diff --git a/.gitattributes b/.gitattributes index 1ef325f1b111266a6b26e0196871bd78baa8c2f3..56343214f0a05a669391db00db9f1774a27e8e82 100644 --- a/.gitattributes +++ b/.gitattributes @@ -57,3 +57,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text # Video files - compressed *.mp4 filter=lfs diff=lfs merge=lfs -text *.webm filter=lfs diff=lfs merge=lfs -text +papers.jsonl filter=lfs diff=lfs merge=lfs -text +train/citations.jsonl filter=lfs diff=lfs merge=lfs -text diff --git a/paper_embeddings.parquet b/paper_embeddings.parquet new file mode 100644 index 0000000000000000000000000000000000000000..95a39cdeb99bfca696a0e6dba44b76eabbf0260f --- /dev/null +++ b/paper_embeddings.parquet @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0acc3e090cd2c5a36c900c83aaeede22afc7d0ae61e251b30be94b88715d970 +size 11433019 diff --git a/papers.jsonl b/papers.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9873a1602d1a2d359e808eac5fb0fcabdc3cae9f --- /dev/null +++ b/papers.jsonl @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5839c614b459022a703bb7d6064160f90c4f9e4609d65b761b70f0356c87a375 +size 143458210 diff --git a/test/citation_embeddings_0.parquet b/test/citation_embeddings_0.parquet new file mode 100644 index 0000000000000000000000000000000000000000..2d07bb426e959b1170b60d53694c5e8a9cc9ae28 --- /dev/null +++ b/test/citation_embeddings_0.parquet @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad6fc4900112b2a848e974b00addfb2f4811d54e623230af3ddaa8c385db033e +size 983365934 diff --git a/test/citation_embeddings_5000.parquet b/test/citation_embeddings_5000.parquet new file mode 100644 index 0000000000000000000000000000000000000000..5c45d653534c59ac89f5ed8a2f9080dc91e0d421 --- /dev/null +++ b/test/citation_embeddings_5000.parquet @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6afdccff109e549c70da527210c78629290f9799de589ae7501fdada15ebbf2 +size 613566225 diff --git a/test/citations.jsonl b/test/citations.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..c508c6f7ddbc48e5a9af11dea82ef291a58885f0 --- /dev/null +++ b/test/citations.jsonl @@ -0,0 +1,8042 @@ +{"index":0,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2209.14491","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2505.14683","reference_contexts":"We tackle both challenges using latent adversarial diffusion distillation (LADD) [47, 48, 49] , reducing the number of sampling steps while increasing the quality of the samples through adversarial training.","reference_id":"MTg0MzI0NjY5MjMwMzc0MDkwNzM="} +{"index":2,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2505.14683","reference_contexts":"The advent of large-scale generative models has changed this landscape, enabling purely text-driven image synthesis and modifications that were previously impractical or impossible [11, 17, 41, 42, 40, 2, 12, 49, 21] .","reference_id":"MTQyODUxMTk3Mjk3NDk4MDA4NDI="} +{"index":3,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2403.03206","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2208.01618","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2410.23775","reference_contexts":"[20] introduce in-context LoRAs for diffusion transformers on specific tasks, where each task needs to train dedicated LoRA weights","reference_id":"MTIyMzU2ODMzMDcyOTYxNDQ2Mg=="} +{"index":6,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2210.09276","reference_contexts":"InstructPix2Pix [5] and subsequent work [4] demonstrated the promise of synthetic instruction-response pairs for fine-tuning a diffusion model for image editing, while learning-free methods for personalized text-to-image synthesis [13, 43, 24] enable image modification with off-theshelf, high-performance image generation models [42, 40]","reference_id":"ODQ2ODY4NTQzODk4OTU2MDQw"} +{"index":7,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2406.13743","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":8,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2410.06511","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":9,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2504.17761","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":10,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2406.16855","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":11,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2208.12242","reference_contexts":"InstructPix2Pix [5] and subsequent work [4] demonstrated the promise of synthetic instruction-response pairs for fine-tuning a diffusion model for image editing, while learning-free methods for personalized text-to-image synthesis [13, 43, 24] enable image modification with off-theshelf, high-performance image generation models [42, 40]","reference_id":"ODQ2ODY4NTQzODk4OTU2MDQw"} +{"index":12,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2311.17042","reference_contexts":"We tackle both challenges using latent adversarial diffusion distillation (LADD) [47, 48, 49] , reducing the number of sampling steps while increasing the quality of the samples through adversarial training.","reference_id":"MTg0MzI0NjY5MjMwMzc0MDkwNzM="} +{"index":13,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2311.17042","reference_contexts":"FLUX.1 Kontext [pro] is trained with the flow objective followed by LADD [48]","reference_id":"OTUxNTczODkzMDA5MDk5MTk0NA=="} +{"index":14,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2403.12015","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":15,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2311.10089","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":16,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"1409.1556","reference_contexts":"We compare our VAE with related models using three reconstruction metrics, namely, SSIM, PSNR, and the Perceptual Distance (PDist) in VGG [52] feature space","reference_id":"MzczOTAxODg2NTI3MTU2NDA5NQ=="} +{"index":17,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2409.11340","reference_contexts":"Subsequent instruction-driven editors such as Emu Edit [51], OmniGen [56] , HiDream-E1 [15] and ICEdit [62] -extend these ideas to refined datasets and model architectures","reference_id":"MTExNzA2MTMxMTU1ODEzMzY4MzQ="} +{"index":18,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2308.06721","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":19,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2206.10789","reference_contexts":"We evaluate on 1 000 diverse test prompts compiled from academic benchmarks (DrawBench [46] , PartiPrompts [59] ) and real user queries","reference_id":"MTczMDMyODIwMjc5NTc1NDA4Mzg="} +{"index":20,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2306.10012","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":21,"citer_arxiv_id":"2506.15742","cited_arxiv_id":"2504.20690","reference_contexts":"Subsequent instruction-driven editors such as Emu Edit [51], OmniGen [56] , HiDream-E1 [15] and ICEdit [62] -extend these ideas to refined datasets and model architectures","reference_id":"MTExNzA2MTMxMTU1ODEzMzY4MzQ="} +{"index":22,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"1703.06975","reference_contexts":"Score-based generative models, and related techniques (Bordes et al., 2017; Goyal et al., 2017; Du & Mordatch, 2019) , have proven effective at generation of images (Song & Ermon, 2019; 2020; Ho et al., 2020) , audio (Chen et al., 2020; Kong et al., 2020) , graphs (Niu et al., 2020) , and shapes (Cai ˚Work partially done during an internship at Google Brain","reference_id":"NTg5NDI2NjIzNDU2Njk0MzE1Ng=="} +{"index":23,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"2009.00713","reference_contexts":"Score-based generative models, and related techniques (Bordes et al., 2017; Goyal et al., 2017; Du & Mordatch, 2019) , have proven effective at generation of images (Song & Ermon, 2019; 2020; Ho et al., 2020) , audio (Chen et al., 2020; Kong et al., 2020) , graphs (Niu et al., 2020) , and shapes (Cai ˚Work partially done during an internship at Google