 + ], + "score": 1.0, + "content": "[12] S. Gu, E. Holly, T. Lillicrap, and S. Levine. Deep reinforcement learning for robotic manipula-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 127, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "tion with asynchronous off-policy updates. In 2017 IEEE international conference on robotics", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 466, + 354, + 480 + ], + "spans": [ + { + "bbox": [ + 127, + 466, + 354, + 480 + ], + "score": 1.0, + "content": "and automation (ICRA), pages 3389–3396. IEEE, 2017.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 92, + 484, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 92, + 484, + 506, + 501 + ], + "score": 1.0, + "content": "45 [13] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine. Soft actor-critic: Off-policy maximum entropy", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 496, + 460, + 510 + ], + "spans": [ + { + "bbox": [ + 127, + 496, + 460, + 510 + ], + "score": 1.0, + "content": "deep reinforcement learning with a stochastic actor. CoRR, abs/1801.01290, 2018.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 91, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 91, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "47 [14] S. Hansen, W. Dabney, A. Barreto, D. Warde-Farley, T. Van de Wiele, and V. Mnih. Fast task", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 524, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 125, + 524, + 506, + 542 + ], + "score": 1.0, + "content": "inference with variational intrinsic successor features. In International Conference on Learning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 538, + 221, + 550 + ], + "spans": [ + { + "bbox": [ + 127, + 538, + 221, + 550 + ], + "score": 1.0, + "content": "Representations, 2019.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 92, + 554, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 92, + 554, + 506, + 573 + ], + "score": 1.0, + "content": "50 [15] E. Hazan, S. Kakade, K. Singh, and A. Van Soest. Provably efficient maximum entropy", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 568, + 484, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 568, + 484, + 582 + ], + "score": 1.0, + "content": "exploration. In International Conference on Machine Learning, pages 2681–2691, 2019.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 104, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "[16] R. Houthooft, X. Chen, Y. Duan, J. Schulman, F. De Turck, and P. Abbeel. Vime: variational", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 127, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "information maximizing exploration. In Proceedings of the 30th International Conference on", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 609, + 391, + 623 + ], + "spans": [ + { + "bbox": [ + 127, + 609, + 391, + 623 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pages 1117–1125, 2016.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "[17] Y. Jinnai, J. W. Park, D. Abel, and G. Konidaris. Discovering options for exploration by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 639, + 507, + 652 + ], + "spans": [ + { + "bbox": [ + 126, + 639, + 507, + 652 + ], + "score": 1.0, + "content": "minimizing cover time. In International Conference on Machine Learning, pages 3130–3139.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 649, + 185, + 664 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 185, + 664 + ], + "score": 1.0, + "content": "PMLR, 2019.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 668, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 668, + 506, + 684 + ], + "score": 1.0, + "content": "[18] Y. Jinnai, J. W. Park, M. C. Machado, and G. Konidaris. Exploration in reinforcement learning", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 681, + 502, + 694 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 502, + 694 + ], + "score": 1.0, + "content": "with deep covering options. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 103, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "[19] L. Lee, B. Eysenbach, E. Parisotto, E. Xing, S. Levine, and R. Salakhutdinov. Efficient", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 125, + 709, + 455, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 709, + 455, + 725 + ], + "score": 1.0, + "content": "exploration via state marginal matching. arXiv preprint arXiv:1906.05274, 2019.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + } + ], + "index": 22, + "bbox_fs": [ + 91, + 70, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 95, + 57, + 507, + 726 + ], + "lines": [ + { + "bbox": [ + 104, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "[20] K. Lu, A. Grover, P. Abbeel, and I. Mordatch. Reset-free lifelong learning with skill-space", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 83, + 331, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 331, + 97 + ], + "score": 1.0, + "content": "planning. arXiv preprint arXiv:2012.03548, 2020.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "[21] M. C. Machado, M. G. Bellemare, and M. Bowling. A laplacian framework for option discovery", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 112, + 508, + 128 + ], + "spans": [ + { + "bbox": [ + 124, + 112, + 508, + 128 + ], + "score": 1.0, + "content": "in reinforcement learning. In International Conference on Machine Learning, pages 2295–2304.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 122, + 187, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 122, + 187, + 138 + ], + "score": 1.0, + "content": "PMLR, 2017.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "[22] M. C. Machado, C. Rosenbaum, X. Guo, M. Liu, G. Tesauro, and M. Campbell. Eigenoption", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 127, + 154, + 506, + 168 + ], + "score": 1.0, + "content": "discovery through the deep successor representation. In International Conference on Learning", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 165, + 220, + 177 + ], + "spans": [ + { + "bbox": [ + 126, + 165, + 220, + 177 + ], + "score": 1.0, + "content": "Representations, 2018.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "score": 1.0, + "content": "[23] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 194, + 507, + 208 + ], + "spans": [ + { + "bbox": [ + 126, + 194, + 507, + 208 + ], + "score": 1.0, + "content": "M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep rein-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 205, + 350, + 219 + ], + "spans": [ + { + "bbox": [ + 126, + 205, + 350, + 219 + ], + "score": 1.0, + "content": "forcement learning. nature, 518(7540):529–533, 2015.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 224, + 507, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 507, + 238 + ], + "score": 1.0, + "content": "[24] N. Modhe, P. Chattopadhyay, M. Sharma, A. Das, D. Parikh, D. Batra, and R. Vedantam.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 236, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 127, + 236, + 506, + 248 + ], + "score": 1.0, + "content": "Ir-vic: Unsupervised discovery of sub-goals for transfer in rl. In Proceedings of the Twenty-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 245, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 126, + 245, + 506, + 259 + ], + "score": 1.0, + "content": "Ninth International Joint Conference on Artificial Intelligence, IJCAI-20. International Joint", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 256, + 363, + 270 + ], + "spans": [ + { + "bbox": [ + 125, + 256, + 363, + 270 + ], + "score": 1.0, + "content": "Conferences on Artificial Intelligence Organization, 2020.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "[25] S. Mohamed and D. J. Rezende. Variational information maximisation for intrinsically motivated", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 126, + 287, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 126, + 287, + 506, + 301 + ], + "score": 1.0, + "content": "reinforcement learning. In Advances in neural information processing systems, pages 2125–", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 297, + 178, + 311 + ], + "spans": [ + { + "bbox": [ + 127, + 297, + 178, + 311 + ], + "score": 1.0, + "content": "2133, 2015.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 316, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 506, + 330 + ], + "score": 1.0, + "content": "[26] M. Mutti, L. Pratissoli, and M. Restelli. A policy gradient method for task-agnostic exploration.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 327, + 290, + 341 + ], + "spans": [ + { + "bbox": [ + 125, + 327, + 290, + 341 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2007.04640, 2020.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 102, + 344, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 102, + 344, + 507, + 361 + ], + "score": 1.0, + "content": "[27] V. H. Pong, M. Dalal, S. Lin, A. Nair, S. Bahl, and S. Levine. Skew-fit: State-covering", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 124, + 354, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 124, + 354, + 506, + 373 + ], + "score": 1.0, + "content": "self-supervised reinforcement learning. In International Conference on Machine Learning,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 367, + 155, + 382 + ], + "spans": [ + { + "bbox": [ + 126, + 367, + 155, + 382 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 103, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 103, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "[28] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 398, + 339, + 411 + ], + "spans": [ + { + "bbox": [ + 125, + 398, + 339, + 411 + ], + "score": 1.0, + "content": "algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 104, + 416, + 506, + 431 + ], + "score": 1.0, + "content": "[29] R. Sekar, O. Rybkin, K. Daniilidis, P. Abbeel, D. Hafner, and D. Pathak. Planning to explore", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 125, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "via self-supervised world models. In International Conference on Machine Learning, pages", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 438, + 235, + 452 + ], + "spans": [ + { + "bbox": [ + 126, + 438, + 235, + 452 + ], + "score": 1.0, + "content": "8583–8592. PMLR, 2020.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 102, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 102, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "[30] A. Sharma, S. Gu, S. Levine, V. Kumar, and K. Hausman. Dynamics-aware unsupervised", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 469, + 466, + 483 + ], + "spans": [ + { + "bbox": [ + 127, + 469, + 466, + 483 + ], + "score": 1.0, + "content": "discovery of skills. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 103, + 488, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 103, + 488, + 506, + 500 + ], + "score": 1.0, + "content": "[31] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 496, + 508, + 513 + ], + "spans": [ + { + "bbox": [ + 125, + 496, + 508, + 513 + ], + "score": 1.0, + "content": "M. Lai, A. Bolton, et al. Mastering the game of go without human knowledge. nature,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 508, + 237, + 523 + ], + "spans": [ + { + "bbox": [ + 127, + 508, + 237, + 523 + ], + "score": 1.0, + "content": "550(7676):354–359, 2017.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "[32] E. Todorov, T. Erez, and Y. Tassa. Mujoco: A physics engine for model-based control. In 2012", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 127, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 50pou26–5033,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 127, + 550, + 154, + 563 + ], + "spans": [ + { + "bbox": [ + 127, + 550, + 154, + 563 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "[33] D. Warde-Farley, T. Van de Wiele, T. Kulkarni, C. Ionescu, S. Hansen, and V. Mnih. Unsuper-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 127, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "vised control through non-parametric discriminative rewards. In International Conference on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 125, + 590, + 259, + 604 + ], + "spans": [ + { + "bbox": [ + 125, + 590, + 259, + 604 + ], + "score": 1.0, + "content": "Learning Representations, 2019.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 103, + 609, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 103, + 609, + 506, + 623 + ], + "score": 1.0, + "content": "[34] K. Xie, H. Bharadhwaj, D. Hafner, A. Garg, and F. Shkurti. Skill transfer via partially amortized", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 621, + 477, + 633 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 477, + 633 + ], + "score": 1.0, + "content": "hierarchical planning. In International Conference on Learning Representations, 2021.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "[35] K. Xu, S. Verma, C. Finn, and S. Levine. Continual learning of control primitives: Skill", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 650, + 399, + 664 + ], + "spans": [ + { + "bbox": [ + 127, + 650, + 399, + 664 + ], + "score": 1.0, + "content": "discovery via reset-games. arXiv preprint arXiv:2011.05286, 2020.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 103, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "[36] D. Yarats, R. Fergus, A. Lazaric, and L. Pinto. Reinforcement learning with prototypical", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 126, + 680, + 357, + 693 + ], + "spans": [ + { + "bbox": [ + 126, + 680, + 357, + 693 + ], + "score": 1.0, + "content": "representations. arXiv preprint arXiv:2102.11271, 2021.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 103, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 103, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "[37] J. Zhang, H. Yu, and W. Xu. Hierarchical reinforcement learning by discovering intrinsic", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 93, + 709, + 423, + 723 + ], + "spans": [ + { + "bbox": [ + 93, + 714, + 99, + 721 + ], + "score": 1.0, + "content": "08", + "type": "text" + }, + { + "bbox": [ + 126, + 709, + 423, + 723 + ], + "score": 1.0, + "content": "options. In International Conference on Learning Representations, 2021.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 23 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "11", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 95, + 57, + 507, + 726 + ], + "lines": [ + { + "bbox": [ + 104, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "[20] K. Lu, A. Grover, P. Abbeel, and I. Mordatch. Reset-free lifelong learning with skill-space", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 83, + 331, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 331, + 97 + ], + "score": 1.0, + "content": "planning. arXiv preprint arXiv:2012.03548, 2020.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "[21] M. C. Machado, M. G. Bellemare, and M. Bowling. A laplacian framework for option discovery", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 112, + 508, + 128 + ], + "spans": [ + { + "bbox": [ + 124, + 112, + 508, + 128 + ], + "score": 1.0, + "content": "in reinforcement learning. In International Conference on Machine Learning, pages 2295–2304.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 122, + 187, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 122, + 187, + 138 + ], + "score": 1.0, + "content": "PMLR, 2017.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "[22] M. C. Machado, C. Rosenbaum, X. Guo, M. Liu, G. Tesauro, and M. Campbell. Eigenoption", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 127, + 154, + 506, + 168 + ], + "score": 1.0, + "content": "discovery through the deep successor representation. In International Conference on Learning", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 165, + 220, + 177 + ], + "spans": [ + { + "bbox": [ + 126, + 165, + 220, + 177 + ], + "score": 1.0, + "content": "Representations, 2018.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "score": 1.0, + "content": "[23] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 194, + 507, + 208 + ], + "spans": [ + { + "bbox": [ + 126, + 194, + 507, + 208 + ], + "score": 1.0, + "content": "M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep rein-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 205, + 350, + 219 + ], + "spans": [ + { + "bbox": [ + 126, + 205, + 350, + 219 + ], + "score": 1.0, + "content": "forcement learning. nature, 518(7540):529–533, 2015.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 224, + 507, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 507, + 238 + ], + "score": 1.0, + "content": "[24] N. Modhe, P. Chattopadhyay, M. Sharma, A. Das, D. Parikh, D. Batra, and R. Vedantam.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 236, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 127, + 236, + 506, + 248 + ], + "score": 1.0, + "content": "Ir-vic: Unsupervised discovery of sub-goals for transfer in rl. In Proceedings of the Twenty-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 245, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 126, + 245, + 506, + 259 + ], + "score": 1.0, + "content": "Ninth International Joint Conference on Artificial Intelligence, IJCAI-20. International Joint", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 256, + 363, + 270 + ], + "spans": [ + { + "bbox": [ + 125, + 256, + 363, + 270 + ], + "score": 1.0, + "content": "Conferences on Artificial Intelligence Organization, 2020.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "[25] S. Mohamed and D. J. Rezende. Variational information maximisation for intrinsically motivated", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 126, + 287, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 126, + 287, + 506, + 301 + ], + "score": 1.0, + "content": "reinforcement learning. In Advances in neural information processing systems, pages 2125–", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 297, + 178, + 311 + ], + "spans": [ + { + "bbox": [ + 127, + 297, + 178, + 311 + ], + "score": 1.0, + "content": "2133, 2015.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 316, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 506, + 330 + ], + "score": 1.0, + "content": "[26] M. Mutti, L. Pratissoli, and M. Restelli. A policy gradient method for task-agnostic exploration.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 327, + 290, + 341 + ], + "spans": [ + { + "bbox": [ + 125, + 327, + 290, + 341 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2007.04640, 2020.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 102, + 344, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 102, + 344, + 507, + 361 + ], + "score": 1.0, + "content": "[27] V. H. Pong, M. Dalal, S. Lin, A. Nair, S. Bahl, and S. Levine. Skew-fit: State-covering", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 124, + 354, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 124, + 354, + 506, + 373 + ], + "score": 1.0, + "content": "self-supervised reinforcement learning. In International Conference on Machine Learning,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 367, + 155, + 382 + ], + "spans": [ + { + "bbox": [ + 126, + 367, + 155, + 382 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 103, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 103, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "[28] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 398, + 339, + 411 + ], + "spans": [ + { + "bbox": [ + 125, + 398, + 339, + 411 + ], + "score": 1.0, + "content": "algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 104, + 416, + 506, + 431 + ], + "score": 1.0, + "content": "[29] R. Sekar, O. Rybkin, K. Daniilidis, P. Abbeel, D. Hafner, and D. Pathak. Planning to explore", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 125, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "via self-supervised world models. In International Conference on Machine Learning, pages", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 438, + 235, + 452 + ], + "spans": [ + { + "bbox": [ + 126, + 438, + 235, + 452 + ], + "score": 1.0, + "content": "8583–8592. PMLR, 2020.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 102, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 102, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "[30] A. Sharma, S. Gu, S. Levine, V. Kumar, and K. Hausman. Dynamics-aware unsupervised", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 469, + 466, + 483 + ], + "spans": [ + { + "bbox": [ + 127, + 469, + 466, + 483 + ], + "score": 1.0, + "content": "discovery of skills. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 103, + 488, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 103, + 488, + 506, + 500 + ], + "score": 1.0, + "content": "[31] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 496, + 508, + 513 + ], + "spans": [ + { + "bbox": [ + 125, + 496, + 508, + 513 + ], + "score": 1.0, + "content": "M. Lai, A. Bolton, et al. Mastering the game of go without human knowledge. nature,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 508, + 237, + 523 + ], + "spans": [ + { + "bbox": [ + 127, + 508, + 237, + 523 + ], + "score": 1.0, + "content": "550(7676):354–359, 2017.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "[32] E. Todorov, T. Erez, and Y. Tassa. Mujoco: A physics engine for model-based control. In 2012", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 127, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 50pou26–5033,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 127, + 550, + 154, + 563 + ], + "spans": [ + { + "bbox": [ + 127, + 550, + 154, + 563 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "[33] D. Warde-Farley, T. Van de Wiele, T. Kulkarni, C. Ionescu, S. Hansen, and V. Mnih. Unsuper-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 127, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "vised control through non-parametric discriminative rewards. In International Conference on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 125, + 590, + 259, + 604 + ], + "spans": [ + { + "bbox": [ + 125, + 590, + 259, + 604 + ], + "score": 1.0, + "content": "Learning Representations, 2019.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 103, + 609, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 103, + 609, + 506, + 623 + ], + "score": 1.0, + "content": "[34] K. Xie, H. Bharadhwaj, D. Hafner, A. Garg, and F. Shkurti. Skill transfer via partially amortized", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 621, + 477, + 633 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 477, + 633 + ], + "score": 1.0, + "content": "hierarchical planning. In International Conference on Learning Representations, 2021.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 103, + 639, + 506, + 653 + ], + "score": 1.0, + "content": "[35] K. Xu, S. Verma, C. Finn, and S. Levine. Continual learning of control primitives: Skill", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 650, + 399, + 664 + ], + "spans": [ + { + "bbox": [ + 127, + 650, + 399, + 664 + ], + "score": 1.0, + "content": "discovery via reset-games. arXiv preprint arXiv:2011.05286, 2020.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 103, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "[36] D. Yarats, R. Fergus, A. Lazaric, and L. Pinto. Reinforcement learning with prototypical", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 126, + 680, + 357, + 693 + ], + "spans": [ + { + "bbox": [ + 126, + 680, + 357, + 693 + ], + "score": 1.0, + "content": "representations. arXiv preprint arXiv:2102.11271, 2021.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 103, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 103, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "[37] J. Zhang, H. Yu, and W. Xu. Hierarchical reinforcement learning by discovering intrinsic", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 93, + 709, + 423, + 723 + ], + "spans": [ + { + "bbox": [ + 93, + 714, + 99, + 721 + ], + "score": 1.0, + "content": "08", + "type": "text" + }, + { + "bbox": [ + 126, + 709, + 423, + 723 + ], + "score": 1.0, + "content": "options. In International Conference on Learning Representations, 2021.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 23, + "bbox_fs": [ + 93, + 73, + 508, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 103 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 146, + 108, + 505, + 223 + ], + "lines": [ + { + "bbox": [ + 145, + 107, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 145, + 107, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 144, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes] We discuss the limitations and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 142, + 370, + 154 + ], + "spans": [ + { + "bbox": [ + 162, + 142, + 370, + 154 + ], + "score": 1.0, + "content": "directions of further investigation in the conclusion.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 146, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [N/A] We do", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 162, + 167, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 162, + 167, + 505, + 177 + ], + "score": 1.0, + "content": "not foresee any obvious negative societal impacts from our work, which focuses on the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 161, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 161, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "fundamentals of reinforcement learning and proposes a new algorithm for unsupervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 161, + 187, + 225, + 200 + ], + "spans": [ + { + "bbox": [ + 161, + 187, + 225, + 200 + ], + "score": 1.0, + "content": "skill discovery.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 200, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 145, + 200, + 505, + 213 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 161, + 210, + 214, + 224 + ], + "spans": [ + { + "bbox": [ + 161, + 210, + 214, + 224 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 132, + 226, + 302, + 238 + ], + "lines": [ + { + "bbox": [ + 129, + 225, + 304, + 239 + ], + "spans": [ + { + "bbox": [ + 129, + 225, + 304, + 239 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 145, + 241, + 452, + 266 + ], + "lines": [ + { + "bbox": [ + 145, + 241, + 453, + 253 + ], + "spans": [ + { + "bbox": [ + 145, + 241, + 453, + 253 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 145, + 253, + 424, + 267 + ], + "spans": [ + { + "bbox": [ + 145, + 253, + 424, + 267 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 131, + 269, + 241, + 281 + ], + "lines": [ + { + "bbox": [ + 128, + 267, + 243, + 283 + ], + "spans": [ + { + "bbox": [ + 128, + 267, + 243, + 283 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 146, + 284, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 145, + 283, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 145, + 283, + 506, + 297 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main exper-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 161, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 161, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "imental results (either in the supplemental material or as a URL)? [No] We plan to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 162, + 306, + 351, + 318 + ], + "spans": [ + { + "bbox": [ + 162, + 306, + 351, + 318 + ], + "score": 1.0, + "content": "release our code upon acceptance of this work.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 146, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 146, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 161, + 329, + 309, + 342 + ], + "spans": [ + { + "bbox": [ + 161, + 329, + 309, + 342 + ], + "score": 1.0, + "content": "were chosen)? [Yes] See Appendix.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 146, + 341, + 507, + 356 + ], + "spans": [ + { + "bbox": [ + 146, + 341, + 507, + 356 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 162, + 354, + 361, + 365 + ], + "spans": [ + { + "bbox": [ + 162, + 354, + 361, + 365 + ], + "score": 1.0, + "content": "ments multiple times)? [Yes] Yes when possible.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 146, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 146, + 365, + 505, + 380 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 162, + 377, + 367, + 389 + ], + "spans": [ + { + "bbox": [ + 162, + 377, + 367, + 389 + ], + "score": 1.0, + "content": "of GPUs, internal cluster, or cloud provider)? [No]", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 133, + 392, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 130, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 146, + 407, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 145, + 406, + 471, + 419 + ], + "spans": [ + { + "bbox": [ + 145, + 406, + 471, + 419 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] See Sect. 5.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 145, + 420, + 360, + 433 + ], + "spans": [ + { + "bbox": [ + 145, + 420, + 360, + 433 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [N/A]", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 145, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 145, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 145, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 145, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 162, + 468, + 253, + 480 + ], + "spans": [ + { + "bbox": [ + 162, + 468, + 253, + 480 + ], + "score": 1.0, + "content": "using/curating? [N/A]", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 146, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 146, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 162, + 491, + 325, + 504 + ], + "spans": [ + { + "bbox": [ + 162, + 491, + 325, + 504 + ], + "score": 1.0, + "content": "information or offensive content? [N/A]", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 131, + 506, + 432, + 518 + ], + "lines": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "spans": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 146, + 521, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "spans": [ + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "spans": [ + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "spans": [ + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 85, + 74, + 103, + 614 + ], + "lines": [ + { + "bbox": [ + 85, + 74, + 101, + 87 + ], + "spans": [ + { + "bbox": [ + 85, + 74, + 101, + 87 + ], + "score": 1.0, + "content": "409", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 93, + 101, + 107 + ], + "spans": [ + { + "bbox": [ + 85, + 93, + 101, + 107 + ], + "score": 1.0, + "content": "410", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 109, + 100, + 120 + ], + "spans": [ + { + "bbox": [ + 85, + 109, + 100, + 120 + ], + "score": 1.0, + "content": "411", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 118, + 101, + 133 + ], + "spans": [ + { + "bbox": [ + 84, + 118, + 101, + 133 + ], + "score": 1.0, + "content": "412", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 133, + 100, + 145 + ], + "spans": [ + { + "bbox": [ + 85, + 133, + 100, + 145 + ], + "score": 1.0, + "content": "413", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 143, + 101, + 156 + ], + "spans": [ + { + "bbox": [ + 84, + 143, + 101, + 156 + ], + "score": 1.0, + "content": "414", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 156, + 100, + 168 + ], + "spans": [ + { + "bbox": [ + 85, + 156, + 100, + 168 + ], + "score": 1.0, + "content": "415", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 167, + 100, + 179 + ], + "spans": [ + { + "bbox": [ + 85, + 167, + 100, + 179 + ], + "score": 1.0, + "content": "416", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 179, + 100, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 179, + 100, + 190 + ], + "score": 1.0, + "content": "417", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 189, + 101, + 202 + ], + "spans": [ + { + "bbox": [ + 85, + 189, + 101, + 202 + ], + "score": 1.0, + "content": "418", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 203, + 100, + 214 + ], + "spans": [ + { + "bbox": [ + 86, + 203, + 100, + 214 + ], + "score": 1.0, + "content": "419", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 213, + 101, + 227 + ], + "spans": [ + { + "bbox": [ + 85, + 213, + 101, + 227 + ], + "score": 1.0, + "content": "420", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 227, + 100, + 242 + ], + "spans": [ + { + "bbox": [ + 85, + 227, + 100, + 242 + ], + "score": 1.0, + "content": "421", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 243, + 100, + 255 + ], + "spans": [ + { + "bbox": [ + 86, + 243, + 100, + 255 + ], + "score": 1.0, + "content": "422", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 256, + 100, + 268 + ], + "spans": [ + { + "bbox": [ + 86, + 256, + 100, + 268 + ], + "score": 1.0, + "content": "423", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 271, + 100, + 285 + ], + "spans": [ + { + "bbox": [ + 85, + 271, + 100, + 285 + ], + "score": 1.0, + "content": "424", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 286, + 100, + 297 + ], + "spans": [ + { + "bbox": [ + 86, + 286, + 100, + 297 + ], + "score": 1.0, + "content": "425", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 297, + 100, + 308 + ], + "spans": [ + { + "bbox": [ + 85, + 297, + 100, + 308 + ], + "score": 1.0, + "content": "426", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 308, + 100, + 320 + ], + "spans": [ + { + "bbox": [ + 86, + 308, + 100, + 320 + ], + "score": 1.0, + "content": "427", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 320, + 101, + 333 + ], + "spans": [ + { + "bbox": [ + 84, + 320, + 101, + 333 + ], + "score": 1.0, + "content": "428", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 332, + 100, + 343 + ], + "spans": [ + { + "bbox": [ + 86, + 332, + 100, + 343 + ], + "score": 1.0, + "content": "429", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 344, + 100, + 356 + ], + "spans": [ + { + "bbox": [ + 85, + 344, + 100, + 356 + ], + "score": 1.0, + "content": "430", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 355, + 100, + 367 + ], + "spans": [ + { + "bbox": [ + 86, + 355, + 100, + 367 + ], + "score": 1.0, + "content": "431", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 368, + 100, + 379 + ], + "spans": [ + { + "bbox": [ + 86, + 368, + 100, + 379 + ], + "score": 1.0, + "content": "432", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 379, + 101, + 392 + ], + "spans": [ + { + "bbox": [ + 85, + 379, + 101, + 392 + ], + "score": 1.0, + "content": "433", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 394, + 100, + 407 + ], + "spans": [ + { + "bbox": [ + 85, + 394, + 100, + 407 + ], + "score": 1.0, + "content": "434", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 408, + 101, + 421 + ], + "spans": [ + { + "bbox": [ + 85, + 408, + 101, + 421 + ], + "score": 1.0, + "content": "435", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 421, + 101, + 434 + ], + "spans": [ + { + "bbox": [ + 85, + 421, + 101, + 434 + ], + "score": 1.0, + "content": "436", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 434, + 100, + 447 + ], + "spans": [ + { + "bbox": [ + 86, + 434, + 100, + 447 + ], + "score": 1.0, + "content": "437", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 446, + 100, + 457 + ], + "spans": [ + { + "bbox": [ + 85, + 446, + 100, + 457 + ], + "score": 1.0, + "content": "438", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 459, + 100, + 470 + ], + "spans": [ + { + "bbox": [ + 86, + 459, + 100, + 470 + ], + "score": 1.0, + "content": "439", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 469, + 101, + 482 + ], + "spans": [ + { + "bbox": [ + 85, + 469, + 101, + 482 + ], + "score": 1.0, + "content": "440", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 482, + 100, + 493 + ], + "spans": [ + { + "bbox": [ + 86, + 482, + 100, + 493 + ], + "score": 1.0, + "content": "441", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 493, + 100, + 506 + ], + "spans": [ + { + "bbox": [ + 85, + 493, + 100, + 506 + ], + "score": 1.0, + "content": "442", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 508, + 101, + 522 + ], + "spans": [ + { + "bbox": [ + 85, + 508, + 101, + 522 + ], + "score": 1.0, + "content": "443", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 523, + 100, + 534 + ], + "spans": [ + { + "bbox": [ + 85, + 523, + 100, + 534 + ], + "score": 1.0, + "content": "444", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 533, + 101, + 546 + ], + "spans": [ + { + "bbox": [ + 84, + 533, + 101, + 546 + ], + "score": 1.0, + "content": "445", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 546, + 101, + 559 + ], + "spans": [ + { + "bbox": [ + 84, + 546, + 101, + 559 + ], + "score": 1.0, + "content": "446", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 557, + 101, + 571 + ], + "spans": [ + { + "bbox": [ + 85, + 557, + 101, + 571 + ], + "score": 1.0, + "content": "447", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 570, + 101, + 583 + ], + "spans": [ + { + "bbox": [ + 84, + 570, + 101, + 583 + ], + "score": 1.0, + "content": "448", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 581, + 101, + 595 + ], + "spans": [ + { + "bbox": [ + 85, + 581, + 101, + 595 + ], + "score": 1.0, + "content": "449", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 602, + 101, + 615 + ], + "spans": [ + { + "bbox": [ + 85, + 602, + 101, + 615 + ], + "score": 1.0, + "content": "450", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 87, + 71, + 156, + 84 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 158, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 158, + 86 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 103 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 129, + 91, + 210, + 105 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 108, + 505, + 223 + ], + "lines": [ + { + "bbox": [ + 145, + 107, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 145, + 107, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 144, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 144, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes] We discuss the limitations and", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 142, + 370, + 154 + ], + "spans": [ + { + "bbox": [ + 162, + 142, + 370, + 154 + ], + "score": 1.0, + "content": "directions of further investigation in the conclusion.", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 146, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [N/A] We do", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 167, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 162, + 167, + 505, + 177 + ], + "score": 1.0, + "content": "not foresee any obvious negative societal impacts from our work, which focuses on the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 161, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 161, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "fundamentals of reinforcement learning and proposes a new algorithm for unsupervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 161, + 187, + 225, + 200 + ], + "spans": [ + { + "bbox": [ + 161, + 187, + 225, + 200 + ], + "score": 1.0, + "content": "skill discovery.