Brain","reference_id":"NTg5NDI2NjIzNDU2Njk0MzE1Ng=="} +{"index":24,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"1605.08803","reference_contexts":"RealNVP (Dinh et al., 2016) 3.49 -iResNet (Behrmann et al., 2019) 3.45 -Glow (Kingma & Dhariwal, 2018) 3.35 -MintNet (Song et al., 2019b) 3.32 -Residual Flow (Chen et al., 2019) 3.28 46.37 FFJORD (Grathwohl et al., 2018) 3.40 -Flow++ (Ho et al., 2019) 3.29 -DDPM (L) (Ho et al., 2020) ď 3.70 * 13.51 DDPM (Lsimple) (Ho et al., 2020) ď 3.75 * 3.17 3.28 3.37 DDPM cont","reference_id":"MTQ5MjMxODAxODU5NDc5ODkxMTk="} +{"index":25,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"1605.08803","reference_contexts":"As is done with other invertible models such as neural ODEs and normalizing flows (Dinh et al., 2016; Kingma & Dhariwal, 2018) , we can manipulate this latent representation for image editing, such as interpolation, and temperature scaling (see Fig","reference_id":"NDI4MTY2NzIwMDY3MTAwODI4Nw=="} +{"index":26,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"2009.05475","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":27,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"2009.09761","reference_contexts":"Score-based generative models, and related techniques (Bordes et al., 2017; Goyal et al., 2017; Du & Mordatch, 2019) , have proven effective at generation of images (Song & Ermon, 2019; 2020; Ho et al., 2020) , audio (Chen et al., 2020; Kong et al., 2020) , graphs (Niu et al., 2020) , and shapes (Cai ˚Work partially done during an internship at Google Brain","reference_id":"NTg5NDI2NjIzNDU2Njk0MzE1Ng=="} +{"index":28,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"2006.00702","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":29,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"2007.03317","reference_contexts":"( 7 ) uses denoising score matching, but other score matching objectives, such as sliced score matching (Song et al., 2019a) and finite-difference score matching (Pang et al., 2020) are also applicable here.","reference_id":"MTY2ODk1NjQ3MjQzOTU3MjcwOA=="} +{"index":30,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"2007.00810","reference_contexts":"Uniquely identifiable encoding Unlike most current invertible models, our encoding is uniquely identifiable, meaning that with sufficient training data, model capacity, and optimization accuracy, the encoding for an input is uniquely determined by the data distribution (Roeder et al., 2020)","reference_id":"NDU4ODczNzA4Mjc3NzgyODI5Mg=="} +{"index":31,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"1506.03365","reference_contexts":"VE SDE (SMLD) VP SDE (DDPM) Predictor r Sampler P1000 P2000 C2000 PC1000 P1000 P2000 C2000 PC1000 ancestral sampling --0.17 --0.01 reverse diffusion --0.16 --0.01 probability flow --0.22 0.17 --0.27 0.04 256 ˆ256 LSUN bedroom and church outdoor (Yu et al., 2015) datasets with the VE SDE and our continuous objective Eq","reference_id":"MTcyMzE3MzU0MTAzMDIwMjMzNDI="} +{"index":32,"citer_arxiv_id":"2011.13456","cited_arxiv_id":"1605.07146","reference_contexts":"To test this idea, we trained a Wide ResNet (Zagoruyko & Komodakis, 2016 ) (Wide-ResNet-28-10) on CIFAR-10 with VE perturbations","reference_id":"OTk0NzgwNDY2MTk0MTE3MzUwMQ=="} +{"index":33,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1711.00482","reference_contexts":"This process can be seen as a kind of language-guided search (Andreas et al., 2017) , and is also similar to the auxiliary text-image matching loss proposed by Xu et al","reference_id":"MTk1OTczNTY4OTc3NjEyNTM0NQ=="} +{"index":34,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1308.3432","reference_contexts":"(2019) address this using an online cluster assignment procedure coupled with the straight-through estimator (Bengio et al., 2013)","reference_id":"MTAxODk2MTU4MDUzMDY2OTE2NDE="} +{"index":35,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1511.06349","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":36,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1904.10509","reference_contexts":"Here, p ψ is represented by a 12-billion parameter sparse transformer (Child et al., 2019) .","reference_id":"MTc4MDc5OTMxNTMxMDExMDkxNTQ="} +{"index":37,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"2009.11278","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":38,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"2005.00341","reference_contexts":"Specifically, when compute, model size, and data are scaled carefully, autoregressive transformers (Vaswani et al., 2017) have achieved impressive results in several domains such as text (Radford et al., 2019) , images (Chen et al., 2020) , and audio (Dhariwal et al., 2020) .","reference_id":"MTM0MzI5MDQ2ODk5MDU1OTQyNzU="} +{"index":39,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1406.2661","reference_contexts":"(2016b) later demonstrated that using a generative adversarial network (Goodfellow et al., 2014) , rather than a recurrent variational auto-encoder, improved image fidelity","reference_id":"MTM1NTIyNzE5NTU3ODk1OTQyOTA="} +{"index":40,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"2012.05208","reference_contexts":"Prompts like the latter require the model to perform variable binding (Smolensky, 1990; Greff et al., 2020) -it is the hedgehog that is in the christmas sweater, not the dog","reference_id":"Nzk3MzY5NTgyODg2NzQ4NDI0OQ=="} +{"index":41,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1706.08500","reference_contexts":"We evaluate our model zero-shot by comparing it to three prior approaches: AttnGAN (Xu et al., 2018) , DM-GAN (Zhu et al., 2019) , and DF-GAN (Tao et al., 2020) , the last of which reports the best Inception Score (Salimans et al., 2016) and Fréchet Inception Distance (Heusel et al., 2017) on MS-COCO","reference_id":"NTc1NDEzODA2NDgyMjc5NjAw"} +{"index":42,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1611.01144","reference_contexts":"We instead use the gumbel-softmax relaxation (Jang et al., 2016; Maddison et al., 2016) , replacing the expectation over q φ with one over q τ φ , where the relaxation becomes tight as the temperature τ → 0","reference_id":"MTgzODEzOTM3NzQ3MjY4OTY5ODA="} +{"index":43,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1412.6980","reference_contexts":"The relaxed ELB is maximized using Adam (Kingma & Ba, 2014) with exponentially weighted iterate averaging","reference_id":"ODczMDAzNjA0Mjc0MzE3MDk0NQ=="} +{"index":44,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1312.6114","reference_contexts":"The overall procedure can be viewed as maximizing the evidence lower bound (ELB) (Kingma & Welling, 2013; Rezende et al., 2014) on the joint likelihood of the model distribution over images x, captions y, and the tokens z for the encoded RGB image","reference_id":"OTAyODI0MzI5NzA1NzM3MzMyNA=="} +{"index":45,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1711.02213","reference_contexts":"This can be seen as a practical alternative to a more general framework for mixed-precision training called Flexpoint (Köster et al., 2017) , with the advantage that specialized GPU kernels are not required","reference_id":"MTU0MDg1NjkwNDE2MDg4MDE0NTM="} +{"index":46,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"2004.08249","reference_contexts":"It is possible that better initialization schemes (Liu et al., 2020 ) might be able to avoid this, but we did not have success with alternative schemes in our experiments.","reference_id":"MTA4MzYyMDE2NzQ4NzM0NDk2ODk="} +{"index":47,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"2004.08249","reference_contexts":"Similar to prior work (Liu et al., 2020) , we found that the norms of the activation gradients from the resblocks decrease monotonically as we move from the earlier resblocks Table 1","reference_id":"MTA1NDY2MTEwNTU1MTU1MDE5NjI="} +{"index":48,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1711.05101","reference_contexts":"We update the parameters using AdamW (Loshchilov & Hutter, 2017) with β 1 = 0.9, β 2 = 0.999, = 