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 200, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 145, + 200, + 505, + 213 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 210, + 214, + 224 + ], + "spans": [ + { + "bbox": [ + 161, + 210, + 214, + 224 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + } + ], + "index": 5.5, + "bbox_fs": [ + 144, + 107, + 506, + 224 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 226, + 302, + 238 + ], + "lines": [ + { + "bbox": [ + 129, + 225, + 304, + 239 + ], + "spans": [ + { + "bbox": [ + 129, + 225, + 304, + 239 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 129, + 225, + 304, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 145, + 241, + 452, + 266 + ], + "lines": [ + { + "bbox": [ + 145, + 241, + 453, + 253 + ], + "spans": [ + { + "bbox": [ + 145, + 241, + 453, + 253 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 145, + 253, + 424, + 267 + ], + "spans": [ + { + "bbox": [ + 145, + 253, + 424, + 267 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 145, + 241, + 453, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 269, + 241, + 281 + ], + "lines": [ + { + "bbox": [ + 128, + 267, + 243, + 283 + ], + "spans": [ + { + "bbox": [ + 128, + 267, + 243, + 283 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 128, + 267, + 243, + 283 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 284, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 145, + 283, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 145, + 283, + 506, + 297 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main exper-", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 161, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "imental results (either in the supplemental material or as a URL)? [No] We plan to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 162, + 306, + 351, + 318 + ], + "spans": [ + { + "bbox": [ + 162, + 306, + 351, + 318 + ], + "score": 1.0, + "content": "release our code upon acceptance of this work.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 146, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 329, + 309, + 342 + ], + "spans": [ + { + "bbox": [ + 161, + 329, + 309, + 342 + ], + "score": 1.0, + "content": "were chosen)? [Yes] See Appendix.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 341, + 507, + 356 + ], + "spans": [ + { + "bbox": [ + 146, + 341, + 507, + 356 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 354, + 361, + 365 + ], + "spans": [ + { + "bbox": [ + 162, + 354, + 361, + 365 + ], + "score": 1.0, + "content": "ments multiple times)? [Yes] Yes when possible.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 146, + 365, + 505, + 380 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 377, + 367, + 389 + ], + "spans": [ + { + "bbox": [ + 162, + 377, + 367, + 389 + ], + "score": 1.0, + "content": "of GPUs, internal cluster, or cloud provider)? [No]", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + } + ], + "index": 19, + "bbox_fs": [ + 145, + 283, + 507, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 392, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 130, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 130, + 391, + 506, + 406 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 407, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 145, + 406, + 471, + 419 + ], + "spans": [ + { + "bbox": [ + 145, + 406, + 471, + 419 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] See Sect. 5.", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 420, + 360, + 433 + ], + "spans": [ + { + "bbox": [ + 145, + 420, + 360, + 433 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [N/A]", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 145, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 145, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 145, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 468, + 253, + 480 + ], + "spans": [ + { + "bbox": [ + 162, + 468, + 253, + 480 + ], + "score": 1.0, + "content": "using/curating? [N/A]", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 146, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 491, + 325, + 504 + ], + "spans": [ + { + "bbox": [ + 162, + 491, + 325, + 504 + ], + "score": 1.0, + "content": "information or offensive content? [N/A]", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + } + ], + "index": 28, + "bbox_fs": [ + 145, + 406, + 505, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 506, + 432, + 518 + ], + "lines": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "spans": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 128, + 504, + 433, + 520 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 521, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "spans": [ + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "spans": [ + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "spans": [ + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + } + ], + "index": 35.5, + "bbox_fs": [ + 145, + 520, + 506, + 592 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_model.json b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8ee93607f34f63f69981ea9cacb3462ca339df5e --- /dev/null +++ b/parse/train/m7QWxKp2Xqd/m7QWxKp2Xqd_model.json @@ -0,0 +1,23845 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 254, + 765, + 1304, + 765, + 1304, + 1343, + 254, + 1343 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 316, + 271, + 1384, + 271, + 1384, + 382, + 316, + 382 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 251, + 1817, + 1403, + 1817, + 1403, + 1909, + 251, + 1909 + ], + "score": 0.923 + }, + { + "category_id": 8, + "poly": [ + 532, + 1914, + 1163, + 1914, + 1163, + 1991, + 532, + 1991 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 720, + 499, + 986, + 499, + 986, + 620, + 720, + 620 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 788, + 701, + 913, + 701, + 913, + 737, + 788, + 737 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 253, + 1374, + 531, + 1374, + 531, + 1412, + 253, + 1412 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 250, + 1431, + 1404, + 1431, + 1404, + 1676, + 250, + 1676 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 252, + 1685, + 1405, + 1685, + 1405, + 1809, + 252, + 1809 + ], + "score": 0.873 + }, + { + "category_id": 1, + "poly": [ + 288, + 2031, + 1388, + 2031, + 1388, + 2063, + 288, + 2063 + ], + "score": 0.834 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1920, + 1399, + 1920, + 1399, + 1949, + 1366, + 1949 + ], + "score": 0.815 + }, + { + "category_id": 13, + "poly": [ + 1234, + 1691, + 1262, + 1691, + 1262, + 1717, + 1234, + 1717 + ], + "score": 0.86, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 486, + 1880, + 508, + 1880, + 508, + 1906, + 486, + 1906 + ], + "score": 0.84, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 725, + 1880, + 749, + 1880, + 749, + 1906, + 725, + 1906 + ], + "score": 0.82, + "latex": "Z" + }, + { + "category_id": 14, + "poly": [ + 533, + 1915, + 1162, + 1915, + 1162, + 1958, + 533, + 1958 + ], + "score": 0.8, + "latex": "\\begin{array} { r } { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\\\ { \\boldsymbol { \\mathcal { Z } } ( \\boldsymbol { S } ; Z ) = \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } ) } \\end{array} - \\boldsymbol { \\mathcal { H } } ( \\boldsymbol { S } | Z ) = \\boldsymbol { \\mathcal { H } } ( Z ) - \\boldsymbol { \\mathcal { H } } ( Z | \\boldsymbol { S } ) ," + }, + { + "category_id": 15, + "poly": [ + 312.0, + 270.0, + 1385.0, + 270.0, + 1385.0, + 329.0, + 312.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 324.0, + 1351.0, + 324.0, + 1351.0, + 390.0, + 348.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 698.0, + 920.0, + 698.0, + 920.0, + 742.0, + 782.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1372.0, + 535.0, + 1372.0, + 535.0, + 1418.0, + 241.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 260.0, + 775.0, + 275.0, + 775.0, + 275.0, + 794.0, + 260.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 764.0, + 1305.0, + 764.0, + 1305.0, + 801.0, + 392.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 259.0, + 803.0, + 277.0, + 803.0, + 277.0, + 826.0, + 259.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 795.0, + 1306.0, + 795.0, + 1306.0, + 832.0, + 392.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 259.0, + 834.0, + 277.0, + 834.0, + 277.0, + 857.0, + 259.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 826.0, + 1307.0, + 826.0, + 1307.0, + 861.0, + 393.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 864.0, + 277.0, + 864.0, + 277.0, + 889.0, + 258.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 857.0, + 1305.0, + 857.0, + 1305.0, + 890.0, + 396.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 259.0, + 894.0, + 278.0, + 894.0, + 278.0, + 917.0, + 259.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 884.0, + 1307.0, + 884.0, + 1307.0, + 923.0, + 393.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 259.0, + 925.0, + 277.0, + 925.0, + 277.0, + 949.0, + 259.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 919.0, + 1306.0, + 919.0, + 1306.0, + 952.0, + 394.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 955.0, + 277.0, + 955.0, + 277.0, + 979.0, + 258.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 943.0, + 1306.0, + 943.0, + 1306.0, + 986.0, + 393.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 986.0, + 277.0, + 986.0, + 277.0, + 1009.0, + 258.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 975.0, + 1306.0, + 975.0, + 1306.0, + 1015.0, + 392.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 257.0, + 1015.0, + 279.0, + 1015.0, + 279.0, + 1041.0, + 257.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1008.0, + 1306.0, + 1008.0, + 1306.0, + 1045.0, + 393.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1046.0, + 277.0, + 1046.0, + 277.0, + 1070.0, + 251.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1039.0, + 1306.0, + 1039.0, + 1306.0, + 1075.0, + 392.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1075.0, + 276.0, + 1075.0, + 276.0, + 1101.0, + 251.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1068.0, + 1307.0, + 1068.0, + 1307.0, + 1105.0, + 393.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1103.0, + 278.0, + 1103.0, + 278.0, + 1133.0, + 248.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1099.0, + 1306.0, + 1099.0, + 1306.0, + 1135.0, + 392.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1136.0, + 277.0, + 1136.0, + 277.0, + 1162.0, + 251.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1131.0, + 1306.0, + 1131.0, + 1306.0, + 1163.0, + 394.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1166.0, + 279.0, + 1166.0, + 279.0, + 1193.0, + 248.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1158.0, + 1307.0, + 1158.0, + 1307.0, + 1196.0, + 392.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1195.0, + 279.0, + 1195.0, + 279.0, + 1223.0, + 248.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1189.0, + 1305.0, + 1189.0, + 1305.0, + 1225.0, + 392.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1226.0, + 279.0, + 1226.0, + 279.0, + 1254.0, + 248.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1218.0, + 1307.0, + 1218.0, + 1307.0, + 1256.0, + 393.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1256.0, + 278.0, + 1256.0, + 278.0, + 1285.0, + 248.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1248.0, + 1306.0, + 1248.0, + 1306.0, + 1286.0, + 393.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1288.0, + 277.0, + 1288.0, + 277.0, + 1313.0, + 250.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1282.0, + 1305.0, + 1282.0, + 1305.0, + 1315.0, + 396.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1319.0, + 277.0, + 1319.0, + 277.0, + 1344.0, + 250.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1313.0, + 1232.0, + 1313.0, + 1232.0, + 1346.0, + 396.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1823.0, + 279.0, + 1823.0, + 279.0, + 1847.0, + 251.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 1406.0, + 1817.0, + 1406.0, + 1851.0, + 292.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1855.0, + 279.0, + 1855.0, + 279.0, + 1879.0, + 249.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1885.0, + 291.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1885.0, + 279.0, + 1885.0, + 279.0, + 1909.0, + 249.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1879.0, + 485.0, + 1879.0, + 485.0, + 1912.0, + 292.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1879.0, + 724.0, + 1879.0, + 724.0, + 1912.0, + 509.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1879.0, + 1405.0, + 1879.0, + 1405.0, + 1912.0, + 750.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 496.0, + 989.0, + 496.0, + 989.0, + 533.0, + 717.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 527.0, + 913.0, + 527.0, + 913.0, + 563.0, + 787.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 557.0, + 902.0, + 557.0, + 902.0, + 592.0, + 797.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 593.0, + 890.0, + 593.0, + 890.0, + 619.0, + 808.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1436.0, + 277.0, + 1436.0, + 277.0, + 1463.0, + 248.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1432.0, + 1404.0, + 1432.0, + 1404.0, + 1466.0, + 295.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1469.0, + 278.0, + 1469.0, + 278.0, + 1496.0, + 248.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1461.0, + 1406.0, + 1461.0, + 1406.0, + 1498.0, + 294.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1498.0, + 278.0, + 1498.0, + 278.0, + 1525.0, + 248.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 1406.0, + 1490.0, + 1406.0, + 1529.0, + 293.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1528.0, + 278.0, + 1528.0, + 278.0, + 1555.0, + 247.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1521.0, + 1407.0, + 1521.0, + 1407.0, + 1559.0, + 293.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1559.0, + 278.0, + 1559.0, + 278.0, + 1587.0, + 248.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1551.0, + 1406.0, + 1551.0, + 1406.0, + 1589.0, + 294.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1589.0, + 278.0, + 1589.0, + 278.0, + 1615.0, + 247.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1584.0, + 1408.0, + 1584.0, + 1408.0, + 1618.0, + 295.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1620.0, + 278.0, + 1620.0, + 278.0, + 1646.0, + 248.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1614.0, + 1406.0, + 1614.0, + 1406.0, + 1648.0, + 295.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1651.0, + 278.0, + 1651.0, + 278.0, + 1676.0, + 248.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1645.0, + 1197.0, + 1645.0, + 1197.0, + 1679.0, + 294.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1691.0, + 278.0, + 1691.0, + 278.0, + 1718.0, + 249.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1685.0, + 1233.0, + 1685.0, + 1233.0, + 1720.0, + 293.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1720.0, + 1263.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1722.0, + 279.0, + 1722.0, + 279.0, + 1748.0, + 249.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1712.0, + 1407.0, + 1712.0, + 1407.0, + 1753.0, + 292.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1751.0, + 278.0, + 1751.0, + 278.0, + 1779.0, + 248.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1746.0, + 1405.0, + 1746.0, + 1405.0, + 1781.0, + 295.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1782.0, + 280.0, + 1782.0, + 280.0, + 1809.0, + 249.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1775.0, + 1041.0, + 1775.0, + 1041.0, + 1814.0, + 292.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2028.0, + 1389.0, + 2028.0, + 1389.0, + 2065.0, + 296.0, + 2065.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 357, + 207, + 1342, + 207, + 1342, + 516, + 357, + 516 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 296, + 538, + 1406, + 538, + 1406, + 933, + 296, + 933 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 253, + 1782, + 1404, + 1782, + 1404, + 1875, + 253, + 1875 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 255, + 1946, + 1403, + 1946, + 1403, + 2009, + 255, + 2009 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 250, + 954, + 1403, + 954, + 1403, + 1139, + 250, + 1139 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 282, + 1897, + 451, + 1897, + 451, + 1934, + 282, + 1934 + ], + "score": 0.865 + }, + { + "category_id": 1, + "poly": [ + 251, + 1148, + 1404, + 1148, + 1404, + 1363, + 251, + 1363 + ], + "score": 0.844 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2085, + 841, + 2085 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 278, + 1372, + 1405, + 1372, + 1405, + 1773, + 278, + 1773 + ], + "score": 0.599 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.112 + }, + { + "category_id": 13, + "poly": [ + 834, + 1976, + 940, + 1976, + 940, + 2009, + 834, + 2009 + ], + "score": 0.94, + "latex": "p ( s ^ { \\prime } | s , a )" + }, + { + "category_id": 13, + "poly": [ + 885, + 1048, + 999, + 1048, + 999, + 1080, + 885, + 1080 + ], + "score": 0.92, + "latex": "- { \\mathcal { H } } ( S | Z )" + }, + { + "category_id": 13, + "poly": [ + 523, + 1017, + 589, + 1017, + 589, + 1050, + 523, + 1050 + ], + "score": 0.9, + "latex": "\\mathcal { H } ( S )" + }, + { + "category_id": 13, + "poly": [ + 620, + 957, + 646, + 957, + 646, + 984, + 620, + 984 + ], + "score": 0.8, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 677, + 1977, + 701, + 1977, + 701, + 2003, + 677, + 2003 + ], + "score": 0.78, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 1094, + 544, + 1120, + 544, + 1120, + 568, + 1094, + 568 + ], + "score": 0.76, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1018, + 1277, + 1018, + 1277, + 1045, + 1255, + 1045 + ], + "score": 0.76, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 489, + 1977, + 511, + 1977, + 511, + 2004, + 489, + 2004 + ], + "score": 0.75, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 332, + 1047, + 357, + 1047, + 357, + 1075, + 332, + 1075 + ], + "score": 0.74, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 371, + 957, + 394, + 957, + 394, + 983, + 371, + 983 + ], + "score": 0.74, + "latex": "\\mathcal { T }" + }, + { + "category_id": 13, + "poly": [ + 1025, + 784, + 1059, + 784, + 1059, + 811, + 1025, + 811 + ], + "score": 0.64, + "latex": "( E )" + }, + { + "category_id": 13, + "poly": [ + 1367, + 1946, + 1401, + 1946, + 1401, + 1974, + 1367, + 1974 + ], + "score": 0.52, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 931, + 601, + 964, + 601, + 964, + 629, + 931, + 629 + ], + "score": 0.51, + "latex": "( B )" + }, + { + "category_id": 13, + "poly": [ + 660, + 723, + 695, + 723, + 695, + 750, + 660, + 750 + ], + "score": 0.49, + "latex": "( D )" + }, + { + "category_id": 13, + "poly": [ + 412, + 662, + 446, + 662, + 446, + 690, + 412, + 690 + ], + "score": 0.32, + "latex": "( C )" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 231.0, + 1201.0, + 231.0, + 1201.0, + 274.0, + 1162.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 271.0, + 849.0, + 271.0, + 849.0, + 352.0, + 777.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 344.0, + 831.0, + 344.0, + 831.0, + 451.0, + 760.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 497.0, + 462.0, + 497.0, + 462.0, + 516.0, + 443.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 494.0, + 664.0, + 494.0, + 664.0, + 518.0, + 642.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 495.0, + 861.0, + 495.0, + 861.0, + 517.0, + 842.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 495.0, + 1060.0, + 495.0, + 1060.0, + 517.0, + 1042.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 495.0, + 1258.0, + 495.0, + 1258.0, + 517.0, + 1239.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 537.0, + 1093.0, + 537.0, + 1093.0, + 572.0, + 294.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 537.0, + 1404.0, + 537.0, + 1404.0, + 572.0, + 1121.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 567.0, + 1406.0, + 567.0, + 1406.0, + 605.0, + 291.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 600.0, + 930.0, + 600.0, + 930.0, + 632.0, + 295.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 600.0, + 1404.0, + 600.0, + 1404.0, + 632.0, + 965.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 627.0, + 1406.0, + 627.0, + 1406.0, + 666.0, + 291.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 662.0, + 411.0, + 662.0, + 411.0, + 694.0, + 295.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 662.0, + 1404.0, + 662.0, + 1404.0, + 694.0, + 447.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 692.0, + 1407.0, + 692.0, + 1407.0, + 724.0, + 294.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 723.0, + 659.0, + 723.0, + 659.0, + 751.0, + 295.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 723.0, + 1403.0, + 723.0, + 1403.0, + 751.0, + 696.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 751.0, + 1404.0, + 751.0, + 1404.0, + 783.0, + 294.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 784.0, + 1024.0, + 784.0, + 1024.0, + 813.0, + 295.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 784.0, + 1403.0, + 784.0, + 1403.0, + 813.0, + 1060.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 813.0, + 1404.0, + 813.0, + 1404.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 841.0, + 1404.0, + 841.0, + 1404.0, + 875.0, + 294.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 871.0, + 1406.0, + 871.0, + 1406.0, + 908.0, + 293.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 902.0, + 1383.0, + 902.0, + 1383.0, + 938.0, + 293.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1890.0, + 457.0, + 1890.0, + 457.0, + 1945.0, + 290.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2094.0, + 839.0, + 2094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1789.0, + 277.0, + 1789.0, + 277.0, + 1814.0, + 250.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1783.0, + 1403.0, + 1783.0, + 1403.0, + 1816.0, + 293.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1820.0, + 279.0, + 1820.0, + 279.0, + 1844.0, + 250.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1814.0, + 1406.0, + 1814.0, + 1406.0, + 1848.0, + 294.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1851.0, + 279.0, + 1851.0, + 279.0, + 1876.0, + 250.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1844.0, + 1243.0, + 1844.0, + 1243.0, + 1878.0, + 294.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1951.0, + 279.0, + 1951.0, + 279.0, + 1976.0, + 249.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1946.0, + 1366.0, + 1946.0, + 1366.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1975.0, + 488.0, + 1975.0, + 488.0, + 2011.0, + 248.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1975.0, + 676.0, + 1975.0, + 676.0, + 2011.0, + 512.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1975.0, + 833.0, + 1975.0, + 833.0, + 2011.0, + 702.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1975.0, + 1405.0, + 1975.0, + 1405.0, + 2011.0, + 941.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 961.0, + 278.0, + 961.0, + 278.0, + 987.0, + 248.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 956.0, + 370.0, + 956.0, + 370.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 956.0, + 619.0, + 956.0, + 619.0, + 988.0, + 395.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 956.0, + 1405.0, + 956.0, + 1405.0, + 988.0, + 647.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 992.0, + 277.0, + 992.0, + 277.0, + 1019.0, + 248.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 982.0, + 1405.0, + 982.0, + 1405.0, + 1025.0, + 293.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1022.0, + 278.0, + 1022.0, + 278.0, + 1049.0, + 248.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1013.0, + 522.0, + 1013.0, + 522.0, + 1054.0, + 290.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1013.0, + 1254.0, + 1013.0, + 1254.0, + 1054.0, + 590.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1013.0, + 1406.0, + 1013.0, + 1406.0, + 1054.0, + 1278.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1053.0, + 278.0, + 1053.0, + 278.0, + 1080.0, + 248.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 331.0, + 1048.0, + 331.0, + 1080.0, + 294.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1048.0, + 884.0, + 1048.0, + 884.0, + 1080.0, + 358.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1048.0, + 1403.0, + 1048.0, + 1403.0, + 1080.0, + 1000.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1085.0, + 278.0, + 1085.0, + 278.0, + 1109.0, + 248.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1077.0, + 1406.0, + 1077.0, + 1406.0, + 1111.0, + 291.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1114.0, + 277.0, + 1114.0, + 277.0, + 1139.0, + 248.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1109.0, + 832.0, + 1109.0, + 832.0, + 1141.0, + 294.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1153.0, + 278.0, + 1153.0, + 278.0, + 1179.0, + 249.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1148.0, + 1403.0, + 1148.0, + 1403.0, + 1183.0, + 292.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1185.0, + 278.0, + 1185.0, + 278.0, + 1210.0, + 248.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1177.0, + 1407.0, + 1177.0, + 1407.0, + 1215.0, + 290.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1216.0, + 279.0, + 1216.0, + 279.0, + 1241.0, + 248.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1207.0, + 1404.0, + 1207.0, + 1404.0, + 1244.0, + 292.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1246.0, + 278.0, + 1246.0, + 278.0, + 1270.0, + 248.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1240.0, + 1406.0, + 1240.0, + 1406.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1276.0, + 279.0, + 1276.0, + 279.0, + 1302.0, + 249.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1270.0, + 1406.0, + 1270.0, + 1406.0, + 1305.0, + 295.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1307.0, + 279.0, + 1307.0, + 279.0, + 1331.0, + 249.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1302.0, + 1404.0, + 1302.0, + 1404.0, + 1333.0, + 295.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1337.0, + 279.0, + 1337.0, + 279.0, + 1361.0, + 249.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1331.0, + 1381.0, + 1331.0, + 1381.0, + 1365.0, + 292.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1372.0, + 1405.0, + 1372.0, + 1405.0, + 1407.0, + 301.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1437.0, + 325.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1433.0, + 1194.0, + 1433.0, + 1194.0, + 1469.0, + 323.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1464.0, + 1404.0, + 1464.0, + 1404.0, + 1502.0, + 319.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1496.0, + 1406.0, + 1496.0, + 1406.0, + 1533.0, + 323.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1527.0, + 1404.0, + 1527.0, + 1404.0, + 1561.0, + 325.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1558.0, + 1404.0, + 1558.0, + 1404.0, + 1591.0, + 325.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1587.0, + 1405.0, + 1587.0, + 1405.0, + 1621.0, + 323.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1619.0, + 824.0, + 1619.0, + 824.0, + 1652.0, + 324.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1649.0, + 1404.0, + 1649.0, + 1404.0, + 1686.0, + 303.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1681.0, + 1409.0, + 1681.0, + 1409.0, + 1716.0, + 323.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1707.0, + 1406.0, + 1707.0, + 1406.0, + 1749.0, + 323.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1739.0, + 1184.0, + 1739.0, + 1184.0, + 1779.0, + 323.0, + 1779.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 926, + 1405, + 926, + 1405, + 1049, + 296, + 1049 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 248, + 1610, + 1406, + 1610, + 1406, + 1832, + 248, + 1832 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 1091, + 1352, + 1414, + 1352, + 1414, + 1523, + 1091, + 1523 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 249, + 308, + 1405, + 308, + 1405, + 592, + 249, + 592 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 283, + 1109, + 1406, + 1109, + 1406, + 1356, + 283, + 1356 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 261, + 783, + 1403, + 783, + 1403, + 848, + 261, + 848 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 307, + 610, + 1397, + 610, + 1397, + 762, + 307, + 762 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 303, + 1412, + 1072, + 1412, + 1072, + 1600, + 303, + 1600 + ], + "score": 0.942 + }, + { + "category_id": 4, + "poly": [ + 1089, + 1533, + 1404, + 1533, + 1404, + 1595, + 1089, + 1595 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 259, + 202, + 1402, + 202, + 1402, + 295, + 259, + 295 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 292, + 1071, + 754, + 1071, + 754, + 1103, + 292, + 1103 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 293, + 874, + 645, + 874, + 645, + 912, + 293, + 912 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 295, + 1864, + 1405, + 1864, + 1405, + 2008, + 295, + 2008 + ], + "score": 0.845 + }, + { + "category_id": 1, + "poly": [ + 250, + 1369, + 988, + 1369, + 988, + 1402, + 250, + 1402 + ], + "score": 0.833 + }, + { + "category_id": 9, + "poly": [ + 1366, + 700, + 1400, + 700, + 1400, + 730, + 1366, + 730 + ], + "score": 0.729 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.644 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 860, + 2061, + 860, + 2085, + 841, + 2085 + ], + "score": 0.485 + }, + { + "category_id": 13, + "poly": [ + 990, + 1200, + 1103, + 1200, + 1103, + 1237, + 990, + 1237 + ], + "score": 0.93, + "latex": "p _ { \\pi ( z ) } ( s _ { T } )" + }, + { + "category_id": 14, + "poly": [ + 310, + 608, + 1395, + 608, + 1395, + 766, + 310, + 766 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } & { } & { \\displaystyle \\underset { | Z | , \\rho , \\pi } { \\operatorname* { m a x } } \\ : \\mathcal { Z } ( S _ { T } ; Z ) = \\mathcal { H } ( S _ { T } ) - \\mathcal { H } ( S _ { T } | Z ) = - { \\displaystyle \\sum _ { s _ { T } } } \\overline { { p } } ( s _ { T } ) \\log \\overline { { p } } ( s _ { T } ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p _ { \\pi ( z ) } ( s _ { T } ) \\right] } \\\\ & { } & { = \\mathcal { H } ( Z ) - \\mathcal { H } ( Z | S _ { T } ) = - { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\log \\rho ( z ) + { \\displaystyle \\sum _ { z \\in Z } } \\rho ( z ) \\mathbb { E } _ { s _ { T } } \\left[ \\log p ( z | s _ { T } ) \\right] , \\quad ( 1 ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 896, + 1982, + 973, + 1982, + 973, + 2009, + 896, + 2009 + ], + "score": 0.93, + "latex": "q _ { \\phi } ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 836, + 1641, + 908, + 1641, + 908, + 1675, + 836, + 1675 + ], + "score": 0.93, + "latex": "p ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 730, + 204, + 811, + 204, + 811, + 234, + 730, + 234 + ], + "score": 0.93, + "latex": "s _ { 0 } \\in S" + }, + { + "category_id": 13, + "poly": [ + 489, + 1474, + 605, + 1474, + 605, + 1507, + 489, + 1507 + ], + "score": 0.93, + "latex": "- { \\mathcal { H } } ( S | Z )" + }, + { + "category_id": 13, + "poly": [ + 668, + 400, + 723, + 400, + 723, + 433, + 668, + 433 + ], + "score": 0.93, + "latex": "\\pi ( z )" + }, + { + "category_id": 13, + "poly": [ + 557, + 1739, + 705, + 1739, + 705, + 1772, + 557, + 1772 + ], + "score": 0.93, + "latex": "\\mathcal { H } ( \\pi ( \\cdot | z , s _ { t } ) )" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1672, + 1401, + 1672, + 1401, + 1715, + 1133, + 1715 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\mathbb { E } _ { \\pi ( z ) } \\left[ \\sum _ { t = T + 1 } ^ { T + H } r _ { z } ( s _ { t } ) \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 757, + 494, + 869, + 494, + 869, + 528, + 757, + 528 + ], + "score": 0.92, + "latex": "p _ { \\pi ( z ) } ( s _ { T } )" + }, + { + "category_id": 13, + "poly": [ + 569, + 1233, + 708, + 1233, + 708, + 1267, + 569, + 1267 + ], + "score": 0.92, + "latex": "- \\mathcal { H } ( S _ { T } | Z ) )" + }, + { + "category_id": 13, + "poly": [ + 572, + 1641, + 742, + 1641, + 742, + 1675, + 572, + 1675 + ], + "score": 0.92, + "latex": "r _ { z } ( s ) \\approx p ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 333, + 554, + 675, + 554, + 675, + 592, + 333, + 592 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\bar { p } ( s _ { T } ) \\bar { ( s _ { T } ) } = \\bar { \\sum _ { z \\in Z } \\bar { \\rho ( z ) } } p _ { \\pi ( z ) } ( s _ { T } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 580, + 783, + 758, + 783, + 758, + 819, + 580, + 819 + ], + "score": 0.92, + "latex": "s _ { T } \\sim p _ { \\pi ( z ) } ( s _ { T } )" + }, + { + "category_id": 13, + "poly": [ + 769, + 1567, + 835, + 1567, + 835, + 1601, + 769, + 1601 + ], + "score": 0.91, + "latex": "\\mathcal { H } ( S )" + }, + { + "category_id": 13, + "poly": [ + 375, + 1978, + 443, + 1978, + 443, + 2008, + 375, + 2008 + ], + "score": 0.91, + "latex": "p ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 1349, + 462, + 1403, + 462, + 1403, + 494, + 1349, + 494 + ], + "score": 0.91, + "latex": "\\pi ( z )" + }, + { + "category_id": 13, + "poly": [ + 673, + 525, + 762, + 525, + 762, + 558, + 673, + 558 + ], + "score": 0.91, + "latex": "p ( z | s _ { T } )" + }, + { + "category_id": 13, + "poly": [ + 870, + 370, + 909, + 370, + 909, + 402, + 870, + 402 + ], + "score": 0.91, + "latex": "| Z |" + }, + { + "category_id": 13, + "poly": [ + 1291, + 370, + 1330, + 370, + 1330, + 402, + 1291, + 402 + ], + "score": 0.9, + "latex": "| Z |" + }, + { + "category_id": 13, + "poly": [ + 682, + 1678, + 737, + 1678, + 737, + 1711, + 682, + 1711 + ], + "score": 0.9, + "latex": "\\pi ( z )" + }, + { + "category_id": 13, + "poly": [ + 1147, + 556, + 1183, + 556, + 1183, + 585, + 1147, + 585 + ], + "score": 0.9, + "latex": "S _ { T }" + }, + { + "category_id": 13, + "poly": [ + 1351, + 401, + 1403, + 401, + 1403, + 434, + 1351, + 434 + ], + "score": 0.89, + "latex": "\\rho ( z )" + }, + { + "category_id": 13, + "poly": [ + 1286, + 432, + 1323, + 432, + 1323, + 461, + 1286, + 461 + ], + "score": 0.89, + "latex": "S _ { T }" + }, + { + "category_id": 13, + "poly": [ + 440, + 238, + 468, + 238, + 468, + 264, + 440, + 264 + ], + "score": 0.84, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1346, + 929, + 1371, + 929, + 1371, + 954, + 1346, + 954 + ], + "score": 0.84, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 362, + 497, + 391, + 497, + 391, + 523, + 362, + 523 + ], + "score": 0.83, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1269, + 406, + 1288, + 406, + 1288, + 433, + 1269, + 433 + ], + "score": 0.83, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1309, + 311, + 1333, + 311, + 1333, + 337, + 1309, + 337 + ], + "score": 0.82, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 1072, + 556, + 1095, + 556, + 1095, + 582, + 1072, + 582 + ], + "score": 0.82, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 1096, + 989, + 1123, + 989, + 1123, + 1016, + 1096, + 1016 + ], + "score": 0.81, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 923, + 465, + 941, + 465, + 941, + 492, + 923, + 492 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1314, + 528, + 1346, + 528, + 1346, + 555, + 1314, + 555 + ], + "score": 0.81, + "latex": "s _ { T }" + }, + { + "category_id": 13, + "poly": [ + 467, + 989, + 493, + 989, + 493, + 1016, + 467, + 1016 + ], + "score": 0.81, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 840, + 341, + 864, + 341, + 864, + 367, + 840, + 367 + ], + "score": 0.8, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 698, + 436, + 716, + 436, + 716, + 458, + 698, + 458 + ], + "score": 0.8, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1271, + 928, + 1295, + 928, + 1295, + 954, + 1271, + 954 + ], + "score": 0.79, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 622, + 1415, + 645, + 1415, + 645, + 1440, + 622, + 1440 + ], + "score": 0.75, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 634, + 1508, + 663, + 1508, + 663, + 1535, + 634, + 1535 + ], + "score": 0.75, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 841, + 470, + 858, + 470, + 858, + 487, + 841, + 487 + ], + "score": 0.75, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 965, + 530, + 982, + 530, + 982, + 552, + 965, + 552 + ], + "score": 0.73, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1647, + 1401, + 1647, + 1401, + 1669, + 1383, + 1669 + ], + "score": 0.72, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 938, + 407, + 955, + 407, + 955, + 428, + 938, + 428 + ], + "score": 0.72, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1616, + 1210, + 1616, + 1210, + 1638, + 1192, + 1638 + ], + "score": 0.72, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1382, + 792, + 1401, + 792, + 1401, + 811, + 1382, + 811 + ], + "score": 0.69, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 610, + 529, + 627, + 529, + 627, + 552, + 610, + 552 + ], + "score": 0.68, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 470, + 1684, + 486, + 1684, + 486, + 1705, + 470, + 1705 + ], + "score": 0.59, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1258, + 784, + 1327, + 784, + 1327, + 817, + 1258, + 817 + ], + "score": 0.58, + "latex": "| Z | , \\rho ," + }, + { + "category_id": 13, + "poly": [ + 506, + 402, + 544, + 402, + 544, + 432, + 506, + 432 + ], + "score": 0.55, + "latex": "( i i )" + }, + { + "category_id": 13, + "poly": [ + 1016, + 402, + 1061, + 402, + 1061, + 432, + 1016, + 432 + ], + "score": 0.53, + "latex": "( i i i )" + }, + { + "category_id": 13, + "poly": [ + 418, + 371, + 448, + 371, + 448, + 401, + 418, + 401 + ], + "score": 0.41, + "latex": "( i )" + }, + { + "category_id": 13, + "poly": [ + 503, + 1950, + 556, + 1950, + 556, + 1980, + 503, + 1980 + ], + "score": 0.34, + "latex": "[ 0 , 1 ]" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1353.0, + 1372.0, + 1353.0, + 1372.0, + 1374.0, + 1265.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1369.0, + 1418.0, + 1369.0, + 1418.0, + 1520.0, + 1221.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1372.0, + 1347.0, + 1372.0, + 1347.0, + 1389.0, + 1296.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1449.0, + 1234.0, + 1449.0, + 1234.0, + 1472.0, + 1129.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1467.0, + 1213.0, + 1467.0, + 1213.0, + 1487.0, + 1150.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1503.0, + 1111.0, + 1503.0, + 1111.0, + 1520.0, + 1093.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1532.0, + 1408.0, + 1532.0, + 1408.0, + 1563.0, + 1089.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1565.0, + 1301.0, + 1565.0, + 1301.0, + 1593.0, + 1087.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1065.0, + 756.0, + 1065.0, + 756.0, + 1111.0, + 294.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 873.0, + 646.0, + 873.0, + 646.0, + 920.0, + 290.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1860.0, + 1405.0, + 1860.0, + 1405.0, + 1899.0, + 332.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1892.0, + 1331.0, + 1892.0, + 1331.0, + 1923.0, + 295.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1917.0, + 1406.0, + 1917.0, + 1406.0, + 1956.0, + 331.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1949.0, + 502.0, + 1949.0, + 502.0, + 1982.0, + 294.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1949.0, + 1405.0, + 1949.0, + 1405.0, + 1982.0, + 557.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1977.0, + 374.0, + 1977.0, + 374.0, + 2008.0, + 297.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 1977.0, + 895.0, + 1977.0, + 895.0, + 2008.0, + 444.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1977.0, + 1233.0, + 1977.0, + 1233.0, + 2008.0, + 974.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 837.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 925.0, + 1270.0, + 925.0, + 1270.0, + 960.0, + 293.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 925.0, + 1345.0, + 925.0, + 1345.0, + 960.0, + 1296.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 925.0, + 1408.0, + 925.0, + 1408.0, + 960.0, + 1372.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 957.0, + 1407.0, + 957.0, + 1407.0, + 990.0, + 294.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 986.0, + 466.0, + 986.0, + 466.0, + 1021.0, + 294.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 986.0, + 1095.0, + 986.0, + 1095.0, + 1021.0, + 494.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 986.0, + 1408.0, + 986.0, + 1408.0, + 1021.0, + 1124.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1019.0, + 1309.0, + 1019.0, + 1309.0, + 1051.0, + 295.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1617.0, + 278.0, + 1617.0, + 278.0, + 1642.0, + 249.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1607.0, + 1191.0, + 1607.0, + 1191.0, + 1647.0, + 291.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1607.0, + 1410.0, + 1607.0, + 1410.0, + 1647.0, + 1211.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1647.0, + 278.0, + 1647.0, + 278.0, + 1673.0, + 248.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1639.0, + 571.0, + 1639.0, + 571.0, + 1678.0, + 294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1639.0, + 835.0, + 1639.0, + 835.0, + 1678.0, + 743.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1639.0, + 1382.0, + 1639.0, + 1382.0, + 1678.0, + 909.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1639.0, + 1408.0, + 1639.0, + 1408.0, + 1678.0, + 1402.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1668.0, + 469.0, + 1668.0, + 469.0, + 1725.0, + 286.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1668.0, + 681.0, + 1668.0, + 681.0, + 1725.0, + 487.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1668.0, + 1132.0, + 1668.0, + 1132.0, + 1725.0, + 738.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1668.0, + 1411.0, + 1668.0, + 1411.0, + 1725.0, + 1402.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1682.0, + 278.0, + 1682.0, + 278.0, + 1711.0, + 241.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1713.0, + 278.0, + 1713.0, + 278.0, + 1741.0, + 240.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1707.0, + 1407.0, + 1707.0, + 1407.0, + 1742.0, + 291.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1744.0, + 279.0, + 1744.0, + 279.0, + 1771.0, + 240.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1737.0, + 556.0, + 1737.0, + 556.0, + 1773.0, + 292.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1737.0, + 1406.0, + 1737.0, + 1406.0, + 1773.0, + 706.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1774.0, + 278.0, + 1774.0, + 278.0, + 1802.0, + 241.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1769.0, + 1406.0, + 1769.0, + 1406.0, + 1804.0, + 294.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 244.0, + 1807.0, + 278.0, + 1807.0, + 278.0, + 1830.0, + 244.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1143.0, + 1802.0, + 1143.0, + 1833.0, + 294.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 314.0, + 277.0, + 314.0, + 277.0, + 341.0, + 249.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 306.0, + 1308.0, + 306.0, + 1308.0, + 345.0, + 292.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 306.0, + 1407.0, + 306.0, + 1407.0, + 345.0, + 1334.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 345.0, + 279.0, + 345.0, + 279.0, + 371.0, + 249.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 338.0, + 839.0, + 338.0, + 839.0, + 375.0, + 293.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 338.0, + 1407.0, + 338.0, + 1407.0, + 375.0, + 865.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 375.0, + 277.0, + 375.0, + 277.0, + 402.0, + 249.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 417.0, + 367.0, + 417.0, + 405.0, + 293.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 367.0, + 869.0, + 367.0, + 869.0, + 405.0, + 449.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 367.0, + 1290.0, + 367.0, + 1290.0, + 405.0, + 910.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 367.0, + 1407.0, + 367.0, + 1407.0, + 405.0, + 1331.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 406.0, + 279.0, + 406.0, + 279.0, + 433.0, + 249.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 397.0, + 505.0, + 397.0, + 505.0, + 437.0, + 292.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 397.0, + 667.0, + 397.0, + 667.0, + 437.0, + 545.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 397.0, + 937.0, + 397.0, + 937.0, + 437.0, + 724.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 397.0, + 1015.0, + 397.0, + 1015.0, + 437.0, + 956.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 397.0, + 1268.0, + 397.0, + 1268.0, + 437.0, + 1062.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 397.0, + 1350.0, + 397.0, + 1350.0, + 437.0, + 1289.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 397.0, + 1407.0, + 397.0, + 1407.0, + 437.0, + 1404.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 435.0, + 277.0, + 435.0, + 277.0, + 463.0, + 249.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 429.0, + 697.0, + 429.0, + 697.0, + 466.0, + 293.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 429.0, + 1285.0, + 429.0, + 1285.0, + 466.0, + 717.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 429.0, + 1407.0, + 429.0, + 1407.0, + 466.0, + 1324.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 466.0, + 277.0, + 466.0, + 277.0, + 493.0, + 249.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 458.0, + 840.0, + 458.0, + 840.0, + 495.0, + 293.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 458.0, + 922.0, + 458.0, + 922.0, + 495.0, + 859.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 458.0, + 1348.0, + 458.0, + 1348.0, + 495.0, + 942.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 495.0, + 277.0, + 495.0, + 277.0, + 523.0, + 249.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 489.0, + 361.0, + 489.0, + 361.0, + 530.0, + 293.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 489.0, + 756.0, + 489.0, + 756.0, + 530.0, + 392.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 489.0, + 1408.0, + 489.0, + 1408.0, + 530.0, + 870.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 529.0, + 279.0, + 529.0, + 279.0, + 555.0, + 249.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 522.0, + 609.0, + 522.0, + 609.0, + 563.0, + 292.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 522.0, + 672.0, + 522.0, + 672.0, + 563.0, + 628.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 522.0, + 964.0, + 522.0, + 964.0, + 563.0, + 763.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 522.0, + 1313.0, + 522.0, + 1313.0, + 563.0, + 983.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 522.0, + 1408.0, + 522.0, + 1408.0, + 563.0, + 1347.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 559.0, + 277.0, + 559.0, + 277.0, + 586.0, + 249.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 549.0, + 332.0, + 549.0, + 332.0, + 597.0, + 289.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 549.0, + 1071.0, + 549.0, + 1071.0, + 597.0, + 676.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 549.0, + 1146.0, + 549.0, + 1146.0, + 597.0, + 1096.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 549.0, + 1386.0, + 549.0, + 1386.0, + 597.0, + 1184.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1106.0, + 1406.0, + 1106.0, + 1406.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1138.0, + 1409.0, + 1138.0, + 1409.0, + 1176.0, + 294.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1166.0, + 1410.0, + 1166.0, + 1410.0, + 1208.0, + 293.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1197.0, + 989.0, + 1197.0, + 989.0, + 1239.0, + 292.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1197.0, + 1406.0, + 1197.0, + 1406.0, + 1239.0, + 1104.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 568.0, + 1233.0, + 568.0, + 1270.0, + 294.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1233.0, + 1408.0, + 1233.0, + 1408.0, + 1270.0, + 709.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1264.0, + 1408.0, + 1264.0, + 1408.0, + 1298.0, + 295.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1291.0, + 1406.0, + 1291.0, + 1406.0, + 1331.0, + 293.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1323.0, + 546.0, + 1323.0, + 546.0, + 1358.0, + 291.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 781.0, + 579.0, + 781.0, + 579.0, + 821.0, + 250.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 781.0, + 1257.0, + 781.0, + 1257.0, + 821.0, + 759.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 781.0, + 1381.0, + 781.0, + 1381.0, + 821.0, + 1328.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 781.0, + 1405.0, + 781.0, + 1405.0, + 821.0, + 1402.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 810.0, + 789.0, + 810.0, + 789.0, + 850.0, + 249.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1410.0, + 621.0, + 1410.0, + 621.0, + 1447.0, + 302.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1410.0, + 1074.0, + 1410.0, + 1074.0, + 1447.0, + 646.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1444.0, + 1075.0, + 1444.0, + 1075.0, + 1478.0, + 324.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1475.0, + 488.0, + 1475.0, + 488.0, + 1507.0, + 324.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1475.0, + 741.0, + 1475.0, + 741.0, + 1507.0, + 606.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1505.0, + 633.0, + 1505.0, + 633.0, + 1542.0, + 303.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1505.0, + 1074.0, + 1505.0, + 1074.0, + 1542.0, + 664.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1535.0, + 1075.0, + 1535.0, + 1075.0, + 1571.0, + 323.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1567.0, + 768.0, + 1567.0, + 768.0, + 1604.0, + 322.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1567.0, + 972.0, + 1567.0, + 972.0, + 1604.0, + 836.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 212.0, + 279.0, + 212.0, + 279.0, + 233.0, + 256.0, + 233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 201.0, + 729.0, + 201.0, + 729.0, + 239.0, + 291.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 201.0, + 1405.0, + 201.0, + 1405.0, + 239.0, + 812.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 242.0, + 276.0, + 242.0, + 276.0, + 263.0, + 256.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 234.0, + 439.0, + 234.0, + 439.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 234.0, + 1404.0, + 234.0, + 1404.0, + 268.0, + 469.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 273.0, + 276.0, + 273.0, + 276.0, + 294.0, + 256.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 263.0, + 1048.0, + 263.0, + 1048.0, + 300.0, + 292.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1365.0, + 993.0, + 1365.0, + 993.0, + 1409.0, + 242.0, + 1409.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 241, + 1591, + 1405, + 1591, + 1405, + 1902, + 241, + 1902 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 241, + 806, + 1405, + 806, + 1405, + 992, + 241, + 992 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 243, + 402, + 1405, + 402, + 1405, + 497, + 243, + 497 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 239, + 509, + 1404, + 509, + 1404, + 795, + 239, + 795 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 242, + 202, + 1404, + 202, + 1404, + 326, + 242, + 326 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 593, + 1154, + 1110, + 1154, + 1110, + 1219, + 593, + 1219 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 241, + 1238, + 1405, + 1238, + 1405, + 1395, + 241, + 1395 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 435, + 1011, + 1261, + 1011, + 1261, + 1082, + 435, + 1082 + ], + "score": 0.94 + }, + { + "category_id": 8, + "poly": [ + 520, + 1518, + 1181, + 1518, + 1181, + 1573, + 520, + 1573 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 296, + 1921, + 1405, + 1921, + 1405, + 2008, + 296, + 2008 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 242, + 1101, + 1188, + 1101, + 1188, + 1135, + 242, + 1135 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 292, + 355, + 754, + 355, + 754, + 390, + 292, + 390 + ], + "score": 0.912 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1167, + 1400, + 1167, + 1400, + 1198, + 1366, + 1198 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1523, + 1400, + 1523, + 1400, + 1553, + 1366, + 1553 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1022, + 1400, + 1022, + 1400, + 1051, + 1366, + 1051 + ], + "score": 0.861 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2084, + 840, + 2084 + ], + "score": 0.776 + }, + { + "category_id": 1, + "poly": [ + 239, + 1406, + 1409, + 1406, + 1409, + 1500, + 239, + 1500 + ], + "score": 0.511 + }, + { + "category_id": 13, + "poly": [ + 664, + 1658, + 749, + 1658, + 749, + 1688, + 664, + 1688 + ], + "score": 0.94, + "latex": "P _ { 1 } = 0" + }, + { + "category_id": 13, + "poly": [ + 295, + 1656, + 406, + 1656, + 406, + 1691, + 295, + 1691 + ], + "score": 0.93, + "latex": "( P _ { N } ) _ { N \\ge 1 }" + }, + { + "category_id": 14, + "poly": [ + 591, + 1152, + 1108, + 1152, + 1108, + 1219, + 591, + 1219 + ], + "score": 0.93, + "latex": "\\operatorname* { m a x } _ { | Z | = N , \\rho } \\bigg \\{ \\mathcal { H } ( Z ) + \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\bigg \\} ," + }, + { + "category_id": 13, + "poly": [ + 1073, + 1331, + 1338, + 1331, + 1338, + 1366, + 1073, + 1366 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\operatorname* { m a x } _ { \\rho } \\mathcal { H } ( Z ) \\bar { = } \\log ( N ) ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 989, + 640, + 1090, + 640, + 1090, + 674, + 989, + 674 + ], + "score": 0.92, + "latex": "p ( z | s _ { \\mathrm { d i f f } } )" + }, + { + "category_id": 13, + "poly": [ + 753, + 1240, + 847, + 1240, + 847, + 1274, + 753, + 1274 + ], + "score": 0.92, + "latex": "z \\in [ N ]" + }, + { + "category_id": 13, + "poly": [ + 324, + 600, + 749, + 600, + 749, + 643, + 324, + 643 + ], + "score": 0.92, + "latex": "p _ { \\pi ( z ) } ( s _ { \\mathrm { d i f f } } ) = 1 / \\dot { H } \\sum _ { t = T + 1 } ^ { T + H } p _ { \\pi ( z ) } ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1187, + 761, + 1288, + 761, + 1288, + 795, + 1187, + 795 + ], + "score": 0.92, + "latex": "| Z | \\ll S" + }, + { + "category_id": 14, + "poly": [ + 432, + 1008, + 1262, + 1008, + 1262, + 1085, + 432, + 1085 + ], + "score": 0.91, + "latex": "- \\mathcal { H } ( Z | S _ { \\mathrm { d i f f } } ) = \\sum _ { z \\in Z } \\rho ( z ) \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\operatorname* { m i n } _ { z \\in Z } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] ," + }, + { + "category_id": 13, + "poly": [ + 968, + 869, + 1083, + 869, + 1083, + 902, + 968, + 902 + ], + "score": 0.91, + "latex": "- { \\mathcal { H } } ( Z | S )" + }, + { + "category_id": 13, + "poly": [ + 1304, + 1301, + 1372, + 1301, + 1372, + 1334, + 1304, + 1334 + ], + "score": 0.91, + "latex": "\\mathcal { H } ( Z )" + }, + { + "category_id": 13, + "poly": [ + 1345, + 403, + 1397, + 403, + 1397, + 437, + 1345, + 437 + ], + "score": 0.91, + "latex": "\\rho ( z )" + }, + { + "category_id": 14, + "poly": [ + 520, + 1517, + 1182, + 1517, + 1182, + 1573, + 520, + 1573 + ], + "score": 0.9, + "latex": "\\operatorname* { m a x } _ { N \\geq 1 } \\log ( N ) \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] \\geq \\log \\eta ." + }, + { + "category_id": 13, + "poly": [ + 758, + 1925, + 866, + 1925, + 866, + 1953, + 758, + 1953 + ], + "score": 0.89, + "latex": "| Z | = | S |" + }, + { + "category_id": 13, + "poly": [ + 772, + 1624, + 1183, + 1624, + 1183, + 1659, + 772, + 1659 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { P _ { N } : = \\operatorname* { m i n } _ { z \\in [ N ] } \\mathbb { E } _ { s _ { \\mathrm { d i f f } } } \\left[ \\log p ( z | s _ { \\mathrm { d i f f } } ) \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1045, + 404, + 1103, + 404, + 1103, + 436, + 1045, + 436 + ], + "score": 0.88, + "latex": "( | Z | )" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1922, + 1350, + 1922, + 1350, + 1953, + 1011, + 1953 + ], + "score": 0.87, + "latex": "{ \\mathcal { T } } ( X , Y ) = { \\mathcal { T } } ( X , X ) = { \\mathcal { H } } ( X ) )" + }, + { + "category_id": 13, + "poly": [ + 690, + 571, + 740, + 571, + 740, + 600, + 690, + 600 + ], + "score": 0.87, + "latex": "S _ { \\mathrm { d i f f } }" + }, + { + "category_id": 13, + "poly": [ + 624, + 675, + 670, + 675, + 670, + 703, + 624, + 703 + ], + "score": 0.87, + "latex": "s _ { \\mathrm { d i f f } }" + }, + { + "category_id": 13, + "poly": [ + 944, + 1363, + 971, + 1363, + 971, + 1389, + 944, + 1389 + ], + "score": 0.85, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 368, + 1600, + 387, + 1600, + 387, + 1626, + 368, + 1626 + ], + "score": 0.84, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1370, + 1242, + 1398, + 1242, + 1398, + 1268, + 1370, + 1268 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 919, + 1333, + 948, + 1333, + 948, + 1359, + 919, + 1359 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1156, + 1242, + 1180, + 1242, + 1180, + 1268, + 1156, + 1268 + ], + "score": 0.83, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 1134, + 235, + 1158, + 235, + 1158, + 261, + 1134, + 261 + ], + "score": 0.83, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 712, + 1719, + 740, + 1719, + 740, + 1745, + 712, + 1745 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1600, + 1253, + 1600, + 1253, + 1626, + 1235, + 1626 + ], + "score": 0.83, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 785, + 1409, + 813, + 1409, + 813, + 1435, + 785, + 1435 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 542, + 1337, + 561, + 1337, + 561, + 1364, + 542, + 1364 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1723, + 1146, + 1723, + 1146, + 1749, + 1129, + 1749 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1303, + 512, + 1325, + 512, + 1325, + 539, + 1303, + 539 + ], + "score": 0.78, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 326, + 676, + 344, + 676, + 344, + 699, + 326, + 699 + ], + "score": 0.75, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 886, + 1109, + 905, + 1109, + 905, + 1130, + 886, + 1130 + ], + "score": 0.75, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1084, + 1599, + 1101, + 1599, + 1101, + 1621, + 1084, + 1621 + ], + "score": 0.74, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1383, + 575, + 1402, + 575, + 1402, + 598, + 1383, + 598 + ], + "score": 0.73, + "latex": "z" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1917.0, + 757.0, + 1917.0, + 757.0, + 1956.0, + 328.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1917.0, + 1010.0, + 1917.0, + 1010.0, + 1956.0, + 867.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1917.0, + 1410.0, + 1917.0, + 1410.0, + 1956.0, + 1351.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1949.0, + 1405.0, + 1949.0, + 1405.0, + 1982.0, + 294.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1976.0, + 1349.0, + 1976.0, + 1349.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 758.0, + 351.0, + 758.0, + 396.0, + 293.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 863.0, + 2058.0, + 863.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1598.0, + 278.0, + 1598.0, + 278.0, + 1626.0, + 240.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1591.0, + 367.0, + 1591.0, + 367.0, + 1627.0, + 292.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1591.0, + 1083.0, + 1591.0, + 1083.0, + 1627.0, + 388.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1591.0, + 1234.0, + 1591.0, + 1234.0, + 1627.0, + 1102.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1591.0, + 1406.0, + 1591.0, + 1406.0, + 1627.0, + 1254.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1628.0, + 279.0, + 1628.0, + 279.0, + 1657.0, + 238.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1618.0, + 771.0, + 1618.0, + 771.0, + 1664.0, + 288.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1618.0, + 1409.0, + 1618.0, + 1409.0, + 1664.0, + 1184.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1660.0, + 278.0, + 1660.0, + 278.0, + 1690.0, + 238.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1657.0, + 663.0, + 1657.0, + 663.0, + 1693.0, + 407.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1693.0, + 750.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1693.0, + 279.0, + 1693.0, + 279.0, + 1721.0, + 238.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1686.0, + 1406.0, + 1686.0, + 1406.0, + 1722.0, + 291.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1722.0, + 278.0, + 1722.0, + 278.0, + 1752.0, + 238.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1718.0, + 711.0, + 1718.0, + 711.0, + 1750.0, + 295.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1718.0, + 1128.0, + 1718.0, + 1128.0, + 1750.0, + 741.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1718.0, + 1405.0, + 1718.0, + 1405.0, + 1750.0, + 1147.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1752.0, + 279.0, + 1752.0, + 279.0, + 1781.0, + 238.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1749.0, + 1408.0, + 1749.0, + 1408.0, + 1781.0, + 294.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1782.0, + 279.0, + 1782.0, + 279.0, + 1812.0, + 238.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1777.0, + 1406.0, + 1777.0, + 1406.0, + 1813.0, + 294.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1813.0, + 279.0, + 1813.0, + 279.0, + 1842.0, + 238.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1806.0, + 1408.0, + 1806.0, + 1408.0, + 1844.0, + 292.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1843.0, + 279.0, + 1843.0, + 279.0, + 1871.0, + 238.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1836.0, + 1406.0, + 1836.0, + 1406.0, + 1874.0, + 292.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1876.0, + 278.0, + 1876.0, + 278.0, + 1901.0, + 240.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1868.0, + 1368.0, + 1868.0, + 1368.0, + 1906.0, + 292.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 813.0, + 278.0, + 813.0, + 278.0, + 838.0, + 241.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 808.0, + 1404.0, + 808.0, + 1404.0, + 840.0, + 295.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 843.0, + 279.0, + 843.0, + 279.0, + 871.0, + 238.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 835.0, + 1406.0, + 835.0, + 1406.0, + 873.0, + 292.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 875.0, + 278.0, + 875.0, + 278.0, + 899.0, + 241.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 866.0, + 967.0, + 866.0, + 967.0, + 903.0, + 292.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 866.0, + 1408.0, + 866.0, + 1408.0, + 903.0, + 1084.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 903.0, + 279.0, + 903.0, + 279.0, + 931.0, + 238.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 895.0, + 1408.0, + 895.0, + 1408.0, + 936.0, + 291.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 933.0, + 279.0, + 933.0, + 279.0, + 962.0, + 238.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 930.0, + 1406.0, + 930.0, + 1406.0, + 962.0, + 295.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 964.0, + 278.0, + 964.0, + 278.0, + 992.0, + 240.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 958.0, + 873.0, + 958.0, + 873.0, + 996.0, + 292.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 243.0, + 411.0, + 276.0, + 411.0, + 276.0, + 434.0, + 243.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 397.0, + 1044.0, + 397.0, + 1044.0, + 441.0, + 286.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 397.0, + 1344.0, + 397.0, + 1344.0, + 441.0, + 1104.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 397.0, + 1409.0, + 397.0, + 1409.0, + 441.0, + 1398.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 243.0, + 442.0, + 276.0, + 442.0, + 276.0, + 464.0, + 243.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 434.0, + 1408.0, + 434.0, + 1408.0, + 471.0, + 292.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 470.0, + 278.0, + 470.0, + 278.0, + 497.0, + 242.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 463.0, + 1309.0, + 463.0, + 1309.0, + 502.0, + 290.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 514.0, + 279.0, + 514.0, + 279.0, + 543.0, + 239.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 510.0, + 1302.0, + 510.0, + 1302.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 510.0, + 1404.0, + 510.0, + 1404.0, + 543.0, + 1326.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 546.0, + 279.0, + 546.0, + 279.0, + 572.0, + 240.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 541.0, + 1404.0, + 541.0, + 1404.0, + 574.0, + 294.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 574.0, + 279.0, + 574.0, + 279.0, + 604.0, + 239.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 569.0, + 689.0, + 569.0, + 689.0, + 606.0, + 293.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 569.0, + 1382.0, + 569.0, + 1382.0, + 606.0, + 741.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 569.0, + 1407.0, + 569.0, + 1407.0, + 606.0, + 1403.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 610.0, + 279.0, + 610.0, + 279.0, + 640.0, + 239.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 597.0, + 323.0, + 597.0, + 323.0, + 652.0, + 290.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 597.0, + 1409.0, + 597.0, + 1409.0, + 652.0, + 750.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 644.0, + 279.0, + 644.0, + 279.0, + 673.0, + 238.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 637.0, + 988.0, + 637.0, + 988.0, + 677.0, + 292.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 637.0, + 1405.0, + 637.0, + 1405.0, + 677.0, + 1091.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 677.0, + 277.0, + 677.0, + 277.0, + 703.0, + 240.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 668.0, + 325.0, + 668.0, + 325.0, + 707.0, + 292.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 668.0, + 623.0, + 668.0, + 623.0, + 707.0, + 345.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 668.0, + 1407.0, + 668.0, + 1407.0, + 707.0, + 671.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 707.0, + 277.0, + 707.0, + 277.0, + 732.0, + 240.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 702.0, + 1405.0, + 702.0, + 1405.0, + 735.0, + 294.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 735.0, + 279.0, + 735.0, + 279.0, + 765.0, + 239.