10 -8 , and weight decay multiplier 10 -4","reference_id":"Mzg1MTQ4OTExMDAwMTQ5MjI1OQ=="} +{"index":49,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1611.00712","reference_contexts":"We instead use the gumbel-softmax relaxation (Jang et al., 2016; Maddison et al., 2016) , replacing the expectation over q φ with one over q τ φ , where the relaxation becomes tight as the temperature τ → 0","reference_id":"MTgzODEzOTM3NzQ3MjY4OTY5ODA="} +{"index":50,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1511.02793","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":51,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1710.03740","reference_contexts":"Standard loss scaling (Micikevicius et al., 2017) is able to avoid underflow when the range spanned by the smallest and largest activation gradients (in absolute value) fits within the exponent range of the 16-bit format","reference_id":"MzM3NzIyNDI1ODcxMTA4NDAyMw=="} +{"index":52,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1711.00937","reference_contexts":"We address these issues by using a two-stage training procedure, similar to (Oord et al., 2017; Razavi et al., 2019) : • Stage 1","reference_id":"MzI2NTM1ODYzMzQ1Mjc0MTEyOA=="} +{"index":53,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1910.13267","reference_contexts":"5 During training, we apply 10% BPE dropout (Provilkov et al., 2019) , whose use is common in the neural machine translation literature.","reference_id":"MTIyODkzMDE1NDM3NTgzMzAzMTE="} +{"index":54,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1910.02054","reference_contexts":"We address this using parameter sharding (Rajbhandari et al., 2019)","reference_id":"MTM2Mzc5ODkwNTc3MDE2Nzk3Mjk="} +{"index":55,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1906.00446","reference_contexts":"We address these issues by using a two-stage training procedure, similar to (Oord et al., 2017; Razavi et al., 2019) : • Stage 1","reference_id":"MzI2NTM1ODYzMzQ1Mjc0MTEyOA=="} +{"index":56,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1610.02454","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":57,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1606.03498","reference_contexts":"We evaluate our model zero-shot by comparing it to three prior approaches: AttnGAN (Xu et al., 2018) , DM-GAN (Zhu et al., 2019) , and DF-GAN (Tao et al., 2020) , the last of which reports the best Inception Score (Salimans et al., 2016) and Fréchet Inception Distance (Heusel et al., 2017) on MS-COCO","reference_id":"NTc1NDEzODA2NDgyMjc5NjAw"} +{"index":58,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1701.05517","reference_contexts":"Likelihood objectives tend to prioritize modeling short-range dependencies between pixels (Salimans et al., 2017) , so much of the modeling capacity would be spent capturing high-frequency details instead of the low-frequency structure that makes objects visually recognizable to us.","reference_id":"OTY1ODY0OTQzNTk3MzgwNjkyMg=="} +{"index":59,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1508.07909","reference_contexts":"Given a text-image pair, we BPE-encode (Sennrich et al., 2015) the lowercased caption using at most 256 tokens 5 with vocabulary size 16,384, and encode the image using 32 × 32 = 1024 tokens with vocabulary size 8192","reference_id":"NjIwNzEwMjA0MDYxMTYxNjAyMA=="} +{"index":60,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"2008.05865","reference_contexts":"We evaluate our model zero-shot by comparing it to three prior approaches: AttnGAN (Xu et al., 2018) , DM-GAN (Zhu et al., 2019) , and DF-GAN (Tao et al., 2020) , the last of which reports the best Inception Score (Salimans et al., 2016) and Fréchet Inception Distance (Heusel et al., 2017) on MS-COCO","reference_id":"NTc1NDEzODA2NDgyMjc5NjAw"} +{"index":61,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1706.03762","reference_contexts":"Specifically, when compute, model size, and data are scaled carefully, autoregressive transformers (Vaswani et al., 2017) have achieved impressive results in several domains such as text (Radford et al., 2019) , images (Chen et al., 2020) , and audio (Dhariwal et al., 2020) .","reference_id":"MTM0MzI5MDQ2ODk5MDU1OTQyNzU="} +{"index":62,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1706.03762","reference_contexts":"Our goal is to train a transformer (Vaswani et al., 2017) to autoregressively model the text and image tokens as a single stream of data","reference_id":"NzQ0NTU2NzczNjMzMzM2OTAxMQ=="} +{"index":63,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1905.13727","reference_contexts":"We were able to drastically reduce this cost by compressing the gradients using PowerSGD (Vogels et al., 2019) .","reference_id":"MjMwNDcxOTAyOTYxMjM0MDg4NA=="} +{"index":64,"citer_arxiv_id":"2102.12092","cited_arxiv_id":"1905.13727","reference_contexts":"We use PowerSGD (Vogels et al., 2019) to compress the gradients with respect to all parameters except the embeddings, unembeddings, gains, and biases","reference_id":"MTIwMzI5MjczMDgxMzgzMDYwMA=="} +{"index":65,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"1509.00519","reference_contexts":"In principle, it would be desirable to perform an importance-weighted estimation of the log-likelihood, as in importance-weighted autoencoders (Burda et al., 2015) , using multiple samples from the velocity distribution","reference_id":"NzI4ODUzNTIyNTg2MTI4Mzg1MA=="} +{"index":66,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"1509.00519","reference_contexts":"(2021) , the stochasticity of log p ε (x, v) prevents us from performing importance weighted NLL estimation over the velocity distribution (Burda et al., 2015)","reference_id":"NDEwMzA0NjA0OTM1MDM5ODk1OQ=="} +{"index":67,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2109.14518","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":68,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.15282","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":69,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2002.07101","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":70,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2104.01409","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":71,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2102.07074","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":72,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2105.14080","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":73,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.05527","reference_contexts":"This is most evident at t=0: The data and velocity distributions are independent at t=0 and the score of p 0 (v 0 |x 0 )=p 0 (v 0 ) simply corresponds to the score of the Normal distribution p 0 (v 0 ) from which the velocities are initialized, whereas the score of the data distribution p 0 (x 0 ) is highly complex and can even be unbounded (Kim et al., 2021)","reference_id":"MTE2NTQ4NTc3Njk3NjM2MDU4MTY="} +{"index":74,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.05527","reference_contexts":"3.2, this is likely because when using standard DSM our CLD would suffer from unbounded scores close to t=0, similar to previous SDEs (Kim et al., 2021)","reference_id":"NTQ4MDU4NzcxMjE0MzY1MTA4OA=="} +{"index":75,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.05527","reference_contexts":"Furthermore, when drawing a sample u 0 to diffuse in DSM, we are essentially placing an infinitely sharp Normal with unbounded score (Kim et al., 2021) at u 0 , which requires undesirable modifications or truncation tricks for stable training (Song et al., 2021c; Vahdat et al., 