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 727.0, + 1407.0, + 727.0, + 1407.0, + 769.0, + 292.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 767.0, + 277.0, + 767.0, + 277.0, + 793.0, + 240.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 758.0, + 1186.0, + 758.0, + 1186.0, + 797.0, + 292.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 758.0, + 1372.0, + 758.0, + 1372.0, + 797.0, + 1289.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 207.0, + 280.0, + 207.0, + 280.0, + 237.0, + 239.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 237.0, + 293.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 239.0, + 277.0, + 239.0, + 277.0, + 265.0, + 241.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 233.0, + 1133.0, + 233.0, + 1133.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 233.0, + 1405.0, + 233.0, + 1405.0, + 268.0, + 1159.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 267.0, + 279.0, + 267.0, + 279.0, + 298.0, + 239.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 263.0, + 1405.0, + 263.0, + 1405.0, + 298.0, + 295.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 300.0, + 279.0, + 300.0, + 279.0, + 326.0, + 241.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 295.0, + 1036.0, + 295.0, + 1036.0, + 330.0, + 295.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1246.0, + 278.0, + 1246.0, + 278.0, + 1271.0, + 241.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1239.0, + 752.0, + 1239.0, + 752.0, + 1272.0, + 292.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1239.0, + 1155.0, + 1239.0, + 1155.0, + 1272.0, + 848.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1239.0, + 1369.0, + 1239.0, + 1369.0, + 1272.0, + 1181.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1239.0, + 1409.0, + 1239.0, + 1409.0, + 1272.0, + 1399.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1276.0, + 277.0, + 1276.0, + 277.0, + 1302.0, + 241.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1271.0, + 1408.0, + 1271.0, + 1408.0, + 1304.0, + 295.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1307.0, + 275.0, + 1307.0, + 275.0, + 1333.0, + 241.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1303.0, + 1300.0, + 1303.0, + 1338.0, + 294.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1300.0, + 1406.0, + 1300.0, + 1406.0, + 1338.0, + 1373.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1338.0, + 278.0, + 1338.0, + 278.0, + 1363.0, + 240.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 541.0, + 1330.0, + 541.0, + 1367.0, + 294.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1330.0, + 918.0, + 1330.0, + 918.0, + 1367.0, + 562.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1330.0, + 1072.0, + 1330.0, + 1072.0, + 1367.0, + 949.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 1330.0, + 1406.0, + 1330.0, + 1406.0, + 1367.0, + 1339.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1367.0, + 278.0, + 1367.0, + 278.0, + 1393.0, + 242.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1362.0, + 943.0, + 1362.0, + 943.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1362.0, + 982.0, + 1362.0, + 982.0, + 1395.0, + 972.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1099.0, + 885.0, + 1099.0, + 885.0, + 1139.0, + 238.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1099.0, + 1189.0, + 1099.0, + 1189.0, + 1139.0, + 906.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1413.0, + 283.0, + 1413.0, + 283.0, + 1438.0, + 241.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1404.0, + 784.0, + 1404.0, + 784.0, + 1443.0, + 290.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1443.0, + 814.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1445.0, + 278.0, + 1445.0, + 278.0, + 1468.0, + 242.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1435.0, + 1405.0, + 1435.0, + 1405.0, + 1474.0, + 288.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 243.0, + 1475.0, + 279.0, + 1475.0, + 279.0, + 1498.0, + 243.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1468.0, + 839.0, + 1468.0, + 839.0, + 1502.0, + 292.0, + 1502.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1554, + 1403, + 1554, + 1403, + 1711, + 297, + 1711 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 292, + 1099, + 1404, + 1099, + 1404, + 1284, + 292, + 1284 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 287, + 1296, + 1403, + 1296, + 1403, + 1540, + 287, + 1540 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 293, + 1784, + 1399, + 1784, + 1399, + 1847, + 293, + 1847 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 270, + 1860, + 1402, + 1860, + 1402, + 1924, + 270, + 1924 + ], + "score": 0.93 + }, + { + "category_id": 2, + "poly": [ + 296, + 1950, + 1403, + 1950, + 1403, + 2008, + 296, + 2008 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 251, + 1052, + 800, + 1052, + 800, + 1086, + 251, + 1086 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 300, + 1739, + 634, + 1739, + 634, + 1770, + 300, + 1770 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 271, + 249, + 1340, + 249, + 1340, + 985, + 271, + 985 + ], + "score": 0.574 + }, + { + "category_id": 0, + "poly": [ + 298, + 211, + 548, + 211, + 548, + 242, + 298, + 242 + ], + "score": 0.573 + }, + { + "category_id": 5, + "poly": [ + 271, + 249, + 1340, + 249, + 1340, + 985, + 271, + 985 + ], + "score": 0.179, + "html": "
Initialize:Discriminability threshold η ∈(O,1),branching factor No ≥1,patience K Initialize: Tree T initialized as a root node indexed by O,queue of parent nodes W = {0}. while W ≠のdo// tree expansion
Dequeue a node/skill w ∈ W and expand Tat w by adding a set C(w) of No nodes/skills
2Create random policies πz,∀z ∈C(w)
3Initialize discriminator q with |T|classes
4Continue = true;Saturated = false
5while Continue do
6forK iterations do
7Sampleaskill z fromTatrandom
8Extract the sequence of nodes z(1),...,z in T leading to z
9 Execute the composed (directed part) policy (Tz(1),..,πz) followed by the diffusing part
10Add states observed during the diffusion part to state buffer Bz
11Update discriminator q with SGD on Bz to predict label z
12 13if z ∈C(w) then// Update only new policies,other polices kept fixed
14Update policy Tz using SAC to optimize the discriminator reward as in Sect. 3.1. Compute the skil-discriminability d(z)=@B)(z) = [ ∑s∈Bzq(z|s)for allz ∈C(w)
15if minz∈c(w)d(z) <n then// Node removal
16Remove the node/skill z = arg minz∈c(w)d(z) from C(w) and T
17Set Saturate = true
18elseifnot Saturated then
19Add one new node/skill to C(w) and T
20else
21Set Continue = false
22
Enqueue in W the consolidated nodes C(w)
" + }, + { + "category_id": 6, + "poly": [ + 298, + 211, + 548, + 211, + 548, + 242, + 298, + 242 + ], + "score": 0.156 + }, + { + "category_id": 13, + "poly": [ + 733, + 699, + 1130, + 699, + 1130, + 740, + 733, + 740 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { d ( z ) = \\widehat { q } _ { \\phi } ^ { \\scriptscriptstyle ( B ) } ( z ) = \\frac { 1 } { | \\mathscr { B } _ { z } | } \\sum _ { s \\in \\mathscr { B } _ { z } } q _ { \\phi } ( z | s ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 981, + 1892, + 1065, + 1892, + 1065, + 1925, + 981, + 1925 + ], + "score": 0.93, + "latex": "q _ { \\phi } ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 565, + 1892, + 638, + 1892, + 638, + 1924, + 565, + 1924 + ], + "score": 0.92, + "latex": "p ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 1139, + 1191, + 1222, + 1191, + 1222, + 1220, + 1139, + 1220 + ], + "score": 0.91, + "latex": "T + H" + }, + { + "category_id": 13, + "poly": [ + 447, + 643, + 547, + 643, + 547, + 672, + 447, + 672 + ], + "score": 0.91, + "latex": "z \\in \\mathcal { C } ( w )" + }, + { + "category_id": 13, + "poly": [ + 675, + 251, + 779, + 251, + 779, + 281, + 675, + 281 + ], + "score": 0.91, + "latex": "\\eta \\in ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1199, + 702, + 1300, + 702, + 1300, + 733, + 1199, + 733 + ], + "score": 0.91, + "latex": "z \\in \\mathcal { C } ( w )" + }, + { + "category_id": 13, + "poly": [ + 993, + 334, + 1049, + 334, + 1049, + 364, + 993, + 364 + ], + "score": 0.91, + "latex": "\\mathcal { C } ( w )" + }, + { + "category_id": 13, + "poly": [ + 873, + 556, + 1028, + 556, + 1028, + 590, + 873, + 590 + ], + "score": 0.91, + "latex": "( \\pi _ { z _ { ( 1 ) } } , \\ldots , \\pi _ { z } )" + }, + { + "category_id": 13, + "poly": [ + 1101, + 277, + 1203, + 277, + 1203, + 308, + 1101, + 308 + ], + "score": 0.91, + "latex": "\\mathcal { W } = \\{ 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 965, + 766, + 1022, + 766, + 1022, + 796, + 965, + 796 + ], + "score": 0.9, + "latex": "\\mathcal { C } ( w )" + }, + { + "category_id": 13, + "poly": [ + 620, + 361, + 736, + 361, + 736, + 390, + 620, + 390 + ], + "score": 0.9, + "latex": "\\forall z \\in \\mathcal { C } ( \\bar { \\boldsymbol { w } } )" + }, + { + "category_id": 13, + "poly": [ + 296, + 1585, + 440, + 1585, + 440, + 1619, + 296, + 1619 + ], + "score": 0.9, + "latex": "( i T + H ) _ { i \\geq 1 }" + }, + { + "category_id": 13, + "poly": [ + 960, + 252, + 1040, + 252, + 1040, + 279, + 960, + 279 + ], + "score": 0.9, + "latex": "N _ { 0 } \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 694, + 851, + 750, + 851, + 750, + 881, + 694, + 881 + ], + "score": 0.9, + "latex": "\\mathcal { C } ( w )" + }, + { + "category_id": 13, + "poly": [ + 557, + 335, + 639, + 335, + 639, + 360, + 557, + 360 + ], + "score": 0.89, + "latex": "w \\in \\mathcal W" + }, + { + "category_id": 13, + "poly": [ + 729, + 531, + 841, + 531, + 841, + 558, + 729, + 558 + ], + "score": 0.89, + "latex": "z _ { ( 1 ) } , \\dotsc , z" + }, + { + "category_id": 13, + "poly": [ + 656, + 390, + 693, + 390, + 693, + 419, + 656, + 419 + ], + "score": 0.88, + "latex": "| \\tau |" + }, + { + "category_id": 13, + "poly": [ + 811, + 617, + 843, + 617, + 843, + 643, + 811, + 643 + ], + "score": 0.88, + "latex": "B _ { z }" + }, + { + "category_id": 13, + "poly": [ + 574, + 393, + 601, + 393, + 601, + 419, + 574, + 419 + ], + "score": 0.87, + "latex": "q _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 726, + 947, + 783, + 947, + 783, + 978, + 726, + 978 + ], + "score": 0.87, + "latex": "\\mathcal { C } ( w )" + }, + { + "category_id": 13, + "poly": [ + 1024, + 589, + 1054, + 589, + 1054, + 615, + 1024, + 615 + ], + "score": 0.86, + "latex": "B _ { z }" + }, + { + "category_id": 13, + "poly": [ + 643, + 620, + 670, + 620, + 670, + 646, + 643, + 646 + ], + "score": 0.86, + "latex": "q _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 614, + 675, + 643, + 675, + 643, + 699, + 614, + 699 + ], + "score": 0.85, + "latex": "\\pi _ { z }" + }, + { + "category_id": 13, + "poly": [ + 1080, + 335, + 1114, + 335, + 1114, + 361, + 1080, + 361 + ], + "score": 0.85, + "latex": "N _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 577, + 365, + 606, + 365, + 606, + 388, + 577, + 388 + ], + "score": 0.84, + "latex": "\\pi _ { z }" + }, + { + "category_id": 13, + "poly": [ + 1265, + 1892, + 1285, + 1892, + 1285, + 1922, + 1265, + 1922 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1325, + 1422, + 1353, + 1422, + 1353, + 1448, + 1325, + 1448 + ], + "score": 0.84, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 870, + 528, + 895, + 528, + 895, + 552, + 870, + 552 + ], + "score": 0.83, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 629, + 1222, + 658, + 1222, + 658, + 1248, + 629, + 1248 + ], + "score": 0.83, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 1370, + 1195, + 1398, + 1195, + 1398, + 1222, + 1370, + 1222 + ], + "score": 0.83, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 538, + 1680, + 562, + 1680, + 562, + 1706, + 538, + 1706 + ], + "score": 0.83, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 761, + 335, + 785, + 335, + 785, + 360, + 761, + 360 + ], + "score": 0.83, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 612, + 1680, + 640, + 1680, + 640, + 1706, + 612, + 1706 + ], + "score": 0.82, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 655, + 767, + 907, + 767, + 907, + 798, + 655, + 798 + ], + "score": 0.82, + "latex": "z = \\arg \\operatorname* { m i n } _ { z \\in { \\mathcal { C } } ( w ) } d ( z )" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1106, + 1401, + 1106, + 1401, + 1131, + 1373, + 1131 + ], + "score": 0.82, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1172, + 1331, + 1191, + 1331, + 1191, + 1359, + 1172, + 1359 + ], + "score": 0.81, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 457, + 279, + 483, + 279, + 483, + 304, + 457, + 304 + ], + "score": 0.81, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 411, + 736, + 624, + 736, + 624, + 769, + 411, + 769 + ], + "score": 0.8, + "latex": "\\begin{array} { r } { \\operatorname* { m i n } _ { z \\in \\mathcal { C } ( w ) } d ( z ) < \\eta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 647, + 501, + 671, + 501, + 671, + 526, + 647, + 526 + ], + "score": 0.8, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 557, + 1222, + 581, + 1222, + 581, + 1249, + 557, + 1249 + ], + "score": 0.79, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 690, + 421, + 713, + 421, + 713, + 441, + 690, + 441 + ], + "score": 0.78, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 420, + 472, + 447, + 472, + 447, + 497, + 420, + 497 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 796, + 851, + 821, + 851, + 821, + 877, + 796, + 877 + ], + "score": 0.78, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 811, + 339, + 833, + 339, + 833, + 359, + 811, + 359 + ], + "score": 0.78, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 1067, + 766, + 1093, + 766, + 1093, + 792, + 1067, + 792 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 881, + 1619, + 905, + 1619, + 905, + 1646, + 881, + 1646 + ], + "score": 0.77, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 572, + 504, + 589, + 504, + 589, + 525, + 572, + 525 + ], + "score": 0.77, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 364, + 305, + 444, + 305, + 444, + 335, + 364, + 335 + ], + "score": 0.77, + "latex": "\\mathcal { W } \\neq \\emptyset" + }, + { + "category_id": 13, + "poly": [ + 954, + 1619, + 983, + 1619, + 983, + 1646, + 954, + 1646 + ], + "score": 0.76, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 577, + 797, + 601, + 797, + 601, + 817, + 577, + 817 + ], + "score": 0.76, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 454, + 420, + 478, + 420, + 478, + 441, + 454, + 441 + ], + "score": 0.76, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1002, + 621, + 1019, + 621, + 1019, + 641, + 1002, + 641 + ], + "score": 0.75, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 549, + 1483, + 567, + 1483, + 567, + 1506, + 549, + 1506 + ], + "score": 0.75, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1140, + 253, + 1166, + 253, + 1166, + 276, + 1140, + 276 + ], + "score": 0.74, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 768, + 1257, + 788, + 1257, + 788, + 1279, + 768, + 1279 + ], + "score": 0.71, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1003, + 534, + 1018, + 534, + 1018, + 551, + 1003, + 551 + ], + "score": 0.7, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 458, + 947, + 490, + 947, + 490, + 973, + 458, + 973 + ], + "score": 0.6, + "latex": "\\mathcal { W }" + }, + { + "category_id": 13, + "poly": [ + 578, + 911, + 601, + 911, + 601, + 931, + 578, + 931 + ], + "score": 0.58, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 605, + 644, + 636, + 644, + 636, + 670, + 605, + 670 + ], + "score": 0.44, + "latex": "/ /" + }, + { + "category_id": 13, + "poly": [ + 680, + 738, + 711, + 738, + 711, + 764, + 680, + 764 + ], + "score": 0.44, + "latex": "/ /" + }, + { + "category_id": 13, + "poly": [ + 859, + 766, + 907, + 766, + 907, + 796, + 859, + 796 + ], + "score": 0.36, + "latex": "d ( z )" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1943.0, + 1407.0, + 1943.0, + 1407.0, + 1985.0, + 329.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 964.0, + 1975.0, + 964.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 1049.0, + 804.0, + 1049.0, + 804.0, + 1095.0, + 245.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1733.0, + 636.0, + 1733.0, + 636.0, + 1777.0, + 292.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2092.0, + 838.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 208.0, + 551.0, + 208.0, + 551.0, + 247.0, + 295.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 208.0, + 551.0, + 208.0, + 551.0, + 247.0, + 295.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1555.0, + 1404.0, + 1555.0, + 1404.0, + 1588.0, + 293.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1583.0, + 1405.0, + 1583.0, + 1405.0, + 1620.0, + 441.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1614.0, + 880.0, + 1614.0, + 880.0, + 1657.0, + 292.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1614.0, + 953.0, + 1614.0, + 953.0, + 1657.0, + 906.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1614.0, + 1405.0, + 1614.0, + 1405.0, + 1657.0, + 984.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1685.0, + 295.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1678.0, + 537.0, + 1678.0, + 537.0, + 1714.0, + 293.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1678.0, + 611.0, + 1678.0, + 611.0, + 1714.0, + 563.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1678.0, + 868.0, + 1678.0, + 868.0, + 1714.0, + 641.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1096.0, + 1372.0, + 1096.0, + 1372.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1136.0, + 1402.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1128.0, + 1406.0, + 1128.0, + 1406.0, + 1165.0, + 293.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1160.0, + 1406.0, + 1160.0, + 1406.0, + 1195.0, + 294.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1186.0, + 1138.0, + 1186.0, + 1138.0, + 1228.0, + 291.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 1186.0, + 1369.0, + 1186.0, + 1369.0, + 1228.0, + 1223.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1186.0, + 1409.0, + 1186.0, + 1409.0, + 1228.0, + 1399.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1216.0, + 556.0, + 1216.0, + 556.0, + 1259.0, + 291.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1216.0, + 628.0, + 1216.0, + 628.0, + 1259.0, + 582.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1216.0, + 1407.0, + 1216.0, + 1407.0, + 1259.0, + 659.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1251.0, + 767.0, + 1251.0, + 767.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1251.0, + 1363.0, + 1251.0, + 1363.0, + 1287.0, + 789.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1297.0, + 1404.0, + 1297.0, + 1404.0, + 1331.0, + 295.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1327.0, + 1171.0, + 1327.0, + 1171.0, + 1360.0, + 295.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1327.0, + 1405.0, + 1327.0, + 1405.0, + 1360.0, + 1192.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1391.0, + 295.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1417.0, + 1324.0, + 1417.0, + 1324.0, + 1453.0, + 293.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1417.0, + 1406.0, + 1417.0, + 1406.0, + 1453.0, + 1354.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1449.0, + 1405.0, + 1449.0, + 1405.0, + 1482.0, + 297.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1476.0, + 548.0, + 1476.0, + 548.0, + 1515.0, + 293.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1476.0, + 1406.0, + 1476.0, + 1406.0, + 1515.0, + 568.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1507.0, + 1234.0, + 1507.0, + 1234.0, + 1543.0, + 294.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1782.0, + 1404.0, + 1782.0, + 1404.0, + 1819.0, + 295.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1814.0, + 1403.0, + 1814.0, + 1403.0, + 1850.0, + 296.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 264.0, + 1857.0, + 1405.0, + 1857.0, + 1405.0, + 1896.0, + 264.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1887.0, + 564.0, + 1887.0, + 564.0, + 1928.0, + 292.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1887.0, + 980.0, + 1887.0, + 980.0, + 1928.0, + 639.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1887.0, + 1264.0, + 1887.0, + 1264.0, + 1928.0, + 1066.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1887.0, + 1406.0, + 1887.0, + 1406.0, + 1928.0, + 1286.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 248.0, + 674.0, + 248.0, + 674.0, + 281.0, + 294.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 248.0, + 959.0, + 248.0, + 959.0, + 281.0, + 780.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 248.0, + 1139.0, + 248.0, + 1139.0, + 281.0, + 1041.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 248.0, + 1170.0, + 248.0, + 1170.0, + 281.0, + 1167.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 274.0, + 456.0, + 274.0, + 456.0, + 310.0, + 294.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 274.0, + 1100.0, + 274.0, + 1100.0, + 310.0, + 484.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 274.0, + 1212.0, + 274.0, + 1212.0, + 310.0, + 1204.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 302.0, + 363.0, + 302.0, + 363.0, + 337.0, + 294.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 302.0, + 709.0, + 302.0, + 709.0, + 337.0, + 445.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 331.0, + 556.0, + 331.0, + 556.0, + 366.0, + 337.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 331.0, + 760.0, + 331.0, + 760.0, + 366.0, + 640.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 331.0, + 810.0, + 331.0, + 810.0, + 366.0, + 786.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 331.0, + 992.0, + 331.0, + 992.0, + 366.0, + 834.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 331.0, + 1079.0, + 331.0, + 1079.0, + 366.0, + 1050.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 331.0, + 1243.0, + 331.0, + 1243.0, + 366.0, + 1115.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 369.0, + 290.0, + 369.0, + 290.0, + 386.0, + 276.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 359.0, + 576.0, + 359.0, + 576.0, + 394.0, + 335.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 359.0, + 619.0, + 359.0, + 619.0, + 394.0, + 607.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 359.0, + 740.0, + 359.0, + 740.0, + 394.0, + 737.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 397.0, + 290.0, + 397.0, + 290.0, + 414.0, + 276.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 388.0, + 573.0, + 388.0, + 573.0, + 419.0, + 338.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 388.0, + 655.0, + 388.0, + 655.0, + 419.0, + 602.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 388.0, + 773.0, + 388.0, + 773.0, + 419.0, + 694.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 425.0, + 290.0, + 425.0, + 290.0, + 442.0, + 276.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 415.0, + 453.0, + 415.0, + 453.0, + 447.0, + 338.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 415.0, + 689.0, + 415.0, + 689.0, + 447.0, + 479.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 415.0, + 791.0, + 415.0, + 791.0, + 447.0, + 714.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 450.0, + 290.0, + 450.0, + 290.0, + 470.0, + 276.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 444.0, + 546.0, + 444.0, + 546.0, + 471.0, + 338.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 478.0, + 294.0, + 478.0, + 294.0, + 500.0, + 276.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 470.0, + 419.0, + 470.0, + 419.0, + 501.0, + 379.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 470.0, + 583.0, + 470.0, + 583.0, + 501.0, + 448.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 505.0, + 295.0, + 505.0, + 295.0, + 527.0, + 276.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 499.0, + 571.0, + 499.0, + 571.0, + 530.0, + 422.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 499.0, + 646.0, + 499.0, + 646.0, + 530.0, + 590.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 499.0, + 781.0, + 499.0, + 781.0, + 530.0, + 672.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 532.0, + 295.0, + 532.0, + 295.0, + 555.0, + 276.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 522.0, + 728.0, + 522.0, + 728.0, + 565.0, + 419.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 522.0, + 869.0, + 522.0, + 869.0, + 565.0, + 842.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 522.0, + 1002.0, + 522.0, + 1002.0, + 565.0, + 896.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 522.0, + 1028.0, + 522.0, + 1028.0, + 565.0, + 1019.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 560.0, + 295.0, + 560.0, + 295.0, + 583.0, + 276.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 550.0, + 872.0, + 550.0, + 872.0, + 592.0, + 418.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 550.0, + 1338.0, + 550.0, + 1338.0, + 592.0, + 1029.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 268.0, + 591.0, + 296.0, + 591.0, + 296.0, + 617.0, + 268.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 587.0, + 1023.0, + 587.0, + 1023.0, + 618.0, + 424.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 587.0, + 1058.0, + 587.0, + 1058.0, + 618.0, + 1055.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 619.0, + 295.0, + 619.0, + 295.0, + 646.0, + 267.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 616.0, + 642.0, + 616.0, + 642.0, + 647.0, + 424.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 616.0, + 810.0, + 616.0, + 810.0, + 647.0, + 671.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 616.0, + 1001.0, + 616.0, + 1001.0, + 647.0, + 844.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 616.0, + 1023.0, + 616.0, + 1023.0, + 647.0, + 1020.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 268.0, + 647.0, + 295.0, + 647.0, + 295.0, + 673.0, + 268.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 639.0, + 446.0, + 639.0, + 446.0, + 675.0, + 420.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 639.0, + 604.0, + 639.0, + 604.0, + 675.0, + 548.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 639.0, + 1305.0, + 639.0, + 1305.0, + 675.0, + 637.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 673.0, + 298.0, + 673.0, + 298.0, + 703.0, + 267.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 669.0, + 613.0, + 669.0, + 613.0, + 700.0, + 462.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 669.0, + 1280.0, + 669.0, + 1280.0, + 700.0, + 644.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 707.0, + 294.0, + 707.0, + 294.0, + 734.0, + 267.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 698.0, + 732.0, + 698.0, + 732.0, + 740.0, + 379.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 698.0, + 1198.0, + 698.0, + 1198.0, + 740.0, + 1131.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 698.0, + 1305.0, + 698.0, + 1305.0, + 740.0, + 1301.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 741.0, + 294.0, + 741.0, + 294.0, + 768.0, + 267.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 731.0, + 410.0, + 731.0, + 410.0, + 774.0, + 376.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 731.0, + 679.0, + 731.0, + 679.0, + 774.0, + 625.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 731.0, + 884.0, + 731.0, + 884.0, + 774.0, + 712.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 770.0, + 296.0, + 770.0, + 296.0, + 795.0, + 267.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 760.0, + 654.0, + 760.0, + 654.0, + 801.0, + 418.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 760.0, + 964.0, + 760.0, + 964.0, + 801.0, + 908.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 760.0, + 1066.0, + 760.0, + 1066.0, + 801.0, + 1023.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 795.0, + 296.0, + 795.0, + 296.0, + 825.0, + 266.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 791.0, + 576.0, + 791.0, + 576.0, + 824.0, + 423.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 791.0, + 666.0, + 791.0, + 666.0, + 824.0, + 602.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 827.0, + 294.0, + 827.0, + 294.0, + 852.0, + 267.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 821.0, + 679.0, + 821.0, + 679.0, + 853.0, + 379.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 855.0, + 295.0, + 855.0, + 295.0, + 879.0, + 266.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 847.0, + 693.0, + 847.0, + 693.0, + 881.0, + 421.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 847.0, + 795.0, + 847.0, + 795.0, + 881.0, + 751.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 883.0, + 294.0, + 883.0, + 294.0, + 908.0, + 266.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 878.0, + 430.0, + 878.0, + 430.0, + 908.0, + 381.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 909.0, + 296.0, + 909.0, + 296.0, + 940.0, + 266.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 906.0, + 577.0, + 906.0, + 577.0, + 937.0, + 422.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 906.0, + 679.0, + 906.0, + 679.0, + 937.0, + 602.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 264.0, + 949.0, + 294.0, + 949.0, + 294.0, + 979.0, + 264.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 943.0, + 457.0, + 943.0, + 457.0, + 979.0, + 336.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 943.0, + 725.0, + 943.0, + 725.0, + 979.0, + 491.0, + 979.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 237, + 1520, + 1405, + 1520, + 1405, + 2008, + 237, + 2008 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 240, + 422, + 1405, + 422, + 1405, + 819, + 240, + 819 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 239, + 833, + 1405, + 833, + 1405, + 1138, + 239, + 1138 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 240, + 202, + 1403, + 202, + 1403, + 329, + 240, + 329 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 286, + 1236, + 1405, + 1236, + 1405, + 1511, + 286, + 1511 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 636, + 348, + 1069, + 348, + 1069, + 404, + 636, + 404 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 251, + 1164, + 540, + 1164, + 540, + 1201, + 251, + 1201 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 353, + 1400, + 353, + 1400, + 382, + 1366, + 382 + ], + "score": 0.798 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 859, + 2062, + 859, + 2085, + 840, + 2085 + ], + "score": 0.772 + }, + { + "category_id": 9, + "poly": [ + 1366, + 353, + 1400, + 353, + 1400, + 382, + 1366, + 382 + ], + "score": 0.155 + }, + { + "category_id": 13, + "poly": [ + 1194, + 1733, + 1304, + 1733, + 1304, + 1767, + 1194, + 1767 + ], + "score": 0.93, + "latex": "\\mathcal { T } ( s ^ { \\prime } ; z | s )" + }, + { + "category_id": 14, + "poly": [ + 630, + 347, + 1072, + 347, + 1072, + 403, + 630, + 403 + ], + "score": 0.93, + "latex": "\\operatorname* { m a x } _ { N \\geq 1 , \\pi } N \\qquad \\mathrm { s . t . } \\qquad \\operatorname* { m i n } _ { z \\in [ N ] } \\widehat { q } _ { \\phi } ^ { ( B ) } ( z ) \\geq \\eta ." + }, + { + "category_id": 13, + "poly": [ + 1168, + 236, + 1247, + 236, + 1247, + 272, + 1168, + 272 + ], + "score": 0.92, + "latex": "\\widehat { q } _ { \\phi } ^ { ( B ) } ( z )" + }, + { + "category_id": 13, + "poly": [ + 760, + 1824, + 811, + 1824, + 811, + 1857, + 760, + 1857 + ], + "score": 0.91, + "latex": "p ( s )" + }, + { + "category_id": 13, + "poly": [ + 1092, + 1582, + 1131, + 1582, + 1131, + 1614, + 1092, + 1614 + ], + "score": 0.89, + "latex": "| Z |" + }, + { + "category_id": 13, + "poly": [ + 446, + 1764, + 470, + 1764, + 470, + 1791, + 446, + 1791 + ], + "score": 0.85, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 821, + 268, + 847, + 268, + 847, + 295, + 821, + 295 + ], + "score": 0.84, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1231, + 1018, + 1249, + 1018, + 1249, + 1047, + 1231, + 1047 + ], + "score": 0.82, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 466, + 732, + 484, + 732, + 484, + 759, + 466, + 759 + ], + "score": 0.82, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1364, + 487, + 1401, + 487, + 1401, + 515, + 1364, + 515 + ], + "score": 0.82, + "latex": "N _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1016, + 1197, + 1016, + 1197, + 1043, + 1168, + 1043 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 654, + 1767, + 673, + 1767, + 673, + 1791, + 654, + 1791 + ], + "score": 0.77, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1218, + 488, + 1242, + 488, + 1242, + 512, + 1218, + 512 + ], + "score": 0.77, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 328, + 303, + 344, + 303, + 344, + 324, + 328, + 324 + ], + "score": 0.73, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 703, + 1077, + 731, + 1077, + 731, + 1103, + 703, + 1103 + ], + "score": 0.71, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1160, + 204, + 1217, + 204, + 1217, + 236, + 1160, + 236 + ], + "score": 0.7, + "latex": "[ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1288, + 1770, + 1303, + 1770, + 1303, + 1791, + 1288, + 1791 + ], + "score": 0.68, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1115, + 731, + 1139, + 731, + 1139, + 755, + 1115, + 755 + ], + "score": 0.67, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 853, + 240, + 869, + 240, + 869, + 261, + 853, + 261 + ], + "score": 0.59, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1161.0, + 546.0, + 1161.0, + 546.0, + 1205.0, + 241.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 862.0, + 2060.0, + 862.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1526.0, + 277.0, + 1526.0, + 277.0, + 1551.0, + 240.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1519.0, + 1406.0, + 1519.0, + 1406.0, + 1554.0, + 293.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1558.0, + 277.0, + 1558.0, + 277.0, + 1583.0, + 240.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1550.0, + 1406.0, + 1550.0, + 1406.0, + 1585.0, + 291.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1588.0, + 277.0, + 1588.0, + 277.0, + 1613.0, + 240.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1580.0, + 1091.0, + 1580.0, + 1091.0, + 1617.0, + 294.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1580.0, + 1406.0, + 1580.0, + 1406.0, + 1617.0, + 1132.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1618.0, + 277.0, + 1618.0, + 277.0, + 1643.0, + 240.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1605.0, + 1406.0, + 1605.0, + 1406.0, + 1651.0, + 291.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1649.0, + 278.0, + 1649.0, + 278.0, + 1673.0, + 240.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1643.0, + 1406.0, + 1643.0, + 1406.0, + 1676.0, + 295.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1677.0, + 279.0, + 1677.0, + 279.0, + 1706.0, + 239.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1673.0, + 1406.0, + 1673.0, + 1406.0, + 1706.0, + 294.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1708.0, + 279.0, + 1708.0, + 279.0, + 1736.0, + 239.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1702.0, + 1407.0, + 1702.0, + 1407.0, + 1739.0, + 294.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1738.0, + 279.0, + 1738.0, + 279.0, + 1767.0, + 237.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1732.0, + 1193.0, + 1732.0, + 1193.0, + 1769.0, + 294.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1732.0, + 1406.0, + 1732.0, + 1406.0, + 1769.0, + 1305.