2021)","reference_id":"MzA0NzAwOTQ1NDU3OTQ3MDUyOQ=="} +{"index":76,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2107.00630","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":77,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.00132","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":78,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2104.14951","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":79,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2101.02388","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":80,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"1902.00996","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":81,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2108.01073","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":82,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.07582","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":83,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"1909.11573","reference_contexts":"Generative modeling of images has promising applications, for example for digital content creation and artistic expression (Bailey, 2020) , but can also be used for nefarious purposes (Vaccari & Chadwick, 2020; Mirsky & Lee, 2021; Nguyen et al., 2021)","reference_id":"MTE1ODQ5MDE5NjIzMjkxNDIyMDM="} +{"index":84,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"1210.0933","reference_contexts":"GGF uses standard Euler-based methods under the hood (Kloeden & Platen, 1992; Roberts, 2012) , in contrast to our SSCS that is derived from first principles","reference_id":"MTc0MDc4MjMwNTI3NzEyMTY1NTE="} +{"index":85,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2104.07636","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":86,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2104.02600","reference_contexts":"Recently, techniques to accelerate sampling from pre-trained SGMs have been proposed (San-Roman et al., 2021; Watson et al., 2021; Kong & Ping, 2021; Song et al., 2021a)","reference_id":"MTA4NjEyMDU2OTc3MjYyOTg5Mzg="} +{"index":87,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2104.05358","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":88,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.06819","reference_contexts":"SGMs offer high quality synthesis and sample diversity, do not require adversarial objectives, and have found applications in image (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Ho et al., 2021) , speech (Chen et al., 2021; Kong et al., 2021; Jeong et al., 2021) , and music synthesis (Mittal et al., 2021) , image editing (Meng et al., 2021; Sinha et al., 2021; Furusawa et al., 2021) , super-resolution (Saharia et al., 2021; Li et al., 2021) , image-to-image translation (Sasaki et al., 2021) , and 3D shape generation (Luo & Hu, 2021; Zhou et al., 2021)","reference_id":"MTYyNTE0NzEyMjYwMTc2Mjc2ODI="} +{"index":89,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.03802","reference_contexts":"For methods without adaptive stepsize (EM and SSCS), we use evaluation times chosen according to a quadratic function, like previous work (Song et al., 2021a; Kong & Ping, 2021; Watson et al., 2021) (indicated by QS).","reference_id":"NjYzMDgyNjU1MzY5NTAwMTc3Nw=="} +{"index":90,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.03802","reference_contexts":"Recently, techniques to accelerate sampling from pre-trained SGMs have been proposed (San-Roman et al., 2021; Watson et al., 2021; Kong & Ping, 2021; Song et al., 2021a)","reference_id":"MTA4NjEyMDU2OTc3MjYyOTg5Mzg="} +{"index":91,"citer_arxiv_id":"2112.07068","cited_arxiv_id":"2106.03802","reference_contexts":"However, prior work (Song et al., 2021a; Kong & Ping, 2021; Watson et al., 2021) has shown that it can be beneficial to focus function evaluations (neural network calls) on times t \"close to the data\"","reference_id":"MTYzODQwMDkzMzkyMDgyNDQ0MTg="} +{"index":92,"citer_arxiv_id":"2106.05527","cited_arxiv_id":"2203.13751","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":93,"citer_arxiv_id":"2106.05527","cited_arxiv_id":"2110.02037","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":94,"citer_arxiv_id":"2106.05527","cited_arxiv_id":"2205.13699","reference_contexts":"Finally, Table 8 shows that Soft Truncation also works with a nonlinear forward SDE (Kim et al., 2022) , so the scope of Soft Truncation is not limited to a family of linear SDEs.","reference_id":"OTgyNzkwNjI2MTkxMTA0MTE5NQ=="} +{"index":95,"citer_arxiv_id":"2106.05527","cited_arxiv_id":"2205.13699","reference_contexts":"Table 15 : Ablation study of Soft Truncation for CIFAR-10 trained with DDPM++ when a diffusion is combined with a normalizing flow (Kim et al., 2022)","reference_id":"MTQzMzIzMDUwMDA1OTkwOTEwNzE="} +{"index":96,"citer_arxiv_id":"2106.05527","cited_arxiv_id":"2202.09778","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":97,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2211.01324","reference_contexts":"Generative features vary with noise level, providing structured feedback at high noise levels and texture-related feedback at low noise levels [1, 32]","reference_id":"MTEwNzQ5NDU5MDIyMzUzODUwODY="} +{"index":98,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2211.01324","reference_contexts":"While diffusion models [54, 18, 57, 30, 31] have revolutionized both synthesis and editing of images [10, 42, 41, 1, 43, 13, 9, 38] and videos [5, 4, 53, 2, 12, 19] , their iterative nature remains a crucial shortcoming: At inference, a trained diffusion model usually requires dozens of network evaluations to approximate the probability path from noise to data","reference_id":"MTc4NDA5MjY5MTIyODg2OTY4NDg="} +{"index":99,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2303.04248","reference_contexts":"Improved techniques for both Consistency Distillation [56, 33, 15, 68] and Progressive Distillation [35, 28, 3] have since been introduced","reference_id":"MTAwMjEzNTEwNDg5NzIxODUzNjQ="} +{"index":100,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.15127","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":101,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2403.06807","reference_contexts":"Improved techniques for both Consistency Distillation [56, 33, 15, 68] and Progressive Distillation [35, 28, 3] have since been introduced","reference_id":"MTAwMjEzNTEwNDg5NzIxODUzNjQ="} +{"index":102,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2207.12598","reference_contexts":"Classifier-free guidance (CFG) [17] is essential for generating highquality samples","reference_id":"MTU4MDkxOTY3NTQ5MjM2ODQ1MjI="} +{"index":103,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2207.12598","reference_contexts":"During sampling, we guide the edit model with a nested classifier-free guidance formulation [17, 7] , which allows us to utilize different strengths w for the image and text conditioning.","reference_id":"MTI2OTY2Mzk1NzUwMTYzNjI5MDM="} +{"index":104,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.01506","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":105,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.04378","reference_contexts":"A potential solution is deliberately adjusting these parameters during the training phase, coupled with model conditioning on these parameters as proposed in [33]","reference_id":"MTQ3NDE0NzU3MjE5MDIyNDEyMzM="} +{"index":106,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.04378","reference_contexts":"Latent consistency models (LCM) [33, 34] leverage consistency distillation for LDMs where training is conducted exclusively in latent space, similarly to LADD","reference_id":"MTUxNzk5Mzk3Mjc1MDExMzI4NDk="} +{"index":107,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.04378","reference_contexts":"We want to highlight that LCM can potentially achieve more impressive results, as shown by SDXL-LCM [34, 33] to which we compare in Section 5.1","reference_id":"MTgyMDUwODkxNzMyMTc5NjIxMDM="} +{"index":108,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.04378","reference_contexts":"Improved techniques for both Consistency Distillation [56, 33, 15, 68] and Progressive Distillation [35, 28, 3] have since been introduced","reference_id":"MTAwMjEzNTEwNDg5NzIxODUzNjQ="} +{"index":109,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.04378","reference_contexts":"We hypothesize that larger models may facilitate LCM training, as evidenced by the substantial improvement when transitioning from SD1.5-LCM to SDXL-LCM [33]","reference_id":"MTQ1NzAyOTMzODk1MDAzNDgwOTE="} +{"index":110,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.04378","reference_contexts":"For LCM, we run a hyperparameter grid search over the skipping-step [33] , noise schedule, and full-finetuning (with and without EMA target [56] ) vs LoRA-training [34] and select the best checkpoint out of all runs and over the course of training","reference_id":"ODE4Njc3MTY4OTU5MTE3MDQ1Mg=="} +{"index":111,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.05556","reference_contexts":"Latent consistency models (LCM) [33, 34] leverage consistency distillation for LDMs where training is conducted exclusively in latent space, similarly to LADD","reference_id":"MTUxNzk5Mzk3Mjc1MDExMzI4NDk="} +{"index":112,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.05556","reference_contexts":"Naturally, a large body of work focuses on speeding up the sampling of diffusion models -both via improved samplers [55, 11, 66, 51] and distilled models that are trained to match the sample quality of their teacher models in fewer steps [44, 35, 34, 58]","reference_id":"MTA1NTU0ODk2OTQzODY4MDkyNzQ="} +{"index":113,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.05556","reference_contexts":"For LCM, we run a hyperparameter grid search over the skipping-step [33] , noise schedule, and full-finetuning (with and without EMA target [56] ) vs LoRA-training [34] and select the best checkpoint out of all runs and over the course of training","reference_id":"ODE4Njc3MTY4OTU5MTE3MDQ1Mg=="} +{"index":114,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.05556","reference_contexts":"We want to highlight that LCM can potentially achieve more impressive results, as shown by SDXL-LCM [34, 33] to which we compare in Section 5.1","reference_id":"MTgyMDUwODkxNzMyMTc5NjIxMDM="} +{"index":115,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2304.07193","reference_contexts":"Another possible option is training a discriminative feature network in latent space, yet, discriminative pretraining is non-trivial and top-performing approaches require significant engineering [8, 36]","reference_id":"NTQzMjMxOTIwMDA2NTgxNjAwMA=="} +{"index":116,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2304.07193","reference_contexts":"Adversarial Diffusion Distillation (ADD) [49] provides the current state-of-the-art method for singlestep synthesis: By leveraging a pretrained DINOv2 [36] feature extractor as the backbone of the discriminator, ADD manages to distill SDXL [38] into a single-step, real-time text-to-image model","reference_id":"MTY4MzU4NDc5NDE5NjU0OTY3MDg="} +{"index":117,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2307.01952","reference_contexts":"We use the synthetic InstrucPix2Pix dataset, for which we follow [6] and upsample the original 512 2 pixel samples using SDXL [38]","reference_id":"NTI2NDgwNjg0MjIwMTQwMzkz"} +{"index":118,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2307.01952","reference_contexts":"Adversarial Diffusion Distillation (ADD) [49] provides the current state-of-the-art method for singlestep synthesis: By leveraging a pretrained DINOv2 [36] feature extractor as the backbone of the discriminator, ADD manages to distill SDXL [38] into a single-step, real-time text-to-image model","reference_id":"MTY4MzU4NDc5NDE5NjU0OTY3MDg="} +{"index":119,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2307.01952","reference_contexts":"In particular, Adversarial Diffusion Distillation (ADD) [49] , uses a pretrained feature extractors as its discriminator, achieving performance on par with strong diffusion models such as SDXL [38] in only four steps.","reference_id":"MTUzMzA0ODg1NzE3ODc1NTM2OQ=="} +{"index":120,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2307.01952","reference_contexts":"While diffusion models [54, 18, 57, 30, 31] have revolutionized both synthesis and editing of images [10, 42, 41, 1, 43, 13, 9, 38] and videos [5, 4, 53, 2, 12, 19] , their iterative nature remains a crucial shortcoming: At inference, a trained diffusion model usually requires dozens of network evaluations to approximate the probability path from noise to data","reference_id":"MTc4NDA5MjY5MTIyODg2OTY4NDg="} +{"index":121,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2307.01952","reference_contexts":"Lastly, MAR training follows the binning strategy outlined in [38, 13] .","reference_id":"NTU4MjMzNTI3MTA1Nzk3MDk2MA=="} +{"index":122,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2307.01952","reference_contexts":"Since standard performance metrics for generative models [45, 16] measure practically relevant quality image aspects as aesthetics and scene composition only insufficiently [27, 38] we use human evaluation to compare our model against the current state-of-the-art in text-to-image synthesis","reference_id":"MTUyMDg1MDU1OTk3MTk2NTk1NTk="} +{"index":123,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2305.18290","reference_contexts":"For better human preference alignment, we finetune our models via Diffusion DPO ( [61] ), an adaption of the Direct Preference Optimization (DPO) [40] technique to diffusion models","reference_id":"NTUxOTQyOTA3NDE5MTgwMzYyMg=="} +{"index":124,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.17042","reference_contexts":"Very recent distillation works aim at reducing the number of model evaluations to a single step, enabling real-time synthesis [63, 34, 62, 49, 28]","reference_id":"MjE2OTA1MTQ3NTk3MzQ3MTc4Nw=="} +{"index":125,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.17042","reference_contexts":"Adversarial Diffusion Distillation (ADD) [49] provides the current state-of-the-art method for singlestep synthesis: By leveraging a pretrained DINOv2 [36] feature extractor as the backbone of the discriminator, ADD manages to distill SDXL [38] into a single-step, real-time text-to-image model","reference_id":"MTY4MzU4NDc5NDE5NjU0OTY3MDg="} +{"index":126,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.17042","reference_contexts":"In particular, Adversarial Diffusion Distillation (ADD) [49] , uses a pretrained feature extractors as its discriminator, achieving performance on par with strong diffusion models such as SDXL [38] in only four steps.","reference_id":"MTUzMzA0ODg1NzE3ODc1NTM2OQ=="} +{"index":127,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.17042","reference_contexts":"Even more surprisingly, smaller discriminator feature networks often offer better performance than their larger counterparts [49, 48]","reference_id":"NjQ3NDQ2NTQ3MzAwMTA1OTI0Mw=="} +{"index":128,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.17042","reference_contexts":"2 , • a significantly simplified distillation formulation that outperforms LADD's predecessor ADD [49] and a systematic study of LADD's scaling behavior,","reference_id":"MTAxMjc0MTUxNjMwNDI0NzI4NDY="} +{"index":129,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.17042","reference_contexts":"For the discriminator architecture, we mostly follow [48, 49]","reference_id":"NjkwMzY0ODU1NDIwNTk2ODQ3MQ=="} +{"index":130,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.10089","reference_contexts":"Similar to [52] we use additional data from bidirectional controlnet tasks (canny edges, keypoints, semantic segmentation, depth maps, HED lines) as well as object segmentation","reference_id":"MTU4NTI1MzYyMzUzMjUxNzk0OTY="} +{"index":131,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.10089","reference_contexts":"Following [7, 52] , we condition on the input image via channel-wise concatenation and train on paired data with edit instructions","reference_id":"ODQ2MTIzMzQ5ODY3NzY5MzcwNw=="} +{"index":132,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"1503.03585","reference_contexts":"While diffusion models [54, 18, 57, 30, 31] have revolutionized both synthesis and editing of images [10, 42, 41, 1, 43, 13, 9, 38] and videos [5, 4, 53, 2, 