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1768.0, + 278.0, + 1768.0, + 278.0, + 1797.0, + 239.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1763.0, + 445.0, + 1763.0, + 445.0, + 1799.0, + 294.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1763.0, + 653.0, + 1763.0, + 653.0, + 1799.0, + 471.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1763.0, + 1287.0, + 1763.0, + 1287.0, + 1799.0, + 674.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 1763.0, + 1405.0, + 1763.0, + 1405.0, + 1799.0, + 1304.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1796.0, + 280.0, + 1796.0, + 280.0, + 1828.0, + 239.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1793.0, + 1406.0, + 1793.0, + 1406.0, + 1830.0, + 294.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1830.0, + 277.0, + 1830.0, + 277.0, + 1855.0, + 240.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1823.0, + 759.0, + 1823.0, + 759.0, + 1856.0, + 295.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1823.0, + 1403.0, + 1823.0, + 1403.0, + 1856.0, + 812.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1861.0, + 277.0, + 1861.0, + 277.0, + 1886.0, + 240.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1853.0, + 1406.0, + 1853.0, + 1406.0, + 1890.0, + 294.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1889.0, + 277.0, + 1889.0, + 277.0, + 1919.0, + 239.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1885.0, + 1406.0, + 1885.0, + 1406.0, + 1921.0, + 294.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1919.0, + 278.0, + 1919.0, + 278.0, + 1948.0, + 240.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1914.0, + 1406.0, + 1914.0, + 1406.0, + 1952.0, + 293.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1950.0, + 278.0, + 1950.0, + 278.0, + 1979.0, + 239.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 1407.0, + 1944.0, + 1407.0, + 1981.0, + 294.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1982.0, + 277.0, + 1982.0, + 277.0, + 2007.0, + 240.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1973.0, + 1149.0, + 1973.0, + 1149.0, + 2013.0, + 291.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 428.0, + 278.0, + 428.0, + 278.0, + 456.0, + 240.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 422.0, + 1405.0, + 422.0, + 1405.0, + 459.0, + 293.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 459.0, + 278.0, + 459.0, + 278.0, + 487.0, + 240.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 456.0, + 1405.0, + 456.0, + 1405.0, + 488.0, + 294.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 491.0, + 277.0, + 491.0, + 277.0, + 516.0, + 241.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 479.0, + 1217.0, + 479.0, + 1217.0, + 522.0, + 290.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 479.0, + 1363.0, + 479.0, + 1363.0, + 522.0, + 1243.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 479.0, + 1406.0, + 479.0, + 1406.0, + 522.0, + 1402.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 520.0, + 278.0, + 520.0, + 278.0, + 548.0, + 240.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 513.0, + 1406.0, + 513.0, + 1406.0, + 549.0, + 293.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 552.0, + 278.0, + 552.0, + 278.0, + 576.0, + 241.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 547.0, + 1405.0, + 547.0, + 1405.0, + 579.0, + 294.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 580.0, + 277.0, + 580.0, + 277.0, + 609.0, + 239.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 575.0, + 1406.0, + 575.0, + 1406.0, + 611.0, + 293.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 610.0, + 280.0, + 610.0, + 280.0, + 639.0, + 240.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 605.0, + 1406.0, + 605.0, + 1406.0, + 641.0, + 293.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 641.0, + 278.0, + 641.0, + 278.0, + 670.0, + 239.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 636.0, + 1406.0, + 636.0, + 1406.0, + 672.0, + 293.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 672.0, + 278.0, + 672.0, + 278.0, + 697.0, + 240.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 662.0, + 1408.0, + 662.0, + 1408.0, + 706.0, + 290.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 701.0, + 280.0, + 701.0, + 280.0, + 729.0, + 239.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 694.0, + 1406.0, + 694.0, + 1406.0, + 732.0, + 290.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 733.0, + 278.0, + 733.0, + 278.0, + 758.0, + 241.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 725.0, + 465.0, + 725.0, + 465.0, + 762.0, + 293.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 725.0, + 1114.0, + 725.0, + 1114.0, + 762.0, + 485.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 725.0, + 1406.0, + 725.0, + 1406.0, + 762.0, + 1140.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 762.0, + 278.0, + 762.0, + 278.0, + 790.0, + 240.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 756.0, + 1406.0, + 756.0, + 1406.0, + 793.0, + 293.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 794.0, + 278.0, + 794.0, + 278.0, + 819.0, + 241.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 786.0, + 1074.0, + 786.0, + 1074.0, + 824.0, + 292.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 840.0, + 277.0, + 840.0, + 277.0, + 864.0, + 240.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 832.0, + 1406.0, + 832.0, + 1406.0, + 868.0, + 293.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 869.0, + 279.0, + 869.0, + 279.0, + 897.0, + 239.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 863.0, + 1408.0, + 863.0, + 1408.0, + 898.0, + 296.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 898.0, + 277.0, + 898.0, + 277.0, + 926.0, + 239.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 893.0, + 1409.0, + 893.0, + 1409.0, + 929.0, + 293.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 929.0, + 279.0, + 929.0, + 279.0, + 956.0, + 239.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 923.0, + 1408.0, + 923.0, + 1408.0, + 959.0, + 295.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 960.0, + 279.0, + 960.0, + 279.0, + 988.0, + 239.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 954.0, + 1407.0, + 954.0, + 1407.0, + 991.0, + 292.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 992.0, + 277.0, + 992.0, + 277.0, + 1016.0, + 240.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1018.0, + 296.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1020.0, + 279.0, + 1020.0, + 279.0, + 1048.0, + 239.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1015.0, + 1167.0, + 1015.0, + 1167.0, + 1050.0, + 295.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1015.0, + 1230.0, + 1015.0, + 1230.0, + 1050.0, + 1198.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1015.0, + 1406.0, + 1015.0, + 1406.0, + 1050.0, + 1250.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1053.0, + 277.0, + 1053.0, + 277.0, + 1077.0, + 240.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1045.0, + 1406.0, + 1045.0, + 1406.0, + 1080.0, + 295.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1080.0, + 279.0, + 1080.0, + 279.0, + 1108.0, + 239.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1075.0, + 702.0, + 1075.0, + 702.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1075.0, + 1406.0, + 1075.0, + 1406.0, + 1111.0, + 732.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1112.0, + 277.0, + 1112.0, + 277.0, + 1136.0, + 240.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1104.0, + 1235.0, + 1104.0, + 1235.0, + 1140.0, + 291.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 209.0, + 278.0, + 209.0, + 278.0, + 236.0, + 241.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 200.0, + 1159.0, + 200.0, + 1159.0, + 240.0, + 293.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 200.0, + 1407.0, + 200.0, + 1407.0, + 240.0, + 1218.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 239.0, + 278.0, + 239.0, + 278.0, + 266.0, + 241.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 229.0, + 852.0, + 229.0, + 852.0, + 277.0, + 290.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 229.0, + 1167.0, + 229.0, + 1167.0, + 277.0, + 870.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 229.0, + 1407.0, + 229.0, + 1407.0, + 277.0, + 1248.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 273.0, + 277.0, + 273.0, + 277.0, + 300.0, + 241.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 264.0, + 820.0, + 264.0, + 820.0, + 305.0, + 291.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 264.0, + 1406.0, + 264.0, + 1406.0, + 305.0, + 848.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 302.0, + 278.0, + 302.0, + 278.0, + 329.0, + 241.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 300.0, + 327.0, + 300.0, + 327.0, + 331.0, + 295.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 300.0, + 887.0, + 300.0, + 887.0, + 331.0, + 345.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 1406.0, + 1233.0, + 1406.0, + 1276.0, + 294.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1265.0, + 1404.0, + 1265.0, + 1404.0, + 1303.0, + 293.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1297.0, + 1407.0, + 1297.0, + 1407.0, + 1333.0, + 294.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1325.0, + 1408.0, + 1325.0, + 1408.0, + 1363.0, + 292.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1357.0, + 1409.0, + 1357.0, + 1409.0, + 1394.0, + 292.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1388.0, + 1405.0, + 1388.0, + 1405.0, + 1424.0, + 294.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1446.0, + 1407.0, + 1446.0, + 1407.0, + 1486.0, + 293.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1479.0, + 1187.0, + 1479.0, + 1187.0, + 1514.0, + 293.0, + 1514.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 349, + 192, + 1373, + 192, + 1373, + 741, + 349, + 741 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 295, + 767, + 1405, + 767, + 1405, + 890, + 295, + 890 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 1778, + 1404, + 1778, + 1404, + 1871, + 298, + 1871 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 287, + 1884, + 1402, + 1884, + 1402, + 2008, + 287, + 2008 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 281, + 1599, + 1400, + 1599, + 1400, + 1693, + 281, + 1693 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 240, + 1376, + 1404, + 1376, + 1404, + 1590, + 240, + 1590 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 242, + 947, + 1404, + 947, + 1404, + 1071, + 242, + 1071 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 274, + 1212, + 1403, + 1212, + 1403, + 1366, + 274, + 1366 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 293, + 1724, + 533, + 1724, + 533, + 1761, + 293, + 1761 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 250, + 1080, + 1404, + 1080, + 1404, + 1203, + 250, + 1203 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.71 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2084, + 841, + 2084 + ], + "score": 0.107 + }, + { + "category_id": 13, + "poly": [ + 1069, + 1976, + 1168, + 1976, + 1168, + 2009, + 1069, + 2009 + ], + "score": 0.92, + "latex": "[ - 1 ; + 1 ]" + }, + { + "category_id": 13, + "poly": [ + 708, + 1437, + 789, + 1437, + 789, + 1465, + 708, + 1465 + ], + "score": 0.9, + "latex": "Z = S" + }, + { + "category_id": 13, + "poly": [ + 1148, + 1279, + 1177, + 1279, + 1177, + 1305, + 1148, + 1305 + ], + "score": 0.85, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 450, + 1919, + 469, + 1919, + 469, + 1946, + 450, + 1946 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 381, + 1919, + 400, + 1919, + 400, + 1942, + 381, + 1942 + ], + "score": 0.78, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 641, + 1529, + 663, + 1529, + 663, + 1556, + 641, + 1556 + ], + "score": 0.78, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 459, + 1529, + 483, + 1529, + 483, + 1556, + 459, + 1556 + ], + "score": 0.74, + "latex": "Z" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 197.0, + 539.0, + 197.0, + 539.0, + 228.0, + 442.0, + 228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 197.0, + 891.0, + 197.0, + 891.0, + 228.0, + 808.0, + 228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 197.0, + 1237.0, + 197.0, + 1237.0, + 227.0, + 1184.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 235.0, + 460.0, + 235.0, + 460.0, + 258.0, + 413.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 685.0, + 1178.0, + 685.0, + 1178.0, + 693.0, + 1154.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 693.0, + 471.0, + 693.0, + 471.0, + 705.0, + 450.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 690.0, + 1136.0, + 690.0, + 1136.0, + 705.0, + 1120.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 439.5, + 989.0, + 439.5, + 989.0, + 455.0, + 946.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.75, + 563.5, + 935.75, + 563.5, + 935.75, + 571.5, + 878.75, + 571.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 766.0, + 1407.0, + 766.0, + 1407.0, + 802.0, + 293.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 795.0, + 1405.0, + 795.0, + 1405.0, + 832.0, + 293.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 829.0, + 1405.0, + 829.0, + 1405.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 856.0, + 1062.0, + 856.0, + 1062.0, + 896.0, + 293.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1717.0, + 537.0, + 1717.0, + 537.0, + 1773.0, + 290.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2092.0, + 840.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1777.0, + 1409.0, + 1777.0, + 1409.0, + 1815.0, + 293.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1808.0, + 1406.0, + 1808.0, + 1406.0, + 1845.0, + 293.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1838.0, + 1347.0, + 1838.0, + 1347.0, + 1879.0, + 293.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1884.0, + 1404.0, + 1884.0, + 1404.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 380.0, + 1913.0, + 380.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1913.0, + 449.0, + 1913.0, + 449.0, + 1947.0, + 401.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1947.0, + 470.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1981.0, + 294.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1975.0, + 1068.0, + 1975.0, + 1068.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1975.0, + 1181.0, + 1975.0, + 1181.0, + 2011.0, + 1169.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1599.0, + 1404.0, + 1599.0, + 1404.0, + 1633.0, + 294.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1630.0, + 1404.0, + 1630.0, + 1404.0, + 1664.0, + 294.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1661.0, + 1403.0, + 1661.0, + 1403.0, + 1694.0, + 296.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1382.0, + 277.0, + 1382.0, + 277.0, + 1409.0, + 240.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1412.0, + 293.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1410.0, + 280.0, + 1410.0, + 280.0, + 1440.0, + 237.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1405.0, + 1407.0, + 1405.0, + 1407.0, + 1442.0, + 290.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1441.0, + 280.0, + 1441.0, + 280.0, + 1470.0, + 237.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1434.0, + 707.0, + 1434.0, + 707.0, + 1470.0, + 291.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1434.0, + 1409.0, + 1434.0, + 1409.0, + 1470.0, + 790.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1473.0, + 277.0, + 1473.0, + 277.0, + 1500.0, + 239.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1465.0, + 1408.0, + 1465.0, + 1408.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1501.0, + 280.0, + 1501.0, + 280.0, + 1531.0, + 237.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1532.0, + 294.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1531.0, + 280.0, + 1531.0, + 280.0, + 1562.0, + 237.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1526.0, + 458.0, + 1526.0, + 458.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1526.0, + 640.0, + 1526.0, + 640.0, + 1562.0, + 484.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1526.0, + 1405.0, + 1526.0, + 1405.0, + 1562.0, + 664.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1562.0, + 280.0, + 1562.0, + 280.0, + 1591.0, + 237.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1559.0, + 1055.0, + 1559.0, + 1055.0, + 1594.0, + 294.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 953.0, + 277.0, + 953.0, + 277.0, + 980.0, + 239.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 947.0, + 1404.0, + 947.0, + 1404.0, + 982.0, + 295.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 985.0, + 277.0, + 985.0, + 277.0, + 1012.0, + 239.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 978.0, + 1404.0, + 978.0, + 1404.0, + 1015.0, + 295.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1015.0, + 277.0, + 1015.0, + 277.0, + 1041.0, + 239.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1008.0, + 1405.0, + 1008.0, + 1405.0, + 1045.0, + 291.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1045.0, + 277.0, + 1045.0, + 277.0, + 1071.0, + 241.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1040.0, + 985.0, + 1040.0, + 985.0, + 1075.0, + 293.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1214.0, + 1404.0, + 1214.0, + 1404.0, + 1247.0, + 296.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1244.0, + 1404.0, + 1244.0, + 1404.0, + 1277.0, + 296.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1274.0, + 1147.0, + 1274.0, + 1147.0, + 1309.0, + 294.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1274.0, + 1407.0, + 1274.0, + 1407.0, + 1309.0, + 1178.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1302.0, + 1406.0, + 1302.0, + 1406.0, + 1342.0, + 292.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1335.0, + 928.0, + 1335.0, + 928.0, + 1368.0, + 296.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 244.0, + 1087.0, + 277.0, + 1087.0, + 277.0, + 1112.0, + 244.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1077.0, + 1404.0, + 1077.0, + 1404.0, + 1119.0, + 293.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 243.0, + 1116.0, + 277.0, + 1116.0, + 277.0, + 1143.0, + 243.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1111.0, + 1404.0, + 1111.0, + 1404.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 244.0, + 1147.0, + 277.0, + 1147.0, + 277.0, + 1174.0, + 244.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1141.0, + 1404.0, + 1141.0, + 1404.0, + 1176.0, + 295.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 244.0, + 1178.0, + 277.0, + 1178.0, + 277.0, + 1204.0, + 244.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1172.0, + 775.0, + 1172.0, + 775.0, + 1208.0, + 293.0, + 1208.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 239, + 626, + 1403, + 626, + 1403, + 870, + 239, + 870 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 300, + 197, + 1338, + 197, + 1338, + 527, + 300, + 527 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 292, + 1399, + 1403, + 1399, + 1403, + 1583, + 292, + 1583 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 294, + 1823, + 1402, + 1823, + 1402, + 2008, + 294, + 2008 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 243, + 1141, + 1404, + 1141, + 1404, + 1385, + 243, + 1385 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 245, + 883, + 1405, + 883, + 1405, + 1128, + 245, + 1128 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 283, + 1596, + 1404, + 1596, + 1404, + 1810, + 283, + 1810 + ], + "score": 0.96 + }, + { + "category_id": 4, + "poly": [ + 522, + 541, + 1173, + 541, + 1173, + 573, + 522, + 573 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 858, + 2062, + 858, + 2085, + 840, + 2085 + ], + "score": 0.795 + }, + { + "category_id": 13, + "poly": [ + 1259, + 1521, + 1353, + 1521, + 1353, + 1549, + 1259, + 1549 + ], + "score": 0.91, + "latex": "T = 1 0" + }, + { + "category_id": 13, + "poly": [ + 1160, + 1597, + 1250, + 1597, + 1250, + 1625, + 1160, + 1625 + ], + "score": 0.89, + "latex": "T = 2 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 1552, + 390, + 1552, + 390, + 1581, + 297, + 1581 + ], + "score": 0.88, + "latex": "H = 1 0" + }, + { + "category_id": 13, + "poly": [ + 1300, + 1597, + 1394, + 1597, + 1394, + 1625, + 1300, + 1625 + ], + "score": 0.87, + "latex": "H = 2 0" + }, + { + "category_id": 13, + "poly": [ + 896, + 627, + 925, + 627, + 925, + 654, + 896, + 654 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 197.0, + 569.0, + 197.0, + 569.0, + 224.0, + 448.0, + 224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 195.0, + 930.0, + 195.0, + 930.0, + 224.0, + 864.0, + 224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 233.0, + 353.0, + 233.0, + 353.0, + 248.0, + 338.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 233.0, + 742.0, + 233.0, + 742.0, + 248.0, + 728.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 272.0, + 364.0, + 272.0, + 364.0, + 298.0, + 321.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 272.0, + 747.0, + 272.0, + 747.0, + 298.0, + 710.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 269.0, + 1244.0, + 269.0, + 1244.0, + 304.0, + 1149.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 293.0, + 1241.0, + 293.0, + 1241.0, + 329.0, + 1147.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 312.0, + 363.0, + 312.0, + 363.0, + 390.0, + 302.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 313.0, + 746.0, + 313.0, + 746.0, + 391.0, + 689.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 319.0, + 1242.0, + 319.0, + 1242.0, + 351.0, + 1147.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 344.0, + 1260.0, + 344.0, + 1260.0, + 375.0, + 1147.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 362.0, + 359.0, + 362.0, + 359.0, + 389.0, + 321.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 363.0, + 744.0, + 363.0, + 744.0, + 386.0, + 711.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 369.0, + 1259.0, + 369.0, + 1259.0, + 397.0, + 1150.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 405.0, + 364.0, + 405.0, + 364.0, + 435.0, + 320.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 407.0, + 748.0, + 407.0, + 748.0, + 434.0, + 710.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 390.0, + 1331.0, + 390.0, + 1331.0, + 450.0, + 1148.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 452.0, + 366.0, + 452.0, + 366.0, + 479.0, + 333.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 455.0, + 747.0, + 455.0, + 747.0, + 478.0, + 724.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 454.0, + 779.0, + 454.0, + 779.0, + 467.0, + 770.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 473.0, + 397.0, + 473.0, + 397.0, + 505.0, + 374.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 480.0, + 444.0, + 480.0, + 444.0, + 504.0, + 411.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 480.0, + 487.0, + 480.0, + 487.0, + 504.0, + 454.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 480.0, + 530.0, + 480.0, + 530.0, + 504.0, + 497.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 479.0, + 572.0, + 479.0, + 572.0, + 506.0, + 537.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 480.0, + 606.0, + 480.0, + 606.0, + 504.0, + 588.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 478.0, + 658.0, + 478.0, + 658.0, + 506.0, + 622.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 482.0, + 783.0, + 482.0, + 783.0, + 504.0, + 764.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 479.0, + 836.0, + 479.0, + 836.0, + 506.0, + 800.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 480.0, + 876.0, + 480.0, + 876.0, + 504.0, + 843.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 479.0, + 920.0, + 479.0, + 920.0, + 506.0, + 884.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 479.0, + 961.0, + 479.0, + 961.0, + 506.0, + 926.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 482.0, + 994.0, + 482.0, + 994.0, + 502.0, + 977.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 479.0, + 1047.0, + 479.0, + 1047.0, + 506.0, + 1012.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 497.0, + 572.0, + 497.0, + 572.0, + 528.0, + 447.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 501.0, + 649.0, + 501.0, + 649.0, + 533.0, + 607.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 500.0, + 959.0, + 500.0, + 959.0, + 527.0, + 837.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 501.0, + 1038.0, + 501.0, + 1038.0, + 533.0, + 996.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 539.0, + 1175.0, + 539.0, + 1175.0, + 576.0, + 523.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2090.0, + 838.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 632.0, + 277.0, + 632.0, + 277.0, + 658.0, + 239.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 628.0, + 895.0, + 628.0, + 895.0, + 658.0, + 296.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 628.0, + 1403.0, + 628.0, + 1403.0, + 658.0, + 926.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 661.0, + 279.0, + 661.0, + 279.0, + 690.0, + 238.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 655.0, + 1404.0, + 655.0, + 1404.0, + 692.0, + 294.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 694.0, + 277.0, + 694.0, + 277.0, + 718.0, + 239.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 1406.0, + 686.0, + 1406.0, + 724.0, + 293.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 724.0, + 277.0, + 724.0, + 277.0, + 749.0, + 239.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 714.0, + 1406.0, + 714.0, + 1406.0, + 753.0, + 292.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 749.0, + 279.0, + 749.0, + 279.0, + 782.0, + 238.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 745.0, + 1406.0, + 745.0, + 1406.0, + 784.0, + 290.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 786.0, + 276.0, + 786.0, + 276.0, + 810.0, + 239.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 775.0, + 1406.0, + 775.0, + 1406.0, + 817.0, + 292.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 815.0, + 277.0, + 815.0, + 277.0, + 841.0, + 239.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 808.0, + 1404.0, + 808.0, + 1404.0, + 842.0, + 294.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 845.0, + 277.0, + 845.0, + 277.0, + 869.0, + 239.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 839.0, + 1259.0, + 839.0, + 1259.0, + 873.0, + 294.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1396.0, + 1408.0, + 1396.0, + 1408.0, + 1436.0, + 292.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1428.0, + 1406.0, + 1428.0, + 1406.0, + 1466.0, + 292.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1459.0, + 1406.0, + 1459.0, + 1406.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1488.0, + 1405.0, + 1488.0, + 1405.0, + 1525.0, + 294.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1520.0, + 1258.0, + 1520.0, + 1258.0, + 1554.0, + 294.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1520.0, + 1406.0, + 1520.0, + 1406.0, + 1554.0, + 1354.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1551.0, + 1356.0, + 1551.0, + 1356.0, + 1586.0, + 391.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1824.0, + 1404.0, + 1824.0, + 1404.0, + 1856.0, + 296.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1853.0, + 1405.0, + 1853.0, + 1405.0, + 1889.0, + 293.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1886.0, + 1405.0, + 1886.0, + 1405.0, + 1918.0, + 296.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1915.0, + 1404.0, + 1915.0, + 1404.0, + 1947.0, + 296.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1947.0, + 1403.0, + 1947.0, + 1403.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 1405.0, + 1975.0, + 1405.0, + 2011.0, + 294.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1147.0, + 277.0, + 1147.0, + 277.0, + 1173.0, + 239.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1141.0, + 1408.0, + 1141.0, + 1408.0, + 1176.0, + 294.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1178.0, + 277.0, + 1178.0, + 277.0, + 1204.0, + 239.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1173.0, + 1405.0, + 1173.0, + 1405.0, + 1207.0, + 295.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1208.0, + 277.0, + 1208.0, + 277.0, + 1233.0, + 239.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1203.0, + 1405.0, + 1203.0, + 1405.0, + 1236.0, + 295.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1236.0, + 278.0, + 1236.0, + 278.0, + 1266.0, + 238.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1231.0, + 1404.0, + 1231.0, + 1404.0, + 1267.0, + 293.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1270.0, + 277.0, + 1270.0, + 277.0, + 1295.0, + 239.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1262.0, + 1407.0, + 1262.0, + 1407.0, + 1299.0, + 293.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1299.0, + 276.0, + 1299.0, + 276.0, + 1325.0, + 239.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1294.0, + 1405.0, + 1294.0, + 1405.0, + 1327.0, + 294.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1330.0, + 277.0, + 1330.0, + 277.0, + 1356.0, + 239.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1321.0, + 1405.0, + 1321.0, + 1405.0, + 1361.0, + 292.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1360.0, + 277.0, + 1360.0, + 277.0, + 1385.0, + 240.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1354.0, + 788.0, + 1354.0, + 788.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 891.0, + 278.0, + 891.0, + 278.0, + 917.0, + 240.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 881.0, + 1405.0, + 881.0, + 1405.0, + 921.0, + 292.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 921.0, + 278.0, + 921.0, + 278.0, + 946.0, + 240.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 1404.0, + 917.0, + 1404.0, + 946.0, + 295.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 949.0, + 279.0, + 949.0, + 279.0, + 979.0, + 238.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 942.0, + 1406.0, + 942.0, + 1406.0, + 981.0, + 292.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 980.0, + 278.0, + 980.0, + 278.0, + 1010.0, + 237.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 973.0, + 1406.0, + 973.0, + 1406.0, + 1011.0, + 293.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1010.0, + 279.0, + 1010.0, + 279.0, + 1039.0, + 238.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1005.0, + 1406.0, + 1005.0, + 1406.0, + 1039.0, + 295.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1042.0, + 276.0, + 1042.0, + 276.0, + 1067.0, + 240.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1035.0, + 1405.0, + 1035.0, + 1405.0, + 1069.0, + 293.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1070.0, + 278.0, + 1070.0, + 278.0, + 1100.0, + 238.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1068.0, + 1405.0, + 1068.0, + 1405.0, + 1098.0, + 293.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1103.0, + 276.0, + 1103.0, + 276.0, + 1128.0, + 240.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1097.0, + 1249.0, + 1097.0, + 1249.0, + 1131.0, + 295.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1594.0, + 1159.0, + 1594.0, + 1159.0, + 1631.0, + 294.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1594.0, + 1299.0, + 1594.0, + 1299.0, + 1631.0, + 1251.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1594.0, + 1407.0, + 1594.0, + 1407.0, + 1631.0, + 1395.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1627.0, + 1406.0, + 1627.0, + 1406.0, + 1661.0, + 294.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1658.0, + 1406.0, + 1658.0, + 1406.0, + 1692.0, + 295.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1690.0, + 1404.0, + 1690.0, + 1404.0, + 1720.0, + 297.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1718.0, + 1404.0, + 1718.0, + 1404.0, + 1752.0, + 295.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1745.0, + 1406.0, + 1745.0, + 1406.0, + 1786.0, + 293.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1779.0, + 887.0, + 1779.0, + 887.0, + 1813.0, + 294.0, + 1813.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1654, + 1404, + 1654, + 1404, + 1929, + 295, + 1929 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 239, + 587, + 1405, + 587, + 1405, + 834, + 239, + 834 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 298, + 197, + 1393, + 197, + 1393, + 482, + 298, + 482 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 306, + 860, + 1376, + 860, + 1376, + 1467, + 306, + 1467 + ], + "score": 0.945 + }, + { + "category_id": 4, + "poly": [ + 295, + 1475, + 1073, + 1475, + 1073, + 1568, + 295, + 1568 + ], + "score": 0.94 + }, + { + "category_id": 4, + "poly": [ + 293, + 489, + 1408, + 489, + 1408, + 553, + 293, + 553 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 275, + 1606, + 509, + 1606, + 509, + 1642, + 275, + 1642 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 1949, + 1401, + 1949, + 1401, + 2007, + 300, + 2007 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2084, + 840, + 2084 + ], + "score": 0.803 + }, + { + "category_id": 2, + "poly": [ + 238, + 1635, + 277, + 1635, + 277, + 1928, + 238, + 1928 + ], + "score": 0.144 + }, + { + "category_id": 13, + "poly": [ + 355, + 624, + 383, + 624, + 383, + 650, + 355, + 650 + ], + "score": 0.83, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 637, + 233, + 647, + 233, + 647, + 243, + 637, + 243 + ], + "score": 0.32, + "latex": "^ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 332, + 234, + 342, + 234, + 342, + 244, + 332, + 244 + ], + "score": 0.25, + "latex": "1" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 196.0, + 509.0, + 196.0, + 509.0, + 220.0, + 447.0, + 220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 197.0, + 811.0, + 197.0, + 811.0, + 222.0, + 754.0, + 222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 198.0, + 1116.0, + 198.0, + 1116.0, + 222.0, + 1059.0, + 222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 230.0, + 346.0, + 230.0, + 346.0, + 246.0, + 343.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 228.0, + 651.0, + 228.0, + 651.0, + 244.0, + 648.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 225.0, + 961.0, + 225.0, + 961.0, + 243.0, + 940.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 264.0, + 351.0, + 264.0, + 351.0, + 287.0, + 318.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 261.0, + 661.0, + 261.0, + 661.0, + 287.0, + 621.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 263.0, + 959.0, + 263.0, + 959.0, + 401.0, + 908.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 285.0, + 1325.0, + 285.0, + 1325.0, + 306.0, + 1258.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 299.0, + 350.0, + 299.0, + 350.0, + 365.0, + 300.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 297.0, + 660.0, + 297.0, + 660.0, + 366.0, + 603.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 300.0, + 960.0, + 300.0, + 960.0, + 323.0, + 928.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 301.0, + 1336.0, + 301.0, + 1336.0, + 324.0, + 1254.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 318.0, + 1337.0, + 318.0, + 1337.0, + 342.0, + 1255.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 339.0, + 351.0, + 339.0, + 351.0, + 366.0, + 318.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 338.0, + 661.0, + 338.0, + 661.0, + 364.0, + 621.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 339.0, + 959.0, + 339.0, + 959.0, + 361.0, + 928.