12, 19] , their iterative nature remains a crucial shortcoming: At inference, a trained diffusion model usually requires dozens of network evaluations to approximate the probability path from noise to data","reference_id":"MTc4NDA5MjY5MTIyODg2OTY4NDg="} +{"index":133,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.14189","reference_contexts":"Improved techniques for both Consistency Distillation [56, 33, 15, 68] and Progressive Distillation [35, 28, 3] have since been introduced","reference_id":"MTAwMjEzNTEwNDg5NzIxODUzNjQ="} +{"index":134,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2310.14189","reference_contexts":"For LCM, we run a hyperparameter grid search over the skipping-step [33] , noise schedule, and full-finetuning (with and without EMA target [56] ) vs LoRA-training [34] and select the best checkpoint out of all runs and over the course of training","reference_id":"ODE4Njc3MTY4OTU5MTE3MDQ1Mg=="} +{"index":135,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2011.13456","reference_contexts":"While diffusion models [54, 18, 57, 30, 31] have revolutionized both synthesis and editing of images [10, 42, 41, 1, 43, 13, 9, 38] and videos [5, 4, 53, 2, 12, 19] , their iterative nature remains a crucial shortcoming: At inference, a trained diffusion model usually requires dozens of network evaluations to approximate the probability path from noise to data","reference_id":"MTc4NDA5MjY5MTIyODg2OTY4NDg="} +{"index":136,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2011.13456","reference_contexts":"After training, we can generate realistic samples by numerically solving a (stochastic) differential equation (backwards from t=1 to t=0) [57, 26] , iteratively evaluating the learned denoiser D.","reference_id":"NjI0MDQzNTUwNjExMDM3MDY4NA=="} +{"index":137,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2311.12908","reference_contexts":"For better human preference alignment, we finetune our models via Diffusion DPO ( [61] ), an adaption of the Direct Preference Optimization (DPO) [40] technique to diffusion models","reference_id":"NTUxOTQyOTA3NDE5MTgwMzYyMg=="} +{"index":138,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2206.10789","reference_contexts":"For evaluation, we compute the CLIP score on all prompts from DrawBench [43] and PartiPrompts [64]","reference_id":"MzcyNDI1OTQ2ODI5MDY0NjU4NA=="} +{"index":139,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2206.10789","reference_contexts":"Both for ours and the competing models, we generate samples based on prompts from the commonly used PartiPrompts benchmark [64]","reference_id":"NTQxNzg4MDc3MjMyNzI5MzkxNA=="} +{"index":140,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2303.09618","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":141,"citer_arxiv_id":"2403.12015","cited_arxiv_id":"2402.19159","reference_contexts":"Improved techniques for both Consistency Distillation [56, 33, 15, 68] and Progressive Distillation [35, 28, 3] have since been introduced","reference_id":"MTAwMjEzNTEwNDg5NzIxODUzNjQ="} +{"index":142,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1701.07875","reference_contexts":"However, GANs require very specific choices in optimization and architectures in order to stabilize training (Arjovsky et al., 2017; Gulrajani et al., 2017; Karras et al., 2018; Brock et al., 2018) , and could fail to cover modes of the data distribution (Zhao et al., 2018) .","reference_id":"NzM5MzIyMTg4NTA4MDc3MjQzOQ=="} +{"index":143,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1809.11096","reference_contexts":"However, GANs require very specific choices in optimization and architectures in order to stabilize training (Arjovsky et al., 2017; Gulrajani et al., 2017; Karras et al., 2018; Brock et al., 2018) , and could fail to cover modes of the data distribution (Zhao et al., 2018) .","reference_id":"NzM5MzIyMTg4NTA4MDc3MjQzOQ=="} +{"index":144,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1809.11096","reference_contexts":"(2020) ) and noise conditional score networks (NCSN, Song & Ermon (2019; 2020) ) have recently achieved high sample quality comparable to GANs (Brock et al., 2018; Karras et al., 2018)","reference_id":"MTU0NjkxMjEzMzE5NTczNjkxODY="} +{"index":145,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"2009.00713","reference_contexts":"Therefore, the trained model could consider many more steps than what is considered in (Ho et al., 2020) or even a continuous time variable t (Chen et al., 2020)","reference_id":"MTY4MTg2NjI0ODkyNjkxOTk5MA=="} +{"index":146,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1806.07366","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":147,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1605.08803","reference_contexts":"(2014) ) currently exhibits higher sample quality than likelihood-based methods such as variational autoencoders (Kingma & Welling, 2013) , autoregressive models (van den Oord et al., 2016b) and normalizing flows (Rezende & Mohamed, 2015; Dinh et al., 2016)","reference_id":"NDc4NDgzOTM0MDE3MDc1MTA5"} +{"index":148,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1605.08803","reference_contexts":"Hence, DDIM has certain properties that resemble GANs (Goodfellow et al., 2014) and invertible flows (Dinh et al., 2016) , such as the ability to produce semantically meaningful interpolations","reference_id":"MTQwOTI4OTU3MDE2NDMxNTIzODE="} +{"index":149,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1810.01367","reference_contexts":"The sampling procedure of DDIM is also reminiscent of neural networks with continuous depth (Chen et al., 2018; Grathwohl et al., 2018) , since the samples it produces from the same latent variable have similar high-level visual features, regardless of the specific sample trajectory.","reference_id":"MTIyMjE0NjAxNzczMjAxODQ1MTU="} +{"index":150,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1706.08500","reference_contexts":"In Table 1 , we report the quality of the generated samples with models trained on CIFAR10 and CelebA, as measured by Frechet Inception Distance (FID (Heusel et al., 2017) ), where we vary the number of timesteps used to generate a sample (dim(τ )) and the stochasticity of the process (η)","reference_id":"ODcyMzg2ODkxMjk1NTk3MDM2Ng=="} +{"index":151,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"2006.11239","reference_contexts":"For all datasets, we set the hyperparameters α according to the heuristic in (Ho et al., 2020) to make the results directly comparable","reference_id":"MTU2Mzc1MjI4NTE1MjM4MjQ4NTg="} +{"index":152,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"2006.11239","reference_contexts":"Therefore, the trained model could consider many more steps than what is considered in (Ho et al., 2020) or even a continuous time variable t (Chen et al., 2020)","reference_id":"MTY4MTg2NjI0ODkyNjkxOTk5MA=="} +{"index":153,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1812.04948","reference_contexts":"However, GANs require very specific choices in optimization and architectures in order to stabilize training (Arjovsky et al., 2017; Gulrajani et al., 2017; Karras et al., 2018; Brock et al., 2018) , and could fail to cover modes of the data distribution (Zhao et al., 2018) .","reference_id":"NzM5MzIyMTg4NTA4MDc3MjQzOQ=="} +{"index":154,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1812.04948","reference_contexts":"Our CelebA model has five feature map resolutions from 64 × 64 to 4 × 4, and we use the original CelebA dataset (not CelebA-HQ) using the pre-processing technique from the StyleGAN (Karras et al., 2018) repository","reference_id":"MTM3MzU1MjcyODcwMTAyMzg0Njk="} +{"index":155,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1812.04948","reference_contexts":"(2020) ) and noise conditional score networks (NCSN, Song & Ermon (2019; 2020) ) have recently achieved high sample quality comparable to GANs (Brock et al., 2018; Karras et al., 2018)","reference_id":"MTU0NjkxMjEzMzE5NTczNjkxODY="} +{"index":156,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1312.6114","reference_contexts":"(2014) ) currently