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 331.0, + 1387.0, + 331.0, + 1387.0, + 361.0, + 1254.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 352.0, + 1322.0, + 352.0, + 1322.0, + 376.0, + 1254.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 379.0, + 353.0, + 379.0, + 353.0, + 402.0, + 319.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 378.0, + 659.0, + 378.0, + 659.0, + 400.0, + 623.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 380.0, + 958.0, + 380.0, + 958.0, + 398.0, + 929.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 368.0, + 1359.0, + 368.0, + 1359.0, + 395.0, + 1254.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 420.0, + 355.0, + 420.0, + 355.0, + 439.0, + 329.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 419.0, + 651.0, + 419.0, + 651.0, + 435.0, + 635.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 419.0, + 959.0, + 419.0, + 959.0, + 435.0, + 941.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 445.0, + 374.0, + 445.0, + 374.0, + 461.0, + 359.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 444.0, + 427.0, + 444.0, + 427.0, + 462.0, + 412.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 447.0, + 481.0, + 447.0, + 481.0, + 462.0, + 466.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 443.0, + 537.0, + 443.0, + 537.0, + 463.0, + 518.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 443.0, + 590.0, + 443.0, + 590.0, + 463.0, + 572.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 443.0, + 680.0, + 443.0, + 680.0, + 460.0, + 664.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 442.0, + 736.0, + 442.0, + 736.0, + 459.0, + 722.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 443.0, + 793.0, + 443.0, + 793.0, + 460.0, + 780.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 443.0, + 850.0, + 443.0, + 850.0, + 460.0, + 836.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 443.0, + 908.0, + 443.0, + 908.0, + 460.0, + 893.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 443.0, + 979.0, + 443.0, + 979.0, + 459.0, + 964.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 441.0, + 1060.0, + 441.0, + 1060.0, + 462.0, + 1030.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 447.0, + 1122.0, + 447.0, + 1122.0, + 457.0, + 1112.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 437.0, + 1205.0, + 437.0, + 1205.0, + 465.0, + 1173.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 459.0, + 532.0, + 459.0, + 532.0, + 481.0, + 426.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 462.0, + 597.0, + 462.0, + 597.0, + 485.0, + 563.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 456.0, + 836.0, + 456.0, + 836.0, + 483.0, + 729.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 457.0, + 1141.0, + 457.0, + 1141.0, + 480.0, + 1035.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 459.0, + 1206.0, + 459.0, + 1206.0, + 485.0, + 1172.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 461.0, + 904.0, + 461.0, + 904.0, + 483.0, + 864.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 865.0, + 592.0, + 865.0, + 592.0, + 885.0, + 401.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 865.0, + 965.0, + 865.0, + 965.0, + 885.0, + 781.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 897.0, + 352.0, + 897.0, + 352.0, + 909.0, + 344.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 896.0, + 733.0, + 896.0, + 733.0, + 911.0, + 718.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 933.0, + 360.0, + 933.0, + 360.0, + 956.0, + 328.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 933.0, + 735.0, + 933.0, + 735.0, + 955.0, + 704.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 959.0, + 360.0, + 959.0, + 360.0, + 1050.0, + 310.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 961.0, + 736.0, + 961.0, + 736.0, + 1048.0, + 685.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1014.0, + 358.0, + 1014.0, + 358.0, + 1036.0, + 327.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1014.0, + 736.0, + 1014.0, + 736.0, + 1036.0, + 704.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1057.0, + 360.0, + 1057.0, + 360.0, + 1075.0, + 329.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1054.0, + 740.0, + 1054.0, + 740.0, + 1078.0, + 704.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1098.0, + 357.0, + 1098.0, + 357.0, + 1116.0, + 341.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1100.0, + 733.0, + 1100.0, + 733.0, + 1116.0, + 718.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1122.0, + 386.0, + 1122.0, + 386.0, + 1143.0, + 368.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1121.0, + 450.0, + 1121.0, + 450.0, + 1143.0, + 432.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1122.0, + 513.0, + 1122.0, + 513.0, + 1143.0, + 495.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1123.0, + 574.0, + 1123.0, + 574.0, + 1140.0, + 558.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1122.0, + 639.0, + 1122.0, + 639.0, + 1143.0, + 620.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1122.0, + 769.0, + 1122.0, + 769.0, + 1143.0, + 751.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1124.0, + 813.0, + 1124.0, + 813.0, + 1140.0, + 798.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1122.0, + 862.0, + 1122.0, + 862.0, + 1143.0, + 843.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1122.0, + 908.0, + 1122.0, + 908.0, + 1143.0, + 889.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1122.0, + 954.0, + 1122.0, + 954.0, + 1141.0, + 935.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1122.0, + 999.0, + 1122.0, + 999.0, + 1143.0, + 980.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1138.0, + 554.0, + 1138.0, + 554.0, + 1164.0, + 440.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1143.0, + 623.0, + 1143.0, + 623.0, + 1166.0, + 587.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1137.0, + 930.0, + 1137.0, + 930.0, + 1165.0, + 817.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1143.0, + 999.0, + 1143.0, + 999.0, + 1166.0, + 964.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1164.0, + 622.0, + 1164.0, + 622.0, + 1183.0, + 377.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1163.0, + 987.0, + 1163.0, + 987.0, + 1187.0, + 746.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1195.0, + 357.0, + 1195.0, + 357.0, + 1211.0, + 344.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1197.0, + 722.0, + 1197.0, + 722.0, + 1209.0, + 712.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1233.0, + 361.0, + 1233.0, + 361.0, + 1256.0, + 330.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1233.0, + 733.0, + 1233.0, + 733.0, + 1256.0, + 697.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1259.0, + 361.0, + 1259.0, + 361.0, + 1348.0, + 312.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1260.0, + 728.0, + 1260.0, + 728.0, + 1349.0, + 679.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1314.0, + 361.0, + 1314.0, + 361.0, + 1337.0, + 330.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1314.0, + 728.0, + 1314.0, + 728.0, + 1338.0, + 697.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1355.0, + 362.0, + 1355.0, + 362.0, + 1377.0, + 330.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1355.0, + 728.0, + 1355.0, + 728.0, + 1377.0, + 697.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1399.0, + 358.0, + 1399.0, + 358.0, + 1415.0, + 344.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1399.0, + 725.0, + 1399.0, + 725.0, + 1416.0, + 711.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1425.0, + 386.0, + 1425.0, + 386.0, + 1441.0, + 373.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1424.0, + 447.0, + 1424.0, + 447.0, + 1442.0, + 433.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 1425.0, + 508.0, + 1425.0, + 508.0, + 1441.0, + 494.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1425.0, + 569.0, + 1425.0, + 569.0, + 1442.0, + 555.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1425.0, + 630.0, + 1425.0, + 630.0, + 1442.0, + 616.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1425.0, + 759.0, + 1425.0, + 759.0, + 1442.0, + 745.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1426.0, + 798.0, + 1426.0, + 798.0, + 1440.0, + 789.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1421.0, + 845.0, + 1421.0, + 845.0, + 1443.0, + 826.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1425.0, + 884.0, + 1425.0, + 884.0, + 1442.0, + 869.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1425.0, + 925.0, + 1425.0, + 925.0, + 1442.0, + 910.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1421.0, + 969.0, + 1421.0, + 969.0, + 1443.0, + 951.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1425.0, + 1009.0, + 1425.0, + 1009.0, + 1442.0, + 993.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1438.0, + 555.0, + 1438.0, + 555.0, + 1462.0, + 445.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1438.0, + 922.0, + 1438.0, + 922.0, + 1462.0, + 812.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1444.0, + 627.0, + 1444.0, + 627.0, + 1466.0, + 588.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1444.0, + 993.0, + 1444.0, + 993.0, + 1466.0, + 953.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1473.0, + 1074.0, + 1473.0, + 1074.0, + 1508.0, + 404.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1502.0, + 1076.0, + 1502.0, + 1076.0, + 1540.0, + 293.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1535.0, + 1017.0, + 1535.0, + 1017.0, + 1571.0, + 404.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 486.0, + 1407.0, + 486.0, + 1407.0, + 526.0, + 294.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 518.0, + 837.0, + 518.0, + 837.0, + 556.0, + 294.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1602.0, + 515.0, + 1602.0, + 515.0, + 1650.0, + 292.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1983.0, + 329.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1977.0, + 583.0, + 1977.0, + 583.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 861.0, + 2060.0, + 861.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1661.0, + 278.0, + 1661.0, + 278.0, + 1688.0, + 238.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1688.0, + 279.0, + 1688.0, + 279.0, + 1719.0, + 238.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1719.0, + 279.0, + 1719.0, + 279.0, + 1750.0, + 237.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1750.0, + 278.0, + 1750.0, + 278.0, + 1781.0, + 237.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1780.0, + 279.0, + 1780.0, + 279.0, + 1812.0, + 237.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1810.0, + 279.0, + 1810.0, + 279.0, + 1842.0, + 237.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1843.0, + 278.0, + 1843.0, + 278.0, + 1871.0, + 238.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1872.0, + 279.0, + 1872.0, + 279.0, + 1902.0, + 237.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1901.0, + 279.0, + 1901.0, + 279.0, + 1932.0, + 237.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1653.0, + 1405.0, + 1653.0, + 1405.0, + 1689.0, + 294.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1682.0, + 1404.0, + 1682.0, + 1404.0, + 1720.0, + 293.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1716.0, + 1405.0, + 1716.0, + 1405.0, + 1749.0, + 295.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1781.0, + 293.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1772.0, + 1407.0, + 1772.0, + 1407.0, + 1814.0, + 290.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1806.0, + 1404.0, + 1806.0, + 1404.0, + 1842.0, + 294.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1836.0, + 1405.0, + 1836.0, + 1405.0, + 1871.0, + 293.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1866.0, + 1407.0, + 1866.0, + 1407.0, + 1902.0, + 292.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1897.0, + 1205.0, + 1897.0, + 1205.0, + 1933.0, + 294.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 594.0, + 277.0, + 594.0, + 277.0, + 620.0, + 239.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 586.0, + 1406.0, + 586.0, + 1406.0, + 622.0, + 292.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 625.0, + 277.0, + 625.0, + 277.0, + 651.0, + 239.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 620.0, + 354.0, + 620.0, + 354.0, + 654.0, + 293.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 620.0, + 1405.0, + 620.0, + 1405.0, + 654.0, + 384.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 654.0, + 279.0, + 654.0, + 279.0, + 682.0, + 236.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 647.0, + 1406.0, + 647.0, + 1406.0, + 685.0, + 293.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 683.0, + 279.0, + 683.0, + 279.0, + 713.0, + 238.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 678.0, + 1406.0, + 678.0, + 1406.0, + 716.0, + 293.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 716.0, + 277.0, + 716.0, + 277.0, + 742.0, + 239.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 711.0, + 1405.0, + 711.0, + 1405.0, + 744.0, + 296.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 747.0, + 277.0, + 747.0, + 277.0, + 771.0, + 239.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 774.0, + 296.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 774.0, + 279.0, + 774.0, + 279.0, + 804.0, + 238.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 769.0, + 1406.0, + 769.0, + 1406.0, + 807.0, + 293.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 808.0, + 277.0, + 808.0, + 277.0, + 834.0, + 239.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 803.0, + 1005.0, + 803.0, + 1005.0, + 837.0, + 295.0, + 837.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 261, + 191, + 1411, + 191, + 1411, + 2025, + 261, + 2025 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 835, + 2060, + 867, + 2060, + 867, + 2087, + 835, + 2087 + ], + "score": 0.836 + }, + { + "category_id": 0, + "poly": [ + 268, + 199, + 456, + 199, + 456, + 235, + 268, + 235 + ], + "score": 0.736 + }, + { + "category_id": 2, + "poly": [ + 238, + 213, + 287, + 213, + 287, + 2040, + 238, + 2040 + ], + "score": 0.189 + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2058.0, + 871.0, + 2058.0, + 871.0, + 2095.0, + 831.0, + 2095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 273.0, + 196.0, + 459.0, + 196.0, + 459.0, + 240.0, + 273.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 232.0, + 205.0, + 283.0, + 205.0, + 283.0, + 243.0, + 232.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 255.0, + 281.0, + 255.0, + 281.0, + 287.0, + 235.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 232.0, + 285.0, + 283.0, + 285.0, + 283.0, + 324.0, + 232.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 340.0, + 281.0, + 340.0, + 281.0, + 372.0, + 237.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 370.0, + 281.0, + 370.0, + 281.0, + 414.0, + 235.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 424.0, + 281.0, + 424.0, + 281.0, + 458.0, + 237.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 452.0, + 283.0, + 452.0, + 283.0, + 494.0, + 235.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 508.0, + 281.0, + 508.0, + 281.0, + 540.0, + 237.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 536.0, + 283.0, + 536.0, + 283.0, + 581.0, + 235.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 593.0, + 281.0, + 593.0, + 281.0, + 625.0, + 237.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 621.0, + 283.0, + 621.0, + 283.0, + 665.0, + 235.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 677.0, + 281.0, + 677.0, + 281.0, + 711.0, + 237.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 705.0, + 281.0, + 705.0, + 281.0, + 739.0, + 237.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 735.0, + 283.0, + 735.0, + 283.0, + 775.0, + 235.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 789.0, + 279.0, + 789.0, + 279.0, + 823.0, + 237.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 819.0, + 283.0, + 819.0, + 283.0, + 864.0, + 235.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 872.0, + 283.0, + 872.0, + 283.0, + 912.0, + 235.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 904.0, + 281.0, + 904.0, + 281.0, + 938.0, + 237.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 934.0, + 283.0, + 934.0, + 283.0, + 976.0, + 235.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 988.0, + 281.0, + 988.0, + 281.0, + 1022.0, + 237.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1016.0, + 283.0, + 1016.0, + 283.0, + 1056.0, + 235.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1068.0, + 283.0, + 1068.0, + 283.0, + 1108.0, + 235.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1100.0, + 283.0, + 1100.0, + 283.0, + 1145.0, + 235.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1155.0, + 281.0, + 1155.0, + 281.0, + 1189.0, + 237.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1187.0, + 281.0, + 1187.0, + 281.0, + 1225.0, + 237.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1239.0, + 281.0, + 1239.0, + 281.0, + 1273.0, + 237.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1271.0, + 281.0, + 1271.0, + 281.0, + 1303.0, + 237.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1299.0, + 283.0, + 1299.0, + 283.0, + 1345.0, + 235.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1355.0, + 281.0, + 1355.0, + 281.0, + 1387.0, + 235.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1383.0, + 283.0, + 1383.0, + 283.0, + 1426.0, + 235.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1438.0, + 281.0, + 1438.0, + 281.0, + 1472.0, + 237.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 232.0, + 1466.0, + 283.0, + 1466.0, + 283.0, + 1504.0, + 232.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1502.0, + 281.0, + 1502.0, + 281.0, + 1536.0, + 237.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1552.0, + 281.0, + 1552.0, + 281.0, + 1586.0, + 237.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1582.0, + 281.0, + 1582.0, + 281.0, + 1626.0, + 235.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1636.0, + 281.0, + 1636.0, + 281.0, + 1668.0, + 235.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1668.0, + 281.0, + 1668.0, + 281.0, + 1701.0, + 235.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1699.0, + 281.0, + 1699.0, + 281.0, + 1737.0, + 237.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 232.0, + 1747.0, + 283.0, + 1747.0, + 283.0, + 1787.0, + 232.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 232.0, + 1779.0, + 283.0, + 1779.0, + 283.0, + 1817.0, + 232.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1811.0, + 283.0, + 1811.0, + 283.0, + 1855.0, + 235.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1865.0, + 281.0, + 1865.0, + 281.0, + 1899.0, + 237.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1893.0, + 283.0, + 1893.0, + 283.0, + 1933.0, + 235.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1949.0, + 281.0, + 1949.0, + 281.0, + 1982.0, + 235.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1980.0, + 279.0, + 1980.0, + 279.0, + 2014.0, + 237.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 212.0, + 281.0, + 212.0, + 281.0, + 228.0, + 267.0, + 228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 197.0, + 462.0, + 197.0, + 462.0, + 240.0, + 288.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 258.0, + 281.0, + 258.0, + 281.0, + 284.0, + 258.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 252.0, + 1410.0, + 252.0, + 1410.0, + 290.0, + 306.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 288.0, + 279.0, + 288.0, + 279.0, + 316.0, + 256.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 282.0, + 811.0, + 282.0, + 811.0, + 320.0, + 353.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 342.0, + 281.0, + 342.0, + 281.0, + 369.0, + 254.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 336.0, + 1410.0, + 336.0, + 1410.0, + 375.0, + 306.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 377.0, + 275.0, + 377.0, + 275.0, + 397.0, + 256.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 367.0, + 1175.0, + 367.0, + 1175.0, + 405.0, + 353.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 427.0, + 281.0, + 427.0, + 281.0, + 453.0, + 254.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 417.0, + 1408.0, + 417.0, + 1408.0, + 461.0, + 304.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 457.0, + 279.0, + 457.0, + 279.0, + 485.0, + 254.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 453.0, + 1037.0, + 453.0, + 1037.0, + 485.0, + 355.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 508.0, + 281.0, + 508.0, + 281.0, + 536.0, + 254.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 504.0, + 1405.0, + 504.0, + 1405.0, + 542.0, + 306.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 540.0, + 279.0, + 540.0, + 279.0, + 568.0, + 254.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 534.0, + 809.0, + 534.0, + 809.0, + 574.0, + 349.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 594.0, + 281.0, + 594.0, + 281.0, + 620.0, + 256.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 588.0, + 1410.0, + 588.0, + 1410.0, + 627.0, + 306.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 627.0, + 279.0, + 627.0, + 279.0, + 651.0, + 254.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 619.0, + 809.0, + 619.0, + 809.0, + 655.0, + 351.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 679.0, + 281.0, + 679.0, + 281.0, + 705.0, + 256.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 671.0, + 1405.0, + 671.0, + 1405.0, + 709.0, + 306.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 709.0, + 279.0, + 709.0, + 279.0, + 735.0, + 254.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 701.0, + 1405.0, + 701.0, + 1405.0, + 739.0, + 353.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 739.0, + 279.0, + 739.0, + 279.0, + 766.0, + 254.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 731.0, + 431.0, + 731.0, + 431.0, + 772.0, + 353.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 792.0, + 279.0, + 792.0, + 279.0, + 820.0, + 254.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 786.0, + 1408.0, + 786.0, + 1408.0, + 824.0, + 306.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 824.0, + 279.0, + 824.0, + 279.0, + 850.0, + 254.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 818.0, + 578.0, + 818.0, + 578.0, + 852.0, + 353.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 876.0, + 281.0, + 876.0, + 281.0, + 905.0, + 254.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 868.0, + 1410.0, + 868.0, + 1410.0, + 913.0, + 304.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 907.0, + 279.0, + 907.0, + 279.0, + 933.0, + 254.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 898.0, + 1408.0, + 898.0, + 1408.0, + 939.0, + 351.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 937.0, + 279.0, + 937.0, + 279.0, + 965.0, + 254.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 931.0, + 542.0, + 931.0, + 542.0, + 969.0, + 353.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 989.0, + 281.0, + 989.0, + 281.0, + 1017.0, + 254.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 983.0, + 1405.0, + 983.0, + 1405.0, + 1021.0, + 306.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1021.0, + 279.0, + 1021.0, + 279.0, + 1050.0, + 254.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1015.0, + 1381.0, + 1015.0, + 1381.0, + 1054.0, + 353.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1074.0, + 283.0, + 1074.0, + 283.0, + 1100.0, + 256.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1068.0, + 1408.0, + 1068.0, + 1408.0, + 1106.0, + 289.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1104.0, + 281.0, + 1104.0, + 281.0, + 1132.0, + 254.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1098.0, + 916.0, + 1098.0, + 916.0, + 1136.0, + 351.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1158.0, + 283.0, + 1158.0, + 283.0, + 1185.0, + 256.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1154.0, + 1405.0, + 1154.0, + 1405.0, + 1187.0, + 291.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1193.0, + 275.0, + 1193.0, + 275.0, + 1213.0, + 256.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1183.0, + 645.0, + 1183.0, + 645.0, + 1221.0, + 353.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 1243.0, + 283.0, + 1243.0, + 283.0, + 1267.0, + 258.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1239.0, + 1408.0, + 1239.0, + 1408.0, + 1271.0, + 289.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1273.0, + 279.0, + 1273.0, + 279.0, + 1299.0, + 254.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1269.0, + 1405.0, + 1269.0, + 1405.0, + 1301.0, + 355.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 1307.0, + 277.0, + 1307.0, + 277.0, + 1328.0, + 258.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1297.0, + 986.0, + 1297.0, + 986.0, + 1336.0, + 353.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 1346.0, + 1406.0, + 1346.0, + 1406.0, + 1392.0, + 258.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1388.0, + 279.0, + 1388.0, + 279.0, + 1414.0, + 256.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1380.0, + 1278.0, + 1380.0, + 1278.0, + 1418.0, + 353.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1436.0, + 1405.0, + 1436.0, + 1405.0, + 1469.0, + 254.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1471.0, + 279.0, + 1471.0, + 279.0, + 1497.0, + 254.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1458.0, + 1408.0, + 1458.0, + 1408.0, + 1507.0, + 349.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 1505.0, + 277.0, + 1505.0, + 277.0, + 1525.0, + 258.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1497.0, + 614.0, + 1497.0, + 614.0, + 1529.0, + 353.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1539.0, + 1408.0, + 1539.0, + 1408.0, + 1594.0, + 256.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1585.0, + 279.0, + 1585.0, + 279.0, + 1612.0, + 254.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1579.0, + 1346.0, + 1579.0, + 1346.0, + 1618.0, + 355.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1638.0, + 283.0, + 1638.0, + 283.0, + 1664.0, + 256.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1632.0, + 1405.0, + 1632.0, + 1405.0, + 1670.0, + 291.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1670.0, + 279.0, + 1670.0, + 279.0, + 1698.0, + 254.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1664.0, + 1408.0, + 1664.0, + 1408.0, + 1702.0, + 353.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1700.0, + 279.0, + 1700.0, + 279.0, + 1726.0, + 254.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1694.0, + 1087.0, + 1694.0, + 1087.0, + 1733.0, + 353.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1755.0, + 283.0, + 1755.0, + 283.0, + 1779.0, + 256.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1747.0, + 1405.0, + 1747.0, + 1405.0, + 1785.0, + 289.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1783.0, + 279.0, + 1783.0, + 279.0, + 1811.0, + 254.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1777.0, + 1410.0, + 1777.0, + 1410.0, + 1813.0, + 351.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1813.0, + 279.0, + 1813.0, + 279.0, + 1839.0, + 254.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1803.0, + 516.0, + 1803.0, + 516.0, + 1846.0, + 352.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1867.0, + 283.0, + 1867.0, + 283.0, + 1894.0, + 254.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1857.0, + 1408.0, + 1857.0, + 1408.0, + 1902.0, + 289.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1900.0, + 279.0, + 1900.0, + 279.0, + 1926.0, + 254.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1892.0, + 1395.0, + 1892.0, + 1395.0, + 1930.0, + 353.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 1952.0, + 285.0, + 1952.0, + 285.0, + 1978.0, + 256.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1944.0, + 1408.0, + 1944.0, + 1408.0, + 1982.0, + 287.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 1980.0, + 279.0, + 1980.0, + 279.0, + 2009.0, + 254.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1972.0, + 1266.0, + 1972.0, + 1266.0, + 2015.0, + 349.0, + 2015.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 266, + 159, + 1411, + 159, + 1411, + 2018, + 266, + 2018 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 835, + 2060, + 864, + 2060, + 864, + 2087, + 835, + 2087 + ], + "score": 0.832 + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2058.0, + 869.0, + 2058.0, + 869.0, + 2097.0, + 831.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 203.0, + 1408.0, + 203.0, + 1408.0, + 242.0, + 290.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 233.0, + 922.0, + 233.0, + 922.0, + 272.0, + 353.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 284.0, + 1406.0, + 284.0, + 1406.0, + 323.0, + 290.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 313.0, + 1412.0, + 313.0, + 1412.0, + 356.0, + 347.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 341.0, + 522.0, + 341.0, + 522.0, + 385.0, + 352.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 397.0, + 1406.0, + 397.0, + 1406.0, + 436.0, + 294.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 429.0, + 1406.0, + 429.0, + 1406.0, + 468.0, + 353.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 460.0, + 613.0, + 460.0, + 613.0, + 493.0, + 351.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 511.0, + 1408.0, + 511.0, + 1408.0, + 550.0, + 294.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 540.0, + 1410.0, + 540.0, + 1410.0, + 580.0, + 351.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 572.0, + 974.0, + 572.0, + 974.0, + 611.0, + 351.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 623.0, + 1410.0, + 623.0, + 1410.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 658.0, + 1408.0, + 658.0, + 1408.0, + 691.0, + 355.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 682.0, + 1406.0, + 682.0, + 1406.0, + 721.0, + 351.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 713.0, + 1010.0, + 713.0, + 1010.0, + 752.0, + 349.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 770.0, + 1406.0, + 770.0, + 1406.0, + 803.0, + 296.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 799.0, + 1408.0, + 799.0, + 1408.0, + 838.0, + 351.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 827.0, + 497.0, + 827.0, + 497.0, + 866.0, + 353.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 878.0, + 1408.0, + 878.0, + 1408.0, + 917.0, + 290.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 909.0, + 808.0, + 909.0, + 808.0, + 948.0, + 349.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 958.0, + 1410.0, + 958.0, + 1410.0, + 1005.0, + 286.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 986.0, + 1408.0, + 986.0, + 1408.0, + 1038.0, + 347.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1021.0, + 431.0, + 1021.0, + 431.0, + 1062.0, + 351.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1076.0, + 1404.0, + 1076.0, + 1404.0, + 1115.0, + 288.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1107.0, + 943.0, + 1107.0, + 943.0, + 1144.0, + 349.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1158.0, + 1406.0, + 1158.0, + 1406.0, + 1199.0, + 290.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1188.0, + 1408.0, + 1188.0, + 1408.0, + 1229.0, + 349.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1217.0, + 654.0, + 1217.0, + 654.0, + 1258.0, + 351.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1272.0, + 1406.0, + 1272.0, + 1406.0, + 1311.0, + 286.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1303.0, + 1297.0, + 1303.0, + 1297.0, + 1342.0, + 353.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1358.0, + 1406.0, + 1358.0, + 1406.0, + 1391.0, + 288.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1380.0, + 1412.0, + 1380.0, + 1412.0, + 1427.0, + 349.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1413.0, + 660.0, + 1413.0, + 660.0, + 1454.0, + 353.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1464.0, + 1404.0, + 1464.0, + 1404.0, + 1503.0, + 290.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1499.0, + 1408.0, + 1499.0, + 1408.0, + 1537.0, + 353.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1529.0, + 429.0, + 1529.0, + 429.0, + 1564.0, + 355.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1580.0, + 1408.0, + 1580.0, + 1408.0, + 1619.0, + 290.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1615.0, + 1404.0, + 1615.0, + 1404.0, + 1648.0, + 355.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1639.0, + 722.0, + 1639.0, + 722.0, + 1678.0, + 349.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1693.0, + 1406.0, + 1693.0, + 1406.0, + 1731.0, + 288.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1727.0, + 1326.0, + 1727.0, + 1326.0, + 1760.0, + 353.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1776.0, + 1406.0, + 1776.0, + 1406.0, + 1815.0, + 288.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1807.0, + 1109.0, + 1807.0, + 1109.0, + 1846.0, + 353.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1862.0, + 1404.0, + 1862.0, + 1404.0, + 1895.0, + 288.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1891.0, + 992.0, + 1891.0, + 992.0, + 1925.0, + 351.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1941.0, + 1408.0, + 1941.0, + 1408.0, + 1980.0, + 288.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 261.0, + 1984.0, + 277.0, + 1984.0, + 277.0, + 2003.0, + 261.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1972.0, + 1176.0, + 1972.0, + 1176.0, + 2011.0, + 351.0, + 2011.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 407, + 789, + 1405, + 789, + 1405, + 1081, + 407, + 1081 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 406, + 300, + 1404, + 300, + 1404, + 621, + 406, + 621 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 407, + 1448, + 1405, + 1448, + 1405, + 1645, + 407, + 1645 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 406, + 1131, + 1404, + 1131, + 1404, + 1398, + 406, + 1398 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 404, + 671, + 1258, + 671, + 1258, + 739, + 404, + 739 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 367, + 629, + 841, + 629, + 841, + 662, + 367, + 662 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 364, + 749, + 671, + 749, + 671, + 781, + 364, + 781 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 365, + 258, + 580, + 258, + 580, + 288, + 365, + 288 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 365, + 1407, + 1201, + 1407, + 1201, + 1439, + 365, + 1439 + ], + "score": 0.868 + }, + { + "category_id": 1, + "poly": [ + 372, + 1091, + 1402, + 1091, + 1402, + 1124, + 372, + 1124 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.803 + }, + { + "category_id": 2, + "poly": [ + 242, + 1679, + 274, + 1679, + 274, + 1699, + 242, + 1699 + ], + "score": 0.761 + }, + { + "category_id": 0, + "poly": [ + 242, + 199, + 435, + 199, + 435, + 235, + 242, + 235 + ], + "score": 0.637 + }, + { + "category_id": 2, + "poly": [ + 237, + 206, + 287, + 206, + 287, + 1707, + 237, + 1707 + ], + "score": 0.286 + }, + { + "category_id": 2, + "poly": [ + 242, + 199, + 435, + 199, + 435, + 235, + 242, + 235 + ], + "score": 0.272 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.113 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1677.0, + 278.0, + 1677.0, + 278.0, + 1702.0, + 239.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 207.0, + 278.0, + 207.0, + 278.0, + 235.0, + 240.0, + 235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 196.0, + 440.0, + 196.0, + 440.0, + 239.0, + 290.