exhibits higher sample quality than likelihood-based methods such as variational autoencoders (Kingma & Welling, 2013) , autoregressive models (van den Oord et al., 2016b) and normalizing flows (Rezende & Mohamed, 2015; Dinh et al., 2016)","reference_id":"NDc4NDgzOTM0MDE3MDc1MTA5"} +{"index":157,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1711.09268","reference_contexts":"Our work is based on a large family of existing methods on learning generative models as transition operators of Markov chains (Sohl-Dickstein et al., 2015; Bengio et al., 2014; Salimans et al., 2014; Song et al., 2017; Goyal et al., 2017; Levy et al., 2017)","reference_id":"NTQyOTI4NDM4Mjk4NjE3Njc1OQ=="} +{"index":158,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1610.03483","reference_contexts":"The resulting model becomes an implicit probabilistic model (Mohamed & Lakshminarayanan, 2016) , where samples are generated from latent variables with a fixed procedure (from x T to x 0 )","reference_id":"NDQ0MDY1MTYyODM5OTk0MjY0"} +{"index":159,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1610.03483","reference_contexts":"DDIMs are implicit probabilistic models (Mohamed & Lakshminarayanan, 2016) and are closely related to DDPMs, in the sense that they are trained with the same objective function","reference_id":"Njg5NjIyMjQ3NjMwMzYwNjAwMw=="} +{"index":160,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1610.03483","reference_contexts":"DDIM, on the other hand, is an implicit generative model (Mohamed & Lakshminarayanan, 2016) where samples are uniquely determined from the latent variables","reference_id":"MTc2MTg5MzY2MzY3MzMyMTIwMQ=="} +{"index":161,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1505.05770","reference_contexts":"(2014) ) currently exhibits higher sample quality than likelihood-based methods such as variational autoencoders (Kingma & Welling, 2013) , autoregressive models (van den Oord et al., 2016b) and normalizing flows (Rezende & Mohamed, 2015; Dinh et al., 2016)","reference_id":"NDc4NDgzOTM0MDE3MDc1MTA5"} +{"index":162,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1401.4082","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":163,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1410.6460","reference_contexts":"Our work is based on a large family of existing methods on learning generative models as transition operators of Markov chains (Sohl-Dickstein et al., 2015; Bengio et al., 2014; Salimans et al., 2014; Song et al., 2017; Goyal et al., 2017; Levy et al., 2017)","reference_id":"NTQyOTI4NDM4Mjk4NjE3Njc1OQ=="} +{"index":164,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1503.03585","reference_contexts":"From a variational perspective, a large T allows the reverse process to be close to a Gaussian (Sohl-Dickstein et al., 2015) , so that the generative process modeled with Gaussian conditional distributions becomes a good approximation; this motivates the choice of large T values, such as T = 1000 in Ho et al","reference_id":"Mjg1MDE2MjU2MTUyNTkzNjA1OQ=="} +{"index":165,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1503.03585","reference_contexts":"Our work is based on a large family of existing methods on learning generative models as transition operators of Markov chains (Sohl-Dickstein et al., 2015; Bengio et al., 2014; Salimans et al., 2014; Song et al., 2017; Goyal et al., 2017; Levy et al., 2017)","reference_id":"NTQyOTI4NDM4Mjk4NjE3Njc1OQ=="} +{"index":166,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1503.03585","reference_contexts":"This generative Markov Chain process is either based on Langevin dynamics (Song & Ermon, 2019) or obtained by reversing a forward diffusion process that progressively turns an image into noise (Sohl-Dickstein et al., 2015) .","reference_id":"ODk2MzkyNzg3MDE3Nzg1ODg2"} +{"index":167,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1907.05600","reference_contexts":"(2020) ) and noise conditional score networks (NCSN, Song & Ermon (2019) ) have demonstrated the ability to produce samples comparable to that of GANs, without having to perform adversarial training","reference_id":"MTI3NTU1NzUwMzE5NDkxMTg4NQ=="} +{"index":168,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1907.05600","reference_contexts":"(2020) , the objective with γ = 1 is optimized instead to maximize generation performance of the trained model; this is also the same objective used in noise conditional score networks (Song & Ermon, 2019) based on score matching (Hyvärinen, 2005; Vincent, 2011)","reference_id":"MTgzNjg1Nzk5NjQ4MzUxMzAxODE="} +{"index":169,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1907.05600","reference_contexts":"This generative Markov Chain process is either based on Langevin dynamics (Song & Ermon, 2019) or obtained by reversing a forward diffusion process that progressively turns an image into noise (Sohl-Dickstein et al., 2015) .","reference_id":"ODk2MzkyNzg3MDE3Nzg1ODg2"} +{"index":170,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1907.05600","reference_contexts":"(2020) ) and noise conditional score networks (NCSN, Song & Ermon (2019; 2020) ) have recently achieved high sample quality comparable to GANs (Brock et al., 2018; Karras et al., 2018)","reference_id":"MTU0NjkxMjEzMzE5NTczNjkxODY="} +{"index":171,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"2006.09011","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":172,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"2011.13456","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":173,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1609.03499","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":174,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1601.06759","reference_contexts":"(2014) ) currently exhibits higher sample quality than likelihood-based methods such as variational autoencoders (Kingma & Welling, 2013) , autoregressive models (van den Oord et al., 2016b) and normalizing flows (Rezende & Mohamed, 2015; Dinh et al., 2016)","reference_id":"NDc4NDgzOTM0MDE3MDc1MTA5"} +{"index":175,"citer_arxiv_id":"2010.02502","cited_arxiv_id":"1605.07146","reference_contexts":"(2020) , which is a U-Net (Ronneberger et al., 2015) based on a Wide ResNet (Zagoruyko & Komodakis, 2016)","reference_id":"NTY2NjgwMTU3MTI1MDg1NTQzOA=="} +{"index":176,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1506.05751","reference_contexts":"This motivated the authors of LAPGAN (Denton et al., 2015) to develop an alternative approach to iteratively upscale low resolution generated images which can be modeled more reliably","reference_id":"MzQzMzI3NTY3NzAyMDgzNjA4"} +{"index":177,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1506.05751","reference_contexts":"A laplacian pyramid extension to this approach (Denton et al., 2015) showed higher quality images, but they still suffered from the objects looking wobbly because of noise introduced in chaining multiple models","reference_id":"ODk1MjI3NzcxMjQ3Mzk4ODg5Nw=="} +{"index":178,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1411.5928","reference_contexts":"It has been previously demonstrated that conditional generative models can learn to convincingly model object attributes like scale, rotation, and position (Dosovitskiy et al., 2014)","reference_id":"MTc0NDkxMDQ2OTc5NTczOTk5Mjc="} +{"index":179,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1411.5928","reference_contexts":"A recurrent network approach (Gregor et al., 2015) and a deconvolution network approach (Dosovitskiy et al., 2014) have also recently had some success with generating natural images","reference_id":"MjYwOTgwNDY2Mjk4MTg0ODQ3MA=="} +{"index":180,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1302.4389","reference_contexts":"This is in contrast to the original GAN paper, which used the maxout activation (Goodfellow et al., 2013) .","reference_id":"ODY4MTMxMzgzODg1MjY5Njk5Ng=="} +{"index":181,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1502.04623","reference_contexts":"A