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 206.0, + 281.0, + 206.0, + 281.0, + 243.0, + 237.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 261.0, + 281.0, + 261.0, + 281.0, + 299.0, + 237.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 304.0, + 278.0, + 304.0, + 278.0, + 336.0, + 237.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 330.0, + 282.0, + 330.0, + 282.0, + 372.0, + 236.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 371.0, + 279.0, + 371.0, + 279.0, + 403.0, + 237.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 399.0, + 281.0, + 399.0, + 281.0, + 436.0, + 235.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 436.0, + 279.0, + 436.0, + 279.0, + 468.0, + 237.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 466.0, + 279.0, + 466.0, + 279.0, + 498.0, + 237.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 498.0, + 279.0, + 498.0, + 279.0, + 530.0, + 237.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 526.0, + 281.0, + 526.0, + 281.0, + 563.0, + 237.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 565.0, + 279.0, + 565.0, + 279.0, + 595.0, + 239.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 593.0, + 281.0, + 593.0, + 281.0, + 633.0, + 237.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 633.0, + 278.0, + 633.0, + 278.0, + 673.0, + 237.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 676.0, + 278.0, + 676.0, + 278.0, + 710.0, + 239.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 713.0, + 278.0, + 713.0, + 278.0, + 746.0, + 239.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 753.0, + 279.0, + 753.0, + 279.0, + 793.0, + 237.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 795.0, + 278.0, + 795.0, + 278.0, + 826.0, + 239.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 826.0, + 279.0, + 826.0, + 279.0, + 858.0, + 237.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 856.0, + 278.0, + 856.0, + 278.0, + 890.0, + 239.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 890.0, + 281.0, + 890.0, + 281.0, + 926.0, + 235.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 923.0, + 279.0, + 923.0, + 279.0, + 955.0, + 239.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 958.0, + 279.0, + 958.0, + 279.0, + 990.0, + 237.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 988.0, + 278.0, + 988.0, + 278.0, + 1022.0, + 239.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1023.0, + 279.0, + 1023.0, + 279.0, + 1055.0, + 239.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1053.0, + 281.0, + 1053.0, + 281.0, + 1090.0, + 237.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1095.0, + 279.0, + 1095.0, + 279.0, + 1133.0, + 237.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1135.0, + 281.0, + 1135.0, + 281.0, + 1172.0, + 237.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1172.0, + 281.0, + 1172.0, + 281.0, + 1208.0, + 237.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1208.0, + 278.0, + 1208.0, + 278.0, + 1242.0, + 239.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1240.0, + 279.0, + 1240.0, + 279.0, + 1272.0, + 237.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1275.0, + 279.0, + 1275.0, + 279.0, + 1307.0, + 239.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1303.0, + 281.0, + 1303.0, + 281.0, + 1340.0, + 237.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1340.0, + 278.0, + 1340.0, + 278.0, + 1372.0, + 239.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1370.0, + 279.0, + 1370.0, + 279.0, + 1408.0, + 237.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1412.0, + 281.0, + 1412.0, + 281.0, + 1450.0, + 237.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1453.0, + 279.0, + 1453.0, + 279.0, + 1485.0, + 237.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1482.0, + 281.0, + 1482.0, + 281.0, + 1519.0, + 235.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1517.0, + 281.0, + 1517.0, + 281.0, + 1554.0, + 235.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1549.0, + 281.0, + 1549.0, + 281.0, + 1587.0, + 237.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1584.0, + 281.0, + 1584.0, + 281.0, + 1620.0, + 235.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1615.0, + 281.0, + 1615.0, + 281.0, + 1654.0, + 237.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1674.0, + 281.0, + 1674.0, + 281.0, + 1710.0, + 237.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 207.0, + 278.0, + 207.0, + 278.0, + 235.0, + 240.0, + 235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 196.0, + 440.0, + 196.0, + 440.0, + 239.0, + 290.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 788.0, + 1407.0, + 788.0, + 1407.0, + 826.0, + 405.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 819.0, + 1405.0, + 819.0, + 1405.0, + 855.0, + 449.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 850.0, + 977.0, + 850.0, + 977.0, + 885.0, + 450.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 885.0, + 1405.0, + 885.0, + 1405.0, + 922.0, + 406.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 916.0, + 860.0, + 916.0, + 860.0, + 951.0, + 449.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 949.0, + 1410.0, + 949.0, + 1410.0, + 991.0, + 407.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 984.0, + 1003.0, + 984.0, + 1003.0, + 1015.0, + 451.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1014.0, + 1405.0, + 1014.0, + 1405.0, + 1056.0, + 408.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1049.0, + 1022.0, + 1049.0, + 1022.0, + 1083.0, + 451.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 298.0, + 1405.0, + 298.0, + 1405.0, + 334.0, + 404.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 329.0, + 802.0, + 329.0, + 802.0, + 363.0, + 450.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 364.0, + 1406.0, + 364.0, + 1406.0, + 401.0, + 402.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 395.0, + 1030.0, + 395.0, + 1030.0, + 428.0, + 451.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 432.0, + 1406.0, + 432.0, + 1406.0, + 466.0, + 407.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 464.0, + 1405.0, + 464.0, + 1405.0, + 494.0, + 450.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 490.0, + 1405.0, + 490.0, + 1405.0, + 526.0, + 448.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 522.0, + 625.0, + 522.0, + 625.0, + 557.0, + 449.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 556.0, + 1404.0, + 556.0, + 1404.0, + 593.0, + 403.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 584.0, + 595.0, + 584.0, + 595.0, + 624.0, + 448.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1446.0, + 1408.0, + 1446.0, + 1408.0, + 1484.0, + 404.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1478.0, + 661.0, + 1478.0, + 661.0, + 1513.0, + 448.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1514.0, + 1404.0, + 1514.0, + 1404.0, + 1550.0, + 405.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1544.0, + 952.0, + 1544.0, + 952.0, + 1580.0, + 450.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1580.0, + 1405.0, + 1580.0, + 1405.0, + 1616.0, + 408.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1613.0, + 926.0, + 1613.0, + 926.0, + 1647.0, + 448.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1130.0, + 1311.0, + 1130.0, + 1311.0, + 1166.0, + 405.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1167.0, + 1001.0, + 1167.0, + 1001.0, + 1203.0, + 404.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1201.0, + 1405.0, + 1201.0, + 1405.0, + 1240.0, + 405.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1268.0, + 1405.0, + 1268.0, + 1405.0, + 1307.0, + 403.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1301.0, + 705.0, + 1301.0, + 705.0, + 1334.0, + 450.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1333.0, + 1403.0, + 1333.0, + 1403.0, + 1371.0, + 407.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1365.0, + 905.0, + 1365.0, + 905.0, + 1401.0, + 450.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 670.0, + 1259.0, + 670.0, + 1259.0, + 705.0, + 405.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 705.0, + 1180.0, + 705.0, + 1180.0, + 743.0, + 403.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 627.0, + 846.0, + 627.0, + 846.0, + 666.0, + 360.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 743.0, + 675.0, + 743.0, + 675.0, + 788.0, + 358.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 253.0, + 586.0, + 253.0, + 586.0, + 292.0, + 361.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1402.0, + 1204.0, + 1402.0, + 1204.0, + 1446.0, + 357.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1088.0, + 1408.0, + 1088.0, + 1408.0, + 1128.0, + 362.0, + 1128.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/q8qLAbQBupm/images/014eb4ab91ae963c4570120ed7f8282ecac097a50c0b193cdfd506aee144c3de.jpg b/parse/train/q8qLAbQBupm/images/014eb4ab91ae963c4570120ed7f8282ecac097a50c0b193cdfd506aee144c3de.jpg new file mode 100644 index 0000000000000000000000000000000000000000..675698e328d80e232e0a44e255bebb79c3c15133 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/014eb4ab91ae963c4570120ed7f8282ecac097a50c0b193cdfd506aee144c3de.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4d51452d71eece75f40c7f70ae0b3ed0bad888abb78a2db5d5431a046714e44 +size 6168 diff --git a/parse/train/q8qLAbQBupm/images/072e4e20aaf60f904f7b84c0db8f9599c3aa05b29f33f1b4750ddfda55313193.jpg b/parse/train/q8qLAbQBupm/images/072e4e20aaf60f904f7b84c0db8f9599c3aa05b29f33f1b4750ddfda55313193.jpg new file mode 100644 index 0000000000000000000000000000000000000000..057d973d5f002196b49fd3a03c73a73f1b7af452 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/072e4e20aaf60f904f7b84c0db8f9599c3aa05b29f33f1b4750ddfda55313193.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01772fa74f082bc14b4c77ba57d7c85b67d46f59a476505b19cb0d15c92a26bf +size 5531 diff --git a/parse/train/q8qLAbQBupm/images/0b98ae6fd2a4485b720820ef3c77f324fd52428db3cd8c65b9af58a81074bc02.jpg b/parse/train/q8qLAbQBupm/images/0b98ae6fd2a4485b720820ef3c77f324fd52428db3cd8c65b9af58a81074bc02.jpg new file mode 100644 index 0000000000000000000000000000000000000000..195693a7a396c38845aa565876c492a94f7c9894 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/0b98ae6fd2a4485b720820ef3c77f324fd52428db3cd8c65b9af58a81074bc02.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a92da781791d69a7df912d4eccb3ca6759e85a41896e0ed6cd24e7b81f495058 +size 3053 diff --git a/parse/train/q8qLAbQBupm/images/0f83a44d5c688f64197910e970714dbc79cf67d663dc38bec30962df8589355d.jpg b/parse/train/q8qLAbQBupm/images/0f83a44d5c688f64197910e970714dbc79cf67d663dc38bec30962df8589355d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e39f2d4518ee54e4b106cf5c93ead651f9f14c72 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/0f83a44d5c688f64197910e970714dbc79cf67d663dc38bec30962df8589355d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3d991f99f2096410456951a0aaa65d69bc56f16f355280bd41e937f89c6b795 +size 74143 diff --git a/parse/train/q8qLAbQBupm/images/10e3800d9f717b3a1887a1d9f97cbba6dd1652b3b344334c511816b33f435f0a.jpg b/parse/train/q8qLAbQBupm/images/10e3800d9f717b3a1887a1d9f97cbba6dd1652b3b344334c511816b33f435f0a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5b18d9f29473ff0d000943cd48e3ae1c0a79e9df --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/10e3800d9f717b3a1887a1d9f97cbba6dd1652b3b344334c511816b33f435f0a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b0376d2eefa2e98bd5882ee4a8394cb07679892a9b1a743390b36b73fe98492 +size 23209 diff --git a/parse/train/q8qLAbQBupm/images/14588611e4b7bf6b88d0e136ba33ad3cdae4c6019a621890a0c2d0a614447167.jpg b/parse/train/q8qLAbQBupm/images/14588611e4b7bf6b88d0e136ba33ad3cdae4c6019a621890a0c2d0a614447167.jpg new file mode 100644 index 0000000000000000000000000000000000000000..956d7373115bad30598f299b38ddc052a41eec1f --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/14588611e4b7bf6b88d0e136ba33ad3cdae4c6019a621890a0c2d0a614447167.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfe10a5203cecf366858b602c8b7a45f57651a78e8111e4f5dec0394e0ecf8ea +size 5783 diff --git a/parse/train/q8qLAbQBupm/images/14dfb38bdd1904a26380322e51f9188c612d30021a334bdbf5f7da60d0bf6cd8.jpg b/parse/train/q8qLAbQBupm/images/14dfb38bdd1904a26380322e51f9188c612d30021a334bdbf5f7da60d0bf6cd8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d15b71f4d2fc2daa43ecc7a0ce7d4be58c53d763 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/14dfb38bdd1904a26380322e51f9188c612d30021a334bdbf5f7da60d0bf6cd8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:565731df4b499dd441e943392b45628fde3095ff2d832118f8bcec79d02c8b67 +size 3791 diff --git a/parse/train/q8qLAbQBupm/images/21b60468151cdb79f1542f6a21186e904d323515b932b94a25abb0ee2fee734f.jpg b/parse/train/q8qLAbQBupm/images/21b60468151cdb79f1542f6a21186e904d323515b932b94a25abb0ee2fee734f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9755c61bbed3e5e4b9eb75af8f38aaa311417d12 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/21b60468151cdb79f1542f6a21186e904d323515b932b94a25abb0ee2fee734f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9d457cb36b2c3c070ab1dfa2d5876d8056528041684b845b2b84a32451ebd6a +size 68972 diff --git a/parse/train/q8qLAbQBupm/images/2425d6a3f932abfbdde03104445bb9c616cdfb3ff3a53be31c855122bdc3b1d8.jpg b/parse/train/q8qLAbQBupm/images/2425d6a3f932abfbdde03104445bb9c616cdfb3ff3a53be31c855122bdc3b1d8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b06c75e26f1a1745b46f8355ed5e6b7bbbe6cdc5 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/2425d6a3f932abfbdde03104445bb9c616cdfb3ff3a53be31c855122bdc3b1d8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a70d56e33e3e4c378ba8c0219b6a5f36bc1a7a9ac6a99cf97aefb65636cc429 +size 5358 diff --git a/parse/train/q8qLAbQBupm/images/24b95a40eff4320c1444a6c019163606c92dc3587cfa9406f332f397a0dba169.jpg b/parse/train/q8qLAbQBupm/images/24b95a40eff4320c1444a6c019163606c92dc3587cfa9406f332f397a0dba169.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8462d2d010fde3615daf08e4711f5b7817c1abbf --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/24b95a40eff4320c1444a6c019163606c92dc3587cfa9406f332f397a0dba169.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f7cf1327a960d1f4511bbc02c713674a9519fae8a3c5fcbe1e0a68c98534a99 +size 34988 diff --git a/parse/train/q8qLAbQBupm/images/27c014a4326286440095a1d5532c140ab5340ffb55d5b5bdf209dc118f8cac84.jpg b/parse/train/q8qLAbQBupm/images/27c014a4326286440095a1d5532c140ab5340ffb55d5b5bdf209dc118f8cac84.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5b8b6cdc0e190e6f1089db4558bbdf46e371a728 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/27c014a4326286440095a1d5532c140ab5340ffb55d5b5bdf209dc118f8cac84.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9612bb9f936f540f295ad0f503b111beb98ae4a6f588d622c610d0556299b7d2 +size 7625 diff --git a/parse/train/q8qLAbQBupm/images/27dc6aab0f9f11225838c15c56afa53303ea2742db36a2e96efc5e8bae9659fc.jpg b/parse/train/q8qLAbQBupm/images/27dc6aab0f9f11225838c15c56afa53303ea2742db36a2e96efc5e8bae9659fc.jpg new file mode 100644 index 0000000000000000000000000000000000000000..793ef6f59bca96440678d7311b6d945eeeebcfaf --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/27dc6aab0f9f11225838c15c56afa53303ea2742db36a2e96efc5e8bae9659fc.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d4666f5b85870830219eaf1046f6b1dfd40aa0ee8a5fa9216061db484205612 +size 8643 diff --git a/parse/train/q8qLAbQBupm/images/27de2c27c7b09d1b348dfc6ede27fdf25c957389ee1403d551357af6be9b1040.jpg b/parse/train/q8qLAbQBupm/images/27de2c27c7b09d1b348dfc6ede27fdf25c957389ee1403d551357af6be9b1040.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2353a94e671df447823073e9332d665e6f72aa79 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/27de2c27c7b09d1b348dfc6ede27fdf25c957389ee1403d551357af6be9b1040.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f70aa45aa36d283a38fb61b4d52b74d6bfe2a9c9ecfe68712dc030e8555ec7a +size 100186 diff --git a/parse/train/q8qLAbQBupm/images/2aceac675b97d230f237ae17d2dd9d511ba2807bfccc2b993ee3875ce6dfeeb7.jpg b/parse/train/q8qLAbQBupm/images/2aceac675b97d230f237ae17d2dd9d511ba2807bfccc2b993ee3875ce6dfeeb7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..605777643182bf161a092d4d6c78a288cef00c08 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/2aceac675b97d230f237ae17d2dd9d511ba2807bfccc2b993ee3875ce6dfeeb7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26d907cd511cb00cd688f94c3021e354bd3f224601b149cd9178a6e7347857c6 +size 7339 diff --git a/parse/train/q8qLAbQBupm/images/2b3806b48e62002be503b30f12fa7f7e71cd766086d0374ebb6f3d4c969d116a.jpg b/parse/train/q8qLAbQBupm/images/2b3806b48e62002be503b30f12fa7f7e71cd766086d0374ebb6f3d4c969d116a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..31ea7bc0fa6d5702d928ea2b765a2738389808d0 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/2b3806b48e62002be503b30f12fa7f7e71cd766086d0374ebb6f3d4c969d116a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69667b6d48e493ce5d4a8abf6be0d8f3ee6ba2b10052327b5edd5ca936450767 +size 2692 diff --git a/parse/train/q8qLAbQBupm/images/2b44755943eb37c0e6e58dcd2693116a29aaad16eb86a92a2b7daae28821400c.jpg b/parse/train/q8qLAbQBupm/images/2b44755943eb37c0e6e58dcd2693116a29aaad16eb86a92a2b7daae28821400c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6b584f549c0b0089590c5e15953a7e08cfd96e86 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/2b44755943eb37c0e6e58dcd2693116a29aaad16eb86a92a2b7daae28821400c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4c4a5f7561d1772042760f2658fad43990c8aa5e7aeb0d42dabb47ae638198d +size 4269 diff --git a/parse/train/q8qLAbQBupm/images/2b647e203256e6f764e57041fb1916128f72a245b45101af70d35375e0aa0428.jpg b/parse/train/q8qLAbQBupm/images/2b647e203256e6f764e57041fb1916128f72a245b45101af70d35375e0aa0428.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0c1113554894cdb7a2e6e30212068cc5623852f6 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/2b647e203256e6f764e57041fb1916128f72a245b45101af70d35375e0aa0428.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f77faa50ad6b5e73cf9676e28ca56be8256384cee6810508940ceb66e596765f +size 3045 diff --git a/parse/train/q8qLAbQBupm/images/2f38ada903d1774e9844d293c9270ccbf916d782bcff5ad00971f89830535221.jpg b/parse/train/q8qLAbQBupm/images/2f38ada903d1774e9844d293c9270ccbf916d782bcff5ad00971f89830535221.jpg new file mode 100644 index 0000000000000000000000000000000000000000..99a3dbefd529ef03374667af5d2bdd7e8720a921 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/2f38ada903d1774e9844d293c9270ccbf916d782bcff5ad00971f89830535221.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acd7f6d22ad19a7b74dda6acdb758d56d58cd3cc54fee1d0dbb6093a6e660e20 +size 65137 diff --git a/parse/train/q8qLAbQBupm/images/30bc8baeb6da8b198605060b23b4adecb65474ff2a813e3391ea88aa7aa2f9e4.jpg b/parse/train/q8qLAbQBupm/images/30bc8baeb6da8b198605060b23b4adecb65474ff2a813e3391ea88aa7aa2f9e4.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4c3b9a4c6fbd585f7e8f5d9554db918d42a4995f --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/30bc8baeb6da8b198605060b23b4adecb65474ff2a813e3391ea88aa7aa2f9e4.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a55c4c601cb27c0b3af63944821ad65b11ad5adb3edee557ad8c493d6a49c54 +size 3329 diff --git a/parse/train/q8qLAbQBupm/images/313daca5a9876c1516899cb7b462b2decc3b5ee6c1f037966e1f2a05e144fff1.jpg b/parse/train/q8qLAbQBupm/images/313daca5a9876c1516899cb7b462b2decc3b5ee6c1f037966e1f2a05e144fff1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2fd112d0890bc08d2ff30f47889aa0443ce570eb --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/313daca5a9876c1516899cb7b462b2decc3b5ee6c1f037966e1f2a05e144fff1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6eada31f0a6208665e2713779a1ce6351575f91f3b6e39ce6d0795616abc4638 +size 25139 diff --git a/parse/train/q8qLAbQBupm/images/32ae547a17365a43b018cf00371c44289fbda99b2e1a230596dd00d82e53e9e5.jpg b/parse/train/q8qLAbQBupm/images/32ae547a17365a43b018cf00371c44289fbda99b2e1a230596dd00d82e53e9e5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1174b6bef6b021b24f32a34f382f9d69ab0de78e --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/32ae547a17365a43b018cf00371c44289fbda99b2e1a230596dd00d82e53e9e5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4a4139b067c0ec757a91f28744a0ed11d8c93cd230469db0b0161ec61ae8919 +size 5518 diff --git a/parse/train/q8qLAbQBupm/images/33223d5c1735ae62f5759480eb2f16edc6099065b2d6418ba88ce65ce67ba5db.jpg b/parse/train/q8qLAbQBupm/images/33223d5c1735ae62f5759480eb2f16edc6099065b2d6418ba88ce65ce67ba5db.jpg new file mode 100644 index 0000000000000000000000000000000000000000..536d1ba9c7825a5bb09ed944f1b3c5e56db83e1e --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/33223d5c1735ae62f5759480eb2f16edc6099065b2d6418ba88ce65ce67ba5db.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e42eb91f083d040e968612c59cd3c7138c704a9304d16e8832104f89d0ff212 +size 24492 diff --git a/parse/train/q8qLAbQBupm/images/337c42ea29bbf6ee984f8fa9016705f64106cf73e23394cf038de94ef089ccbc.jpg b/parse/train/q8qLAbQBupm/images/337c42ea29bbf6ee984f8fa9016705f64106cf73e23394cf038de94ef089ccbc.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7a97e44cc488077ea7702df638e114c743cf2445 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/337c42ea29bbf6ee984f8fa9016705f64106cf73e23394cf038de94ef089ccbc.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23cc999a4e4d4f83a5892141180398fdb6d579546a4aa8b0d50594d44d6f1b1d +size 21777 diff --git a/parse/train/q8qLAbQBupm/images/33d12c6f91336cb141d6b4737908c5115579629b0974de92c28675c07b17e70b.jpg b/parse/train/q8qLAbQBupm/images/33d12c6f91336cb141d6b4737908c5115579629b0974de92c28675c07b17e70b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..135f42d91518b6803ec3695cb79ac622d26d8313 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/33d12c6f91336cb141d6b4737908c5115579629b0974de92c28675c07b17e70b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9c8c56ea63e42212ab57cd1c7e4fd8b3f27984cb661232f52582a854d65359a +size 7106 diff --git a/parse/train/q8qLAbQBupm/images/361f7f500144d8c8dee9d489b7b19b5a984be4e525e81758d3a6929dcc2b07cb.jpg b/parse/train/q8qLAbQBupm/images/361f7f500144d8c8dee9d489b7b19b5a984be4e525e81758d3a6929dcc2b07cb.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a914f0f6738c12479243f69571a99fc96b51edb3 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/361f7f500144d8c8dee9d489b7b19b5a984be4e525e81758d3a6929dcc2b07cb.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2babda4448e10816047155348dbf8ccd9e4165419e6cbd0f522ea127f6d907d +size 10623 diff --git a/parse/train/q8qLAbQBupm/images/36baf9655cd7cff950657516b9a34ed8e5ee27f853cb104fee8cd87034f56c94.jpg b/parse/train/q8qLAbQBupm/images/36baf9655cd7cff950657516b9a34ed8e5ee27f853cb104fee8cd87034f56c94.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9316c1c08585753b16401ecb714e31351564de03 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/36baf9655cd7cff950657516b9a34ed8e5ee27f853cb104fee8cd87034f56c94.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d224dd67fe2da3201c6a8365dca337b8d3e62d3edace7fd53b33747d4cac47ec +size 6821 diff --git a/parse/train/q8qLAbQBupm/images/37cbcc81a0ffebfcfc0a4b7237647b167e7caf672ef12f362450e2101c150b13.jpg b/parse/train/q8qLAbQBupm/images/37cbcc81a0ffebfcfc0a4b7237647b167e7caf672ef12f362450e2101c150b13.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fb0f24c8f2821ea2a7cb38461cb69c38ab1b268f --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/37cbcc81a0ffebfcfc0a4b7237647b167e7caf672ef12f362450e2101c150b13.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:367bb542168bc5f49843a8f9e07ffc4d80a1f79b881eba351a4a7025dd16b44d +size 34716 diff --git a/parse/train/q8qLAbQBupm/images/3912db89633b1dbdc0d4ff6462c19a44e6f455594b59c51ba93d8243f54a7e83.jpg b/parse/train/q8qLAbQBupm/images/3912db89633b1dbdc0d4ff6462c19a44e6f455594b59c51ba93d8243f54a7e83.jpg new file mode 100644 index 0000000000000000000000000000000000000000..983f7889bcd5cbd8328e4437b7805c26cb8fa989 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/3912db89633b1dbdc0d4ff6462c19a44e6f455594b59c51ba93d8243f54a7e83.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:214504b054ee88ca615728dc87732d5bc5561d481f50ad4577a87b84070cdc07 +size 21084 diff --git a/parse/train/q8qLAbQBupm/images/3cade64d3c53796171a588157b239c3e0d7a58116cfd5606d13c1c4a880b086f.jpg b/parse/train/q8qLAbQBupm/images/3cade64d3c53796171a588157b239c3e0d7a58116cfd5606d13c1c4a880b086f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d579b31907b70732f758e6a717a002ea2308c821 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/3cade64d3c53796171a588157b239c3e0d7a58116cfd5606d13c1c4a880b086f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a900712bf06dc7606da26e5681758c7578b0d919865c89f65734260a8be62ca +size 2515 diff --git a/parse/train/q8qLAbQBupm/images/40a7856880e390c7f9c4d304da63ed4a3b59bdd1c8d2f4accb5f7cccf50e32c9.jpg b/parse/train/q8qLAbQBupm/images/40a7856880e390c7f9c4d304da63ed4a3b59bdd1c8d2f4accb5f7cccf50e32c9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1f4fab7dfeae23d3a4f402371cc402952f09b4d5 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/40a7856880e390c7f9c4d304da63ed4a3b59bdd1c8d2f4accb5f7cccf50e32c9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c7d327cc4e8dce0ad07272da58150e49be198d4e7308ee34dc9b4e63d098870 +size 3876 diff --git a/parse/train/q8qLAbQBupm/images/42ade1aaa2fea3a24f346d0859b22ce827ef7959fa96e6c5276b3c6ffa678731.jpg b/parse/train/q8qLAbQBupm/images/42ade1aaa2fea3a24f346d0859b22ce827ef7959fa96e6c5276b3c6ffa678731.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5dc595e502a25a952325fa35fa19a4a0ba1db997 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/42ade1aaa2fea3a24f346d0859b22ce827ef7959fa96e6c5276b3c6ffa678731.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1aa158ce86b38b0a9198a6779b0f4b38b9309e0cea71ff04a24c68ed103aa1e +size 5214 diff --git a/parse/train/q8qLAbQBupm/images/43fe7578def07664069cf56d3d54817faa173482e778921547d7318ea42b3398.jpg b/parse/train/q8qLAbQBupm/images/43fe7578def07664069cf56d3d54817faa173482e778921547d7318ea42b3398.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7603228cd6c7ee66cdd5173308cfbfcab276cced --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/43fe7578def07664069cf56d3d54817faa173482e778921547d7318ea42b3398.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e36d8823c350de46ea9f5d5e2ed214779779a2c8ab58fe64a7ff388f405dbc9d +size 7997 diff --git a/parse/train/q8qLAbQBupm/images/4648ceb7246033fcbb6d118fdebee5202c67e92b292e81148e841906410cf98b.jpg b/parse/train/q8qLAbQBupm/images/4648ceb7246033fcbb6d118fdebee5202c67e92b292e81148e841906410cf98b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f22d51b7ce04734c70ce1d7c6008b6728715259d --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/4648ceb7246033fcbb6d118fdebee5202c67e92b292e81148e841906410cf98b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c0410965ad914cc1d36125ace8c59f6259103df1640e8e70b2c570ed1d6c725 +size 3018 diff --git a/parse/train/q8qLAbQBupm/images/4c7556d0a83836614d830391e7925a71cc6a142f3ec58e6decda840793bf434d.jpg b/parse/train/q8qLAbQBupm/images/4c7556d0a83836614d830391e7925a71cc6a142f3ec58e6decda840793bf434d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..eb5b06842401514b7f77c5fa3b5ccf7cb466ccba --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/4c7556d0a83836614d830391e7925a71cc6a142f3ec58e6decda840793bf434d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64abe4236943e7d257bbde667c399ae160718ed45eba53d1b47abfe21a3603cf +size 11364 diff --git a/parse/train/q8qLAbQBupm/images/4d69809524b6636889d4d935fc713b36cc0deaa4b6d8fca13ef12717303cefb5.jpg b/parse/train/q8qLAbQBupm/images/4d69809524b6636889d4d935fc713b36cc0deaa4b6d8fca13ef12717303cefb5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..23fda9f7609160b16a305f6205ed56fa598c831e --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/4d69809524b6636889d4d935fc713b36cc0deaa4b6d8fca13ef12717303cefb5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6f5315da7c87fda0b717ff39caffc50583a38a20f343b0ddf7ecc8dc2d2e5da +size 4953 diff --git a/parse/train/q8qLAbQBupm/images/52211e06d9b1958a0b722f1d31542bad68418ac83a895e914eb1a72193beceea.jpg b/parse/train/q8qLAbQBupm/images/52211e06d9b1958a0b722f1d31542bad68418ac83a895e914eb1a72193beceea.jpg new file mode 100644 index 0000000000000000000000000000000000000000..153eb115b341e2c3f01d0abbf6790c6cf1095709 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/52211e06d9b1958a0b722f1d31542bad68418ac83a895e914eb1a72193beceea.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f569ece4d32bb21a48136bbf0711b5a75ee5535bf46fe8ebcf0a9468ae298447 +size 3314 diff --git a/parse/train/q8qLAbQBupm/images/5231fa06f876fefaa9e078994d355ae466d8194fa55b0306a160a137f6a4ab21.jpg b/parse/train/q8qLAbQBupm/images/5231fa06f876fefaa9e078994d355ae466d8194fa55b0306a160a137f6a4ab21.jpg new file mode 100644 index 0000000000000000000000000000000000000000..23aa3889e076f5c650a2b20c3e58a4c8f8531170 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/5231fa06f876fefaa9e078994d355ae466d8194fa55b0306a160a137f6a4ab21.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0902de757ec8fabbb869064e5f23dbddae4b91e8d2c1d8218191454f4ae91fdf +size 17901 diff --git a/parse/train/q8qLAbQBupm/images/54b1c4c025ed605982c18e050b1ba728373121363cf7e8182c429d68ab043c0f.jpg b/parse/train/q8qLAbQBupm/images/54b1c4c025ed605982c18e050b1ba728373121363cf7e8182c429d68ab043c0f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b4665cd40b818654a00d59231324c9aec1d05149 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/54b1c4c025ed605982c18e050b1ba728373121363cf7e8182c429d68ab043c0f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9f832fb9587976c15258b53f5868249f8b0f6f4678bf4c0fe56d604627d553b +size 5323 diff --git a/parse/train/q8qLAbQBupm/images/564e3d22becb69306a94821b686ba46b7d241067d0e3f166752d9fe794d3ba1a.jpg b/parse/train/q8qLAbQBupm/images/564e3d22becb69306a94821b686ba46b7d241067d0e3f166752d9fe794d3ba1a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..19fdb3004a8b83f1b44bcd210e6155177c1dfab3 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/564e3d22becb69306a94821b686ba46b7d241067d0e3f166752d9fe794d3ba1a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9100fd02325d5583144e0ccb4c8e36d1a7e4beb471919f090717b6b361f181a0 +size 5622 diff --git a/parse/train/q8qLAbQBupm/images/58ed73a3d274cacd41aad292c4bfe29f94bc57f266ea29c7ded8aa319329d07d.jpg b/parse/train/q8qLAbQBupm/images/58ed73a3d274cacd41aad292c4bfe29f94bc57f266ea29c7ded8aa319329d07d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3a22d6c14d2e1d99c76448d9899368afdf41d553 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/58ed73a3d274cacd41aad292c4bfe29f94bc57f266ea29c7ded8aa319329d07d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae30a4119c8ae4f44e8cdb5dfd23634dcd1bd7a19cba776881ef71ad5a5985f1 +size 4887 diff --git a/parse/train/q8qLAbQBupm/images/5b527e65cc9772cd5638e156bd3fbddc631c168e4684847d01b6dad4dfacfde3.jpg b/parse/train/q8qLAbQBupm/images/5b527e65cc9772cd5638e156bd3fbddc631c168e4684847d01b6dad4dfacfde3.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9b40e6dae0a68a1b38687914a2fcd9a1a5c64e90 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/5b527e65cc9772cd5638e156bd3fbddc631c168e4684847d01b6dad4dfacfde3.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b42a9a506da737c76d6c6f9539f14282a9a98a977f10ac175e0d55c72da85ce +size 4920 diff --git a/parse/train/q8qLAbQBupm/images/5bad999307fc7c49046c7883ad310704ed432d2b70fd90dcd821a19528453cf0.jpg b/parse/train/q8qLAbQBupm/images/5bad999307fc7c49046c7883ad310704ed432d2b70fd90dcd821a19528453cf0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..45a5d579313cb90f5a0e628bdb36aa63bcbb7d88 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/5bad999307fc7c49046c7883ad310704ed432d2b70fd90dcd821a19528453cf0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:679047b6036d24a5b3450066101ae577e769c0c6bdd9f7ee00b22a5935aa6e74 +size 2883 diff --git a/parse/train/q8qLAbQBupm/images/5c0023591344e92b1c391238f680d88f900f3c6e2cf612bfc55c24ad99c71a00.jpg b/parse/train/q8qLAbQBupm/images/5c0023591344e92b1c391238f680d88f900f3c6e2cf612bfc55c24ad99c71a00.jpg new file mode 100644 index 0000000000000000000000000000000000000000..173dd6d26f9210dbcf41b6d0359c7ef01dd90f50 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/5c0023591344e92b1c391238f680d88f900f3c6e2cf612bfc55c24ad99c71a00.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a7e6869869a84d7751bc1331f213156759442ea88f3c797c4d69b56e47f167f +size 16982 diff --git a/parse/train/q8qLAbQBupm/images/5d4dda326005e6750fd36d9e1920123a56d83d44eb292eec5022aa9118d8d845.jpg b/parse/train/q8qLAbQBupm/images/5d4dda326005e6750fd36d9e1920123a56d83d44eb292eec5022aa9118d8d845.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bb835ceeb2a8517da8515147672e568722f7faec --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/5d4dda326005e6750fd36d9e1920123a56d83d44eb292eec5022aa9118d8d845.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bc4564f9e7fe904b30c726831e6121683a090c5706d0f3ac74e444ff7cab25c +size 9293 diff --git a/parse/train/q8qLAbQBupm/images/60c964594033396f053bd3aef85c0a0260c3bffd7716e993f7ab8abaaad920e0.jpg b/parse/train/q8qLAbQBupm/images/60c964594033396f053bd3aef85c0a0260c3bffd7716e993f7ab8abaaad920e0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a58ece964f05db0db3282fd886f5f4eb6f4be388 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/60c964594033396f053bd3aef85c0a0260c3bffd7716e993f7ab8abaaad920e0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0651b9cc84ebfc195d8739265b568c88fba1438d8b5e98e31babd88bc488d740 +size 6812 diff --git a/parse/train/q8qLAbQBupm/images/65e7127ca286c1d6aaa77b73f89d6ae1fd6b534ae9bd2ef7a9f8cdc0f62b2a61.jpg b/parse/train/q8qLAbQBupm/images/65e7127ca286c1d6aaa77b73f89d6ae1fd6b534ae9bd2ef7a9f8cdc0f62b2a61.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ad0ac39a42a7b133b1a2e2635e39eff6484e4652 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/65e7127ca286c1d6aaa77b73f89d6ae1fd6b534ae9bd2ef7a9f8cdc0f62b2a61.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd1ad9c7a378518a667cd6e0a68a92b3569b2d36f7a12835d4821398a3f74ed6 +size 27372 diff --git a/parse/train/q8qLAbQBupm/images/68e1ac171a33e685ef8104fe29178c0c8e2d615c856ce5aa3cc9f30bc1799711.jpg b/parse/train/q8qLAbQBupm/images/68e1ac171a33e685ef8104fe29178c0c8e2d615c856ce5aa3cc9f30bc1799711.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fd892bb4c735019d42cc702ac2e8976ca8b20b38 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/68e1ac171a33e685ef8104fe29178c0c8e2d615c856ce5aa3cc9f30bc1799711.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1724412281ddb665515e43e2187a2ec461826f4a9855ad6cd35df1d29e345660 +size 8257 diff --git a/parse/train/q8qLAbQBupm/images/6d512f4b5fa254ba24fd81ee6d642f0e1ed066fa0f4e020a88808896b4d6352f.jpg b/parse/train/q8qLAbQBupm/images/6d512f4b5fa254ba24fd81ee6d642f0e1ed066fa0f4e020a88808896b4d6352f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2e45a75b14d38dd07bdf1051f465ba5518b8c17e --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/6d512f4b5fa254ba24fd81ee6d642f0e1ed066fa0f4e020a88808896b4d6352f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75b431d0090cb0f0cb7ae4804cf02c2269f5f2955db12fea90f93ae6dbe43d52 +size 9872 diff --git a/parse/train/q8qLAbQBupm/images/6feb0fca50a81d082231d9dcf5e973aebbe5289c89c77b4c0cb29c00d7ed21ca.jpg b/parse/train/q8qLAbQBupm/images/6feb0fca50a81d082231d9dcf5e973aebbe5289c89c77b4c0cb29c00d7ed21ca.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f63aeb1e31ad66f364525fd819060661af53e729 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/6feb0fca50a81d082231d9dcf5e973aebbe5289c89c77b4c0cb29c00d7ed21ca.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91b3f76052582f7bb432e6899945553c56d65af141d275105c14a4f64111a3b9 +size 19961 diff --git a/parse/train/q8qLAbQBupm/images/73184e387dc767a9dae69b2b991295b2a58bbb37aca780b04a0326bffdda27d4.jpg b/parse/train/q8qLAbQBupm/images/73184e387dc767a9dae69b2b991295b2a58bbb37aca780b04a0326bffdda27d4.jpg new file mode 100644 index 0000000000000000000000000000000000000000..af1c7bd6e8cf74b3a34781a95e3f1bc39f03ad61 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/73184e387dc767a9dae69b2b991295b2a58bbb37aca780b04a0326bffdda27d4.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e89361d4375d69f38c383353efc01a6809bd939853435b6ab88efee943508b2a +size 2880 diff --git a/parse/train/q8qLAbQBupm/images/746ab7951a725eee22fdd6418dae175f174bdf6aee2f817023caf0f106d76d0e.jpg b/parse/train/q8qLAbQBupm/images/746ab7951a725eee22fdd6418dae175f174bdf6aee2f817023caf0f106d76d0e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d8c0d6f93fe025d75a6ccc432cd84a1eb46ee12c --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/746ab7951a725eee22fdd6418dae175f174bdf6aee2f817023caf0f106d76d0e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f2dbd91b7ec5361a72154448dc8afe76fa7e156013fb34430362b42335746f5 +size 18383 diff --git a/parse/train/q8qLAbQBupm/images/7bb16150dd8877f5b91cf743fb510f331013664279bf237e21a9be3c9c9f7b03.jpg b/parse/train/q8qLAbQBupm/images/7bb16150dd8877f5b91cf743fb510f331013664279bf237e21a9be3c9c9f7b03.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f77d318a3dd602b0df2f1c0048757186fef1a8b1 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/7bb16150dd8877f5b91cf743fb510f331013664279bf237e21a9be3c9c9f7b03.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c187ce65abcaf1986e833b92ad14f9104e5cc9038417fd7909fc5a936ea6455d +size 12056 diff --git a/parse/train/q8qLAbQBupm/images/7d75dac099eee90f394606af380d51cba4976551082acdc04454c53390bb9a74.jpg b/parse/train/q8qLAbQBupm/images/7d75dac099eee90f394606af380d51cba4976551082acdc04454c53390bb9a74.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1228e1549fb1daf8effb9760c88241c1de81807b --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/7d75dac099eee90f394606af380d51cba4976551082acdc04454c53390bb9a74.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34661e823e9a0804b5265fe8977116bafd4d8609f13ff783c3d8733506466e64 +size 5595 diff --git a/parse/train/q8qLAbQBupm/images/7f8f94af555b43bc78b4b1f4b3788db453ea3911de8c3595734660af72acfb76.jpg b/parse/train/q8qLAbQBupm/images/7f8f94af555b43bc78b4b1f4b3788db453ea3911de8c3595734660af72acfb76.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fa27d876cbfb74f7c0feb1e6be00e159a728f58b --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/7f8f94af555b43bc78b4b1f4b3788db453ea3911de8c3595734660af72acfb76.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8edf2533b82e7fb64a1e5e96e024875ae0cad656f7c8c075d8bac617f11b689 +size 33377 diff --git a/parse/train/q8qLAbQBupm/images/7fc192c7bcafd3b53ba28e721c802837abc6fc802d7059f61bd26709713646f4.jpg b/parse/train/q8qLAbQBupm/images/7fc192c7bcafd3b53ba28e721c802837abc6fc802d7059f61bd26709713646f4.jpg new file mode 100644 index 0000000000000000000000000000000000000000..44d6ddea8d5f7cd5d00cb1960d0ea9cf54e7585a --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/7fc192c7bcafd3b53ba28e721c802837abc6fc802d7059f61bd26709713646f4.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb1533ede4c9e3d13e9d8d5af97c8dae47fb7c3b2637a51dfaf93f58ca155718 +size 23885 diff --git a/parse/train/q8qLAbQBupm/images/8253f891f366430434455a040a9184eee9233645718b7ba0ac257f4d50a8de1b.jpg b/parse/train/q8qLAbQBupm/images/8253f891f366430434455a040a9184eee9233645718b7ba0ac257f4d50a8de1b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bc7237afa45e6c03e7d3f6ef96d87e7f0743f72f --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/8253f891f366430434455a040a9184eee9233645718b7ba0ac257f4d50a8de1b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbb9b3cadc6405a52b53872d92efe64e1d5fb519142290f197728e351a647399 +size 10952 diff --git a/parse/train/q8qLAbQBupm/images/82b2bb9e6eea6f91f82fc7adf8cddd73c89ba5e7e09dcb0c8b027d8e646c07bc.jpg b/parse/train/q8qLAbQBupm/images/82b2bb9e6eea6f91f82fc7adf8cddd73c89ba5e7e09dcb0c8b027d8e646c07bc.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d03099905372f2816901d58f3fadfcc6af527af0 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/82b2bb9e6eea6f91f82fc7adf8cddd73c89ba5e7e09dcb0c8b027d8e646c07bc.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f2932c2843c6f2ca5c572217a2b02f34ccc23508bc9a9c14c41b936a0ed6171 +size 2727 diff --git a/parse/train/q8qLAbQBupm/images/8403277cb0fc077fcd52166c5673f5ffc0fd57e55ee4618f66596ae01c6f3e1c.jpg b/parse/train/q8qLAbQBupm/images/8403277cb0fc077fcd52166c5673f5ffc0fd57e55ee4618f66596ae01c6f3e1c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1936f89d7fea0b059deb8990b30023a8382f0c1b --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/8403277cb0fc077fcd52166c5673f5ffc0fd57e55ee4618f66596ae01c6f3e1c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8401518b90f4468f64ee1077059d37cf90aa76ef38ce282bd154d231bdda1a4 +size 20995 diff --git a/parse/train/q8qLAbQBupm/images/84abbc8454ca523ff314027a032fbde6e101a84e8b9d0b4518ed827357646138.jpg b/parse/train/q8qLAbQBupm/images/84abbc8454ca523ff314027a032fbde6e101a84e8b9d0b4518ed827357646138.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d631b52cc322aca5d4e1168dadf2cb71b7dee0b0 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/84abbc8454ca523ff314027a032fbde6e101a84e8b9d0b4518ed827357646138.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e6a39493cfec434b46f5885530e61edc7590f33ca0afd8dde6aaac63cac6f3d +size 6697 diff --git a/parse/train/q8qLAbQBupm/images/898e257f95a61192d6cc4634c20188163630027af8474d37d5458352198a0d15.jpg b/parse/train/q8qLAbQBupm/images/898e257f95a61192d6cc4634c20188163630027af8474d37d5458352198a0d15.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2e80dcce1f864c6e65d895c399c08053edb6eb4b --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/898e257f95a61192d6cc4634c20188163630027af8474d37d5458352198a0d15.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a48eea5acfe00993a2c456efe54fe7dbffa74bcd3c473503f2ca2e81b3b4a997 +size 26686 diff --git a/parse/train/q8qLAbQBupm/images/89f317d5ccd238dd8327b48b399084f1a0cfda1bc6c5f868a08cb19cf05309f6.jpg b/parse/train/q8qLAbQBupm/images/89f317d5ccd238dd8327b48b399084f1a0cfda1bc6c5f868a08cb19cf05309f6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..de23a566d3fac98b1502113604d56b03896a60d4 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/89f317d5ccd238dd8327b48b399084f1a0cfda1bc6c5f868a08cb19cf05309f6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a0d3ee0cb49f9fb194776b0cc3bccdf0eceec646f0d34b9867c72c4fd7f94dc +size 61096 diff --git a/parse/train/q8qLAbQBupm/images/8a12603e5156994568067c0ce4a98392147e6c77437ec8796dcfe9d12433ecce.jpg b/parse/train/q8qLAbQBupm/images/8a12603e5156994568067c0ce4a98392147e6c77437ec8796dcfe9d12433ecce.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e8cf575d3b23c97e6e7349001bfee700856c28a9 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/8a12603e5156994568067c0ce4a98392147e6c77437ec8796dcfe9d12433ecce.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c88e9fead4f4760e7fb89bf923741febde4c6c6be1bdd017430ca19b37f6f34 +size 3143 diff --git a/parse/train/q8qLAbQBupm/images/8a99f34338e52043fa544d6c39c44d7816989d32564beeedddca3b0449dc22a5.jpg b/parse/train/q8qLAbQBupm/images/8a99f34338e52043fa544d6c39c44d7816989d32564beeedddca3b0449dc22a5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fad8a98e70352cf5e389ef9f6c9bb2a26215a8ea --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/8a99f34338e52043fa544d6c39c44d7816989d32564beeedddca3b0449dc22a5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0de9aa82085365ea806d7c77fff6851a4ab4a5a26da6f1a000cd2ae386c5ff24 +size 2560 diff --git a/parse/train/q8qLAbQBupm/images/8e63d4ca36bec563840253ba9e339b06b86a848fd88bab7d00f6c03694ecd6e6.jpg b/parse/train/q8qLAbQBupm/images/8e63d4ca36bec563840253ba9e339b06b86a848fd88bab7d00f6c03694ecd6e6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..230979fec77ac04481a9d1e3007ca5931a848e32 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/8e63d4ca36bec563840253ba9e339b06b86a848fd88bab7d00f6c03694ecd6e6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8163eb34956fe2cd0405b6e8ffcc80bbcf34ff48d1b538516d603b743d3ab355 +size 32955 diff --git a/parse/train/q8qLAbQBupm/images/971428dbbdab6f03facda63da6f07594209e176a7b1ab232fae10c13c50b8d4d.jpg b/parse/train/q8qLAbQBupm/images/971428dbbdab6f03facda63da6f07594209e176a7b1ab232fae10c13c50b8d4d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bf367a3c143dc86a37884b653f7c19e047aaea87 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/971428dbbdab6f03facda63da6f07594209e176a7b1ab232fae10c13c50b8d4d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d6129f4481d248a6c828aaef04482f6eeae14a934c09db3b26bf05be3c9f1c8 +size 14291 diff --git a/parse/train/q8qLAbQBupm/images/97368ee2b32a15943090d8a857a2a67cbdf349bb09b4ebab0ac3a93d0fbf4565.jpg b/parse/train/q8qLAbQBupm/images/97368ee2b32a15943090d8a857a2a67cbdf349bb09b4ebab0ac3a93d0fbf4565.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6a7c0f93d47a248f295fa6f67f6dfa828dd8cad6 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/97368ee2b32a15943090d8a857a2a67cbdf349bb09b4ebab0ac3a93d0fbf4565.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80b2429305bb2a380a487568f9e3c9fa5cca1258b280b812d96764dd7775e233 +size 27747 diff --git a/parse/train/q8qLAbQBupm/images/99c0202e7823897ffe45e495ab74ce686922f4e98a4b17ba6ff4d87a97852e8a.jpg b/parse/train/q8qLAbQBupm/images/99c0202e7823897ffe45e495ab74ce686922f4e98a4b17ba6ff4d87a97852e8a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8ce27b41791ae3353f151c91201b0f83a7ebcf3b --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/99c0202e7823897ffe45e495ab74ce686922f4e98a4b17ba6ff4d87a97852e8a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6075a4692f4cbc3b12aefac4dd86b5d6cbd5bc6ce45beb2257cd3bf5413e87b +size 3792 diff --git a/parse/train/q8qLAbQBupm/images/9b32c6681e6d788b662e86cbc617a89e7a085a6cc7638e4fda104fe998a38c58.jpg b/parse/train/q8qLAbQBupm/images/9b32c6681e6d788b662e86cbc617a89e7a085a6cc7638e4fda104fe998a38c58.jpg new file mode 100644 index 0000000000000000000000000000000000000000..65a59d8a3d37d8ce71d010885ee60b733e5eb6e7 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/9b32c6681e6d788b662e86cbc617a89e7a085a6cc7638e4fda104fe998a38c58.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edc29959ddd5f38138ffd8cd89f8a4800d172897c66ec75b7bbd79843a7f169d +size 32701 diff --git a/parse/train/q8qLAbQBupm/images/9e6222aee11ce3e460ccb6fcfb8947bd682fe54df800767a7b363db20e90665d.jpg b/parse/train/q8qLAbQBupm/images/9e6222aee11ce3e460ccb6fcfb8947bd682fe54df800767a7b363db20e90665d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ffeb557a89b4d1d007ced75f1dae429ab5d3e966 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/9e6222aee11ce3e460ccb6fcfb8947bd682fe54df800767a7b363db20e90665d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7a2e8406131737ce898fc355343d03c9d5c3fbb07c59edcc86db04cef5c30a1 +size 32613 diff --git a/parse/train/q8qLAbQBupm/images/a4a6832984d8a4bbe3172f132f82fbb2b1431c00ad430c891b2a2fca30baaa11.jpg b/parse/train/q8qLAbQBupm/images/a4a6832984d8a4bbe3172f132f82fbb2b1431c00ad430c891b2a2fca30baaa11.jpg new file mode 100644 index 0000000000000000000000000000000000000000..504662a505b963f774d794f6a5c19ec762d0d216 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/a4a6832984d8a4bbe3172f132f82fbb2b1431c00ad430c891b2a2fca30baaa11.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67aa317830d63cc75d8b11e27887420062e4da44c58d054e71de1ab90ef3075f +size 13930 diff --git a/parse/train/q8qLAbQBupm/images/a7825b5c981a98917b248f7299abd250a20f991643efb8fa61108765a3c1c993.jpg b/parse/train/q8qLAbQBupm/images/a7825b5c981a98917b248f7299abd250a20f991643efb8fa61108765a3c1c993.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c241257a9d85c4eb555dc8a41fcf046c841d01b6 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/a7825b5c981a98917b248f7299abd250a20f991643efb8fa61108765a3c1c993.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f33a40b0220a690a35e3cf5bf2b8723f5ac879d0ff7e14c7a7bffce4afe65379 +size 63637 diff --git a/parse/train/q8qLAbQBupm/images/a822b364407d49b727f5c41b82138bb4a562313e922cb590a931f0b2843b15af.jpg b/parse/train/q8qLAbQBupm/images/a822b364407d49b727f5c41b82138bb4a562313e922cb590a931f0b2843b15af.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1ca5456de7c4a01c74c81ccca3917a531b0c2d7f --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/a822b364407d49b727f5c41b82138bb4a562313e922cb590a931f0b2843b15af.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e06ff6cd1b4d501a937515502520d2e678bb9ee2decd3bbb2fa20e75ec48b10c +size 2475 diff --git a/parse/train/q8qLAbQBupm/images/a88ae49ef169a948a5dfed81811c394e87f0c5718af8a1e6b7116bd02f7d6797.jpg b/parse/train/q8qLAbQBupm/images/a88ae49ef169a948a5dfed81811c394e87f0c5718af8a1e6b7116bd02f7d6797.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ce342b07fed5dea626a15d38dbe590b94b9e471b --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/a88ae49ef169a948a5dfed81811c394e87f0c5718af8a1e6b7116bd02f7d6797.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbae01fc595a3b08ebe6175526d67b22c2f26a53ace7309c5ab0438bf934eac3 +size 14980 diff --git a/parse/train/q8qLAbQBupm/images/ae9e1c146426ca9e6eeee71489448eedc40ceba5f07a38bca79931bec0df3e87.jpg b/parse/train/q8qLAbQBupm/images/ae9e1c146426ca9e6eeee71489448eedc40ceba5f07a38bca79931bec0df3e87.jpg new file mode 100644 index 0000000000000000000000000000000000000000..995889e083d27466bdd62d43aaba3df153693e33 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/ae9e1c146426ca9e6eeee71489448eedc40ceba5f07a38bca79931bec0df3e87.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28c2bca50ad56835271ea41269c342df75119b7c8b152b44a427408658e33088 +size 25293 diff --git a/parse/train/q8qLAbQBupm/images/b2d145b8c0a301de5cf3519046c3058369e9ce9c5bdac459b226cef784e9d324.jpg b/parse/train/q8qLAbQBupm/images/b2d145b8c0a301de5cf3519046c3058369e9ce9c5bdac459b226cef784e9d324.jpg new file mode 100644 index 0000000000000000000000000000000000000000..986505d621ae9772b9c955b66099686d145ef313 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/b2d145b8c0a301de5cf3519046c3058369e9ce9c5bdac459b226cef784e9d324.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:834f110fd16343621a53b575135344739140ae8bc7b9d0e6c9cd40a68ac198c1 +size 72873 diff --git a/parse/train/q8qLAbQBupm/images/b3ddd761e30b1252304ea65d131020d8378ff3b65b857e40b686f2d99cb9ce5e.jpg b/parse/train/q8qLAbQBupm/images/b3ddd761e30b1252304ea65d131020d8378ff3b65b857e40b686f2d99cb9ce5e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a1cfefc0f69a6928242a85455a7cf5a1541b9dc7 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/b3ddd761e30b1252304ea65d131020d8378ff3b65b857e40b686f2d99cb9ce5e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a41b6c2e455a473e0627944f2a7617e78931e9df456f080aacbc6ffc8f106df3 +size 4739 diff --git a/parse/train/q8qLAbQBupm/images/b50c32733219cb794950a9615e14aac1fe917b56603d8426b8a5a9a7cb95b95b.jpg b/parse/train/q8qLAbQBupm/images/b50c32733219cb794950a9615e14aac1fe917b56603d8426b8a5a9a7cb95b95b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0c782cc72cb597dd432ed293bfc89b00766b134f --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/b50c32733219cb794950a9615e14aac1fe917b56603d8426b8a5a9a7cb95b95b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f58515f23721f9c3e759699a835b92b4bd72cfc0259927fc7525ae16e671552c +size 2750 diff --git a/parse/train/q8qLAbQBupm/images/ba4aea1c1c9dafac355afb8d6df2f70620d3ee1c3f55ad12f8ee27cd8f04c033.jpg b/parse/train/q8qLAbQBupm/images/ba4aea1c1c9dafac355afb8d6df2f70620d3ee1c3f55ad12f8ee27cd8f04c033.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8fdb296cf45b4cefe1067b75c883d6cd6d8c8441 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/ba4aea1c1c9dafac355afb8d6df2f70620d3ee1c3f55ad12f8ee27cd8f04c033.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c4f590ca7b68b97a53258239837c432c53dc07aa4f2278cb9ea847fe468e977 +size 10636 diff --git a/parse/train/q8qLAbQBupm/images/bd69829860d0cb4225437e12efc7d6e66e7baaf9efac1afe7519c06c171e1d97.jpg b/parse/train/q8qLAbQBupm/images/bd69829860d0cb4225437e12efc7d6e66e7baaf9efac1afe7519c06c171e1d97.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3244a16e31ff2200cdf5e27124ac1a98ab78737f --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/bd69829860d0cb4225437e12efc7d6e66e7baaf9efac1afe7519c06c171e1d97.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21c0c26cf6275ab39fc0470edae839497bed704b2383cd244d03b3973cc0992e +size 2830 diff --git a/parse/train/q8qLAbQBupm/images/ca4564fa2ed4a454310225579e35dcb73656358419a662c4d406fc9646d4d81f.jpg b/parse/train/q8qLAbQBupm/images/ca4564fa2ed4a454310225579e35dcb73656358419a662c4d406fc9646d4d81f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ecdfdd54727203ed6d66170ee7299085f45302d5 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/ca4564fa2ed4a454310225579e35dcb73656358419a662c4d406fc9646d4d81f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e343d659afb3d77d444af086270fafe94d5348322150ddefc6b1b6428dc593a7 +size 5477 diff --git a/parse/train/q8qLAbQBupm/images/ccc78643e58e4d811dd2239c218da36e20fe665c12a4df9ec8c067054dfff4cf.jpg b/parse/train/q8qLAbQBupm/images/ccc78643e58e4d811dd2239c218da36e20fe665c12a4df9ec8c067054dfff4cf.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d255e2b6d1de49d1e58bceb30da25b2c947eb4dd --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/ccc78643e58e4d811dd2239c218da36e20fe665c12a4df9ec8c067054dfff4cf.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efdc6965719cafff7fdbde09ddae908e22e113e12bf043090e7d2e12be727621 +size 36375 diff --git a/parse/train/q8qLAbQBupm/images/d23728938f24bacba24d85dedec0d546e4adc05e9df763102a87e3c7095755fe.jpg b/parse/train/q8qLAbQBupm/images/d23728938f24bacba24d85dedec0d546e4adc05e9df763102a87e3c7095755fe.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d4ef1a35814e970f2d93889dfc0d38719d1a3527 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/d23728938f24bacba24d85dedec0d546e4adc05e9df763102a87e3c7095755fe.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d9bf8babfc36ef87a68df75449eee848a2b7d80df4e542f96888731b75aee6a +size 3795 diff --git a/parse/train/q8qLAbQBupm/images/d3a4bed9bafcdc6e8f1b7aa91b343c34166d1ab7cb47f5af6c2764e8f890fdbe.jpg b/parse/train/q8qLAbQBupm/images/d3a4bed9bafcdc6e8f1b7aa91b343c34166d1ab7cb47f5af6c2764e8f890fdbe.jpg new file mode 100644 index 0000000000000000000000000000000000000000..73a46186136e96813f09824543fcb1130b6e2d0c --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/d3a4bed9bafcdc6e8f1b7aa91b343c34166d1ab7cb47f5af6c2764e8f890fdbe.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ef864aa30845c7060b361b163dbb5b10db1700df28eb3e3fc3f4228543635a7 +size 84313 diff --git a/parse/train/q8qLAbQBupm/images/d5371c7a6bb4b92c55f67d7c686c33a9539bfb37857d8e1253b5e6095fe37723.jpg b/parse/train/q8qLAbQBupm/images/d5371c7a6bb4b92c55f67d7c686c33a9539bfb37857d8e1253b5e6095fe37723.jpg new file mode 100644 index 0000000000000000000000000000000000000000..93c5d3d09f4f70b8f56b45e8c253a82c5f29f3f8 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/d5371c7a6bb4b92c55f67d7c686c33a9539bfb37857d8e1253b5e6095fe37723.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54e38d919e34816c64b8bcaceafee3cb7893e02c329a27917004c385f3adc7a9 +size 3410 diff --git a/parse/train/q8qLAbQBupm/images/d83ad82b94e29395c41d4c3396621611edc963a11d46bc534817dbeeeae95c8e.jpg b/parse/train/q8qLAbQBupm/images/d83ad82b94e29395c41d4c3396621611edc963a11d46bc534817dbeeeae95c8e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5960958996cb708ae161990b2c14eca6a2f20957 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/d83ad82b94e29395c41d4c3396621611edc963a11d46bc534817dbeeeae95c8e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a68b870091bbec9fd22f1e2c326f336f8bd8ea21b1a8f42a546c25f350cc073 +size 9282 diff --git a/parse/train/q8qLAbQBupm/images/da86cbf37c93427b3aafeeb030f5bd59b04058994a3b5daad3f311c861708977.jpg b/parse/train/q8qLAbQBupm/images/da86cbf37c93427b3aafeeb030f5bd59b04058994a3b5daad3f311c861708977.jpg new file mode 100644 index 0000000000000000000000000000000000000000..39af5684824eee07869c0d01b0a5d5818a9e6fe0 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/da86cbf37c93427b3aafeeb030f5bd59b04058994a3b5daad3f311c861708977.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9256ac86b06f05ec72172515ba19b6e5c78933ba6161d303c5971e7225e6e44a +size 2515 diff --git a/parse/train/q8qLAbQBupm/images/dd3e5673a37e9987f8a05e3e9fcf6c2646ac9eda91632f6e2f290391669f64e1.jpg b/parse/train/q8qLAbQBupm/images/dd3e5673a37e9987f8a05e3e9fcf6c2646ac9eda91632f6e2f290391669f64e1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b013b027105dfcf7a6383c391fb59f36158a22fc --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/dd3e5673a37e9987f8a05e3e9fcf6c2646ac9eda91632f6e2f290391669f64e1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c34e9d842672583b658f609839ec84213ba0168969484971346ac770691a347d +size 8050 diff --git a/parse/train/q8qLAbQBupm/images/dda9caf6519b7e83f8b483b6207566d3ad0308285bea66bd53f7cbf31880c802.jpg b/parse/train/q8qLAbQBupm/images/dda9caf6519b7e83f8b483b6207566d3ad0308285bea66bd53f7cbf31880c802.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a5a40ac4b5a1b50b9e533dd2fe6823bb553acf7a --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/dda9caf6519b7e83f8b483b6207566d3ad0308285bea66bd53f7cbf31880c802.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e97a457fd1a1cf202bf233f6e010ed860bb07e28f7c68fd5dde6264017efdd4 +size 3206 diff --git a/parse/train/q8qLAbQBupm/images/ddbe3b273c2a9f1810b4b6ba806c7a9a91359efa94090c191ed6dadfa95faedf.jpg b/parse/train/q8qLAbQBupm/images/ddbe3b273c2a9f1810b4b6ba806c7a9a91359efa94090c191ed6dadfa95faedf.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1966c762d594f71131fa47c801e30a9b6bcc23e2 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/ddbe3b273c2a9f1810b4b6ba806c7a9a91359efa94090c191ed6dadfa95faedf.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaaa3699470260738af6bcc73de4fd4039beb7ebbe812e251946765c897b7c41 +size 3156 diff --git a/parse/train/q8qLAbQBupm/images/e02ca351f61abdda4028cb61c0ba1fd3bc51e3768b7f532ab590a9e51226f07d.jpg b/parse/train/q8qLAbQBupm/images/e02ca351f61abdda4028cb61c0ba1fd3bc51e3768b7f532ab590a9e51226f07d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fd553e46072a7a796c90da366741c29a2969acac --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/e02ca351f61abdda4028cb61c0ba1fd3bc51e3768b7f532ab590a9e51226f07d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bbe4d1e3c000812c7ccff186d4665a77566cf4b3716766ce1961ce0a7b5a38b +size 33862 diff --git a/parse/train/q8qLAbQBupm/images/e3aa0ef4faef24943b5e905bbda00fcf2c2d2113939e99cd6d12ce4cbd6f405b.jpg b/parse/train/q8qLAbQBupm/images/e3aa0ef4faef24943b5e905bbda00fcf2c2d2113939e99cd6d12ce4cbd6f405b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2e21031fa89abf303ed89c67a1674555af3efc6a --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/e3aa0ef4faef24943b5e905bbda00fcf2c2d2113939e99cd6d12ce4cbd6f405b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c661582ce554f04768a0181b549c66d832f5fe03018fbac12c344613d58e4cd4 +size 1869 diff --git a/parse/train/q8qLAbQBupm/images/e47ad567a5cffda6736aa10c6576737ff360360e397b69400c8e15c0d55f6a88.jpg b/parse/train/q8qLAbQBupm/images/e47ad567a5cffda6736aa10c6576737ff360360e397b69400c8e15c0d55f6a88.jpg new file mode 100644 index 0000000000000000000000000000000000000000..442e6b1eb1ffc7c3f2ba2d6b584ef3002f628722 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/e47ad567a5cffda6736aa10c6576737ff360360e397b69400c8e15c0d55f6a88.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40d85a51e9aaac3c13af6ab6afb2bfccdf2cbd6cb312ebe55e1458876097ded5 +size 6402 diff --git a/parse/train/q8qLAbQBupm/images/e49f684b022f3a7d4b2281e2f2ca06cb10e4190920b8c101473c3edaaae1e62c.jpg b/parse/train/q8qLAbQBupm/images/e49f684b022f3a7d4b2281e2f2ca06cb10e4190920b8c101473c3edaaae1e62c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..db9422036243a87bafe42b39f888f53f3a2418d1 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/e49f684b022f3a7d4b2281e2f2ca06cb10e4190920b8c101473c3edaaae1e62c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1afab87dc98b1233f5df9f6d47951672f5d98153196b3cb9c3b2ae46989e5a2 +size 6733 diff --git a/parse/train/q8qLAbQBupm/images/e8e4365cc9ac33a587e9fb19754917818649c0e2f83d6151fb320f069f5e7aea.jpg b/parse/train/q8qLAbQBupm/images/e8e4365cc9ac33a587e9fb19754917818649c0e2f83d6151fb320f069f5e7aea.jpg new file mode 100644 index 0000000000000000000000000000000000000000..efcd3b47645771ed5dbc1344563ea9954d2eea6d --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/e8e4365cc9ac33a587e9fb19754917818649c0e2f83d6151fb320f069f5e7aea.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f846b6d5e5c3aa3cb39184be4161a03204246206e5344204d1c708ea734d676 +size 2025 diff --git a/parse/train/q8qLAbQBupm/images/eca701549804a82d6b30851bd162ed57643c34231dcb4ea78f08941ea7bb1a6d.jpg b/parse/train/q8qLAbQBupm/images/eca701549804a82d6b30851bd162ed57643c34231dcb4ea78f08941ea7bb1a6d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..55c3cf3172479c9cf403d67e331b4aa10a5a1a6d --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/eca701549804a82d6b30851bd162ed57643c34231dcb4ea78f08941ea7bb1a6d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f61dd45b302605ec187c999a9c345479e0fdaffe4b349ff06e5452aad046b469 +size 22396 diff --git a/parse/train/q8qLAbQBupm/images/eea00faf72a63f1b3541462c4e6f213e72fa2dc7d43b53933c2505ba81d6e005.jpg b/parse/train/q8qLAbQBupm/images/eea00faf72a63f1b3541462c4e6f213e72fa2dc7d43b53933c2505ba81d6e005.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9e4fc977de07b8952b3da4aa6fbb8af90b86d5d5 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/eea00faf72a63f1b3541462c4e6f213e72fa2dc7d43b53933c2505ba81d6e005.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e89b9da75151c5a37d297f95d3ca8ab987d057d090f0babade690f2ac7c8900 +size 1643 diff --git a/parse/train/q8qLAbQBupm/images/f0ed54df1c11b4b6b2634ef9649ffd0d2b9c35d190acbf34881ca712c49481b2.jpg b/parse/train/q8qLAbQBupm/images/f0ed54df1c11b4b6b2634ef9649ffd0d2b9c35d190acbf34881ca712c49481b2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7845bb0003e3150a195c01aa9261a49bcc5718a6 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/f0ed54df1c11b4b6b2634ef9649ffd0d2b9c35d190acbf34881ca712c49481b2.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51583720d3bc755da1ba0b0659f81b897bf37bb90cf64dca87a1d88b026ce9e9 +size 1918 diff --git a/parse/train/q8qLAbQBupm/images/f555f11943f4c8a92be2676315684ef5989409f273bda9eff44be3a4819eff26.jpg b/parse/train/q8qLAbQBupm/images/f555f11943f4c8a92be2676315684ef5989409f273bda9eff44be3a4819eff26.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4cff418b8ef0183581737ab758b108715acadb5d --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/f555f11943f4c8a92be2676315684ef5989409f273bda9eff44be3a4819eff26.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88b96b06301a9a2c5d8df6ca3c1a4db33f874417fd851fec0743043605ef0431 +size 6665 diff --git a/parse/train/q8qLAbQBupm/images/fd164f9e1d178d7307816c0623a51df6a99af32d28b14bce99a8f43ebff975b0.jpg b/parse/train/q8qLAbQBupm/images/fd164f9e1d178d7307816c0623a51df6a99af32d28b14bce99a8f43ebff975b0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..cf6975fc9e0285712b4bde4bdb7b7180dcdee566 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/fd164f9e1d178d7307816c0623a51df6a99af32d28b14bce99a8f43ebff975b0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffb7c8acaed78c1a9d014eaaf644c7d913fbb83a4b41366e5c095b5bdb4e96c8 +size 4344 diff --git a/parse/train/q8qLAbQBupm/images/feaf0fb1972f67021f766d5b2b9bd7f94cdf4d08f4728dea6cd46d8f8e0cfc63.jpg b/parse/train/q8qLAbQBupm/images/feaf0fb1972f67021f766d5b2b9bd7f94cdf4d08f4728dea6cd46d8f8e0cfc63.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4f2e60ec93eae83425444f5d47eaa61ba1c276a9 --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/feaf0fb1972f67021f766d5b2b9bd7f94cdf4d08f4728dea6cd46d8f8e0cfc63.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26ffe252bcac32c18daf975f86b164e739abad280af4eab02ee47779c9858224 +size 31813 diff --git a/parse/train/q8qLAbQBupm/images/fee33fcd973957112fc46af30d9cfb092b86b79028e3ef7dcc0599362fa68707.jpg b/parse/train/q8qLAbQBupm/images/fee33fcd973957112fc46af30d9cfb092b86b79028e3ef7dcc0599362fa68707.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ca8a56aae375c3aa803501a2e21d821c5eeea9bd --- /dev/null +++ b/parse/train/q8qLAbQBupm/images/fee33fcd973957112fc46af30d9cfb092b86b79028e3ef7dcc0599362fa68707.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03ff5ea345e48cf057feb7f37659763c5cc7522e0b477f213364a072455c4633 +size 13693 diff --git a/parse/train/r1lZgyBYwS/images/0aaca877eccf26d2edd08745b3eabab37ccddddb8c673f0d8034be6cdb262052.jpg b/parse/train/r1lZgyBYwS/images/0aaca877eccf26d2edd08745b3eabab37ccddddb8c673f0d8034be6cdb262052.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4f8ad1d63aae38d0e233208fd5e1dc791038cf00 --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/0aaca877eccf26d2edd08745b3eabab37ccddddb8c673f0d8034be6cdb262052.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:945d5066fcc2471972dd49acf55400dfa14bd6db8d5495cd53a031cdb88b7f55 +size 17095 diff --git a/parse/train/r1lZgyBYwS/images/2a171c4cacd067f3dc6529811906820167037ee379e555c3a1d5d6b8e0541aef.jpg b/parse/train/r1lZgyBYwS/images/2a171c4cacd067f3dc6529811906820167037ee379e555c3a1d5d6b8e0541aef.jpg new file mode 100644 index 0000000000000000000000000000000000000000..434021ca8fee01d2265fdf179fa3a3c83b8eefb2 --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/2a171c4cacd067f3dc6529811906820167037ee379e555c3a1d5d6b8e0541aef.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f17dbcaae2b2bc64c3fa483b6a78e5762d96853401843ca2982adb145ab3d17 +size 29387 diff --git a/parse/train/r1lZgyBYwS/images/2b0b5b25d26919e5317114cf05c798a3d992c7935070155bc8756595fd201024.jpg b/parse/train/r1lZgyBYwS/images/2b0b5b25d26919e5317114cf05c798a3d992c7935070155bc8756595fd201024.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e54ca88ce33eceb0aa0c964dd97d1055af70ccdc --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/2b0b5b25d26919e5317114cf05c798a3d992c7935070155bc8756595fd201024.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d0aa55653465bf7c01ff4b545535d92b4ecacbfdfde07fbef368679e1338176 +size 2827 diff --git a/parse/train/r1lZgyBYwS/images/5cd683ca8e3e562acc905b2b6e32b4270672593d0e1630e8b6ad4f7b468c065f.jpg b/parse/train/r1lZgyBYwS/images/5cd683ca8e3e562acc905b2b6e32b4270672593d0e1630e8b6ad4f7b468c065f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..13688f40d8c3e5f1e763766826198146ed4dc541 --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/5cd683ca8e3e562acc905b2b6e32b4270672593d0e1630e8b6ad4f7b468c065f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dcc10363da4fe77def1151398f6fb4a97051bcfa9bebcce7d71717856832f83 +size 34900 diff --git a/parse/train/r1lZgyBYwS/images/8d507ac75f70f94acfd1a8d4c2b13c2fc7465be89c2079a82570a9d6fb4824a8.jpg b/parse/train/r1lZgyBYwS/images/8d507ac75f70f94acfd1a8d4c2b13c2fc7465be89c2079a82570a9d6fb4824a8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1839b17df21152971faef419d56e9088444885e5 --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/8d507ac75f70f94acfd1a8d4c2b13c2fc7465be89c2079a82570a9d6fb4824a8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46301fcd83f75b514b5ccae64db40ed9ae3515be8b7274fcc6d60d2cebab6893 +size 5807 diff --git a/parse/train/r1lZgyBYwS/images/af251455bc46108914f08e63940c79dcf3f8bb2a901c5916f9ba5dcbd94486b7.jpg b/parse/train/r1lZgyBYwS/images/af251455bc46108914f08e63940c79dcf3f8bb2a901c5916f9ba5dcbd94486b7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6466d10782a05c33fbe627e353c5c7d8d69b196d --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/af251455bc46108914f08e63940c79dcf3f8bb2a901c5916f9ba5dcbd94486b7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15c98471c80d4fc0a8051db46821723fec7f041fd602f03e3f83941033c34c1b +size 12252 diff --git a/parse/train/r1lZgyBYwS/images/d811f7394d37690939320dd1f86c83340301c13d6ab564158a03371160460f0f.jpg b/parse/train/r1lZgyBYwS/images/d811f7394d37690939320dd1f86c83340301c13d6ab564158a03371160460f0f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b21f0e117dd145b78bc70d96e728cc78443559f8 --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/d811f7394d37690939320dd1f86c83340301c13d6ab564158a03371160460f0f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6ba9f274c6f0ad81980f599af9035641fffb2a2e66c468b5c4223bc3194302b +size 5957 diff --git a/parse/train/r1lZgyBYwS/images/d8f940b7e74ac9f1fe5b736894bfc2f97b012b9fbaa5763e634bbd3a5be36641.jpg b/parse/train/r1lZgyBYwS/images/d8f940b7e74ac9f1fe5b736894bfc2f97b012b9fbaa5763e634bbd3a5be36641.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d0e92748277fd58d6e94ed75c499913c2557617a --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/d8f940b7e74ac9f1fe5b736894bfc2f97b012b9fbaa5763e634bbd3a5be36641.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a2c29b99b531300f75ce34428827bc011b17e40cdd558a041e3ef3c51bffb66 +size 5123 diff --git a/parse/train/r1lZgyBYwS/images/daa2c33557f0fe0183151d1aed6ae06194e34722cb6cc2450b910b98c5a309d6.jpg b/parse/train/r1lZgyBYwS/images/daa2c33557f0fe0183151d1aed6ae06194e34722cb6cc2450b910b98c5a309d6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6503d3cd9e38636a17434681ce40a64bf2c8691f --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/daa2c33557f0fe0183151d1aed6ae06194e34722cb6cc2450b910b98c5a309d6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdb376ce321adaea9512008308d07654105f206ffdd5559a7d4e7311bc501b02 +size 38847 diff --git a/parse/train/r1lZgyBYwS/images/dee37a7ae0105f2229b12fcf719a1a550610d641c77804ceb792d63f96869848.jpg b/parse/train/r1lZgyBYwS/images/dee37a7ae0105f2229b12fcf719a1a550610d641c77804ceb792d63f96869848.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0fd5a11c6df66967de05b42b8b471544fee6ca45 --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/dee37a7ae0105f2229b12fcf719a1a550610d641c77804ceb792d63f96869848.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a058610b6fa0f873107adfca3d59a717fc73ffb6b30579e5457d527d77ff542d +size 15317 diff --git a/parse/train/r1lZgyBYwS/images/e701efaca8f376e98297702c0d257202ff41d19521310727b1e4ddda5a54bd55.jpg b/parse/train/r1lZgyBYwS/images/e701efaca8f376e98297702c0d257202ff41d19521310727b1e4ddda5a54bd55.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fdbc4f414ebb8fc7f956eb64396e0d6673947ae3 --- /dev/null +++ b/parse/train/r1lZgyBYwS/images/e701efaca8f376e98297702c0d257202ff41d19521310727b1e4ddda5a54bd55.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1aaf85416e4012a64aeba382379569eb9eccc7594c7d158c8ee524bd5a26dcd0 +size 4755