recurrent network approach (Gregor et al., 2015) and a deconvolution network approach (Dosovitskiy et al., 2014) have also recently had some success with generating natural images","reference_id":"MjYwOTgwNDY2Mjk4MTg0ODQ3MA=="} +{"index":182,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1509.01240","reference_contexts":"Recent analysis has shown that there is a direct link between how fast models learn and their generalization performance (Hardt et al., 2015)","reference_id":"MTc0NTAyMTEwMDI1NDMyNzkzMjM="} +{"index":183,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1502.03167","reference_contexts":"Third is Batch Normalization (Ioffe & Szegedy, 2015) which stabilizes learning by normalizing the input to each unit to have zero mean and unit variance","reference_id":"NTkyMTU5MzQyMjIxMDMyNjU4OA=="} +{"index":184,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1412.6980","reference_contexts":"While previous GAN work has used momentum to accelerate training, we used the Adam optimizer (Kingma & Ba, 2014) with tuned hyperparameters","reference_id":"ODcxOTQ5NTgxOTYyNDc1NTY4Mw=="} +{"index":185,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1312.6114","reference_contexts":"A variational sampling approach to generating images (Kingma & Welling, 2013) has had some success, but the samples often suffer from being blurry","reference_id":"NjY2NzY4NDQ0NjIxNzgxNTk5"} +{"index":186,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1507.02672","reference_contexts":"Another popular method is to train auto-encoders (convolutionally, stacked (Vincent et al., 2010) , separating the what and where components of the code (Zhao et al., 2015) , ladder structures (Rasmus et al., 2015) ) that encode an image into a compact code, and decode the code to reconstruct the image as accurately as possible","reference_id":"MTQ0Nzg5MDAwNzY5NDc2NDI2MDg="} +{"index":187,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1503.03585","reference_contexts":"Another approach generates images using an iterative forward diffusion process (Sohl-Dickstein et al., 2015)","reference_id":"ODU4NzI2NzgyMTgzMDY4MzE3MA=="} +{"index":188,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1412.6806","reference_contexts":"The first is the all convolutional net (Springenberg et al., 2014) which replaces deterministic spatial pooling functions (such as maxpooling) with strided convolutions, allowing the network to learn its own spatial downsampling","reference_id":"NzYxMTgyMzQxNDk2NjU3NzkwNQ=="} +{"index":189,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1412.6806","reference_contexts":"Using guided backpropagation as proposed by (Springenberg et al., 2014) , we show in Fig","reference_id":"MTg1MzQ5MTQ2NzkzNDg0MjkwMQ=="} +{"index":190,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1410.1165","reference_contexts":"The resulting code layer activations are then binarized via thresholding the ReLU activation which has been shown to be an effective information preserving technique (Srivastava et al., 2014) and provides a convenient form of semantic-hashing, allowing for linear time de-duplication","reference_id":"MTQzMDM3NDY4OTcwODU4NDk0Mjc="} +{"index":191,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1511.01844","reference_contexts":"Nearest neighbors in pixel or feature space are trivially fooled (Theis et al., 2015) by small image transforms","reference_id":"MTY5NDQwNTIzNzE2ODE2MTExMzM="} +{"index":192,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1505.00853","reference_contexts":"Within the discriminator we found the leaky rectified activation (Maas et al., 2013 ) (Xu et al., 2015) to work well, especially for higher resolution modeling","reference_id":"MzIwMzA2NzEzOTAwNTU1NDM3Mw=="} +{"index":193,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1506.03365","reference_contexts":"We trained DCGANs on three datasets, Large-scale Scene Understanding (LSUN) (Yu et al., 2015) , Imagenet-1k and a newly assembled Faces dataset","reference_id":"OTc4ODA3OTQ2MjQ1MDgxNDM2Nw=="} +{"index":194,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1506.02351","reference_contexts":"Another popular method is to train auto-encoders (convolutionally, stacked (Vincent et al., 2010) , separating the what and where components of the code (Zhao et al., 2015) , ladder structures (Rasmus et al., 2015) ) that encode an image into a compact code, and decode the code to reconstruct the image as accurately as possible","reference_id":"MTQ0Nzg5MDAwNzY5NDc2NDI2MDg="} +{"index":195,"citer_arxiv_id":"1511.06434","cited_arxiv_id":"1506.02351","reference_contexts":"This achieves state of the art (for classification using 1000 labels) at 22.48% test error, improving upon another modifcation of CNNs designed to leverage unlabled data (Zhao et al., 2015)","reference_id":"NzUxNTcxMDEyODAwMTE1OTcyMA=="} +{"index":196,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1808.04444","reference_contexts":"We found that Transformers were difficult to train with many layers, as noted by (Al-Rfou et al., 2018)","reference_id":"NzM2NDQ3MDY2NjA4NzU3NDk5Nw=="} +{"index":197,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1808.04444","reference_contexts":"Deeper Self-Attention (Al-Rfou et al., 2018) 1.06 Transformer-XL 88M (Dai et al., 2018) 1.03 Transformer-XL 277M (Dai et al., 2018) 0.99 Sparse Transformer 95M (fixed) 0.99","reference_id":"ODQ2MDIwMDY0MDI1Mjg0MDYy"} +{"index":198,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1607.06450","reference_contexts":"The norm function denotes Layer Normalization (Ba et al., 2016) , and ff(x) = W 2 f (W 1 x + b 1 ) + b 2","reference_id":"NTIzMzM0NTM2NjEwMDM4MDQyOA=="} +{"index":199,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1707.00110","reference_contexts":"Outside of generative modeling, there are several works relevant to improving the efficiency of attention based off chunking (Chiu & Raffel, 2017) or using fixed length representations (Britz et al., 2017)","reference_id":"MTU4Nzk3MzQxOTU5MDM1NDE4NTY="} +{"index":200,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1604.06174","reference_contexts":"Gradient checkpointing has been shown to be effective in reducing the memory requirements of training deep neural networks (Chen et al., 2016) , (Gruslys et al., 2016)","reference_id":"OTIzMTYyNDE4ODg3NDQzNzQ0NQ=="} +{"index":201,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1712.09763","reference_contexts":"Recently, neural autoregressive models have achieved impressive results in this domain, achieving state-of-the-art in modeling natural language (Jozefowicz et al., 2016) (Radford et al., 2018) (Dai et al., 2018) , raw audio (Van Den Oord et al., 2016) (Mehri et al., 2016) , and images (Oord et al., 2016) (Menick & Kalchbrenner, 2018) (Salimans et al., 2017) (Reed et al., 2017) (Chen et al., 2017) .","reference_id":"ODA1NDUwODQ3ODgzNDI1NTgxNA=="} +{"index":202,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1712.09763","reference_contexts":"The model achieves 2.80 bits per dim (2.798 ± 0.004 over seeds 1, 2, 3) versus the previous 2.85 state of the art (Chen et al., 2017)","reference_id":"MTE3ODEzMjc2NDcyMjM4NjY1ODA="} +{"index":203,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1712.09763","reference_contexts":"PixelCNN (Oord et al., 2016) 3.03 PixelCNN++ (Salimans et al., 2017) 2.92 Image Transformer (Parmar et al., 2018) 2.90 PixelSNAIL (Chen et al., 2017) 2.85 Sparse Transformer 59M (strided) 2.80","reference_id":"MTA1MDQ0NTY3ODUxNDIyNzI5MjA="} +{"index":204,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1705.03122","reference_contexts":"In addition to the embedding of input symbols, positional embeddings are typically used in Transformers and other location-agnostic architectures to encode the spatial relationships of data (Gehring et al., 2017) , (Parmar et al., 2018)","reference_id":"Mjk1OTQ0MDc5NDI2MjU5OTUzNQ=="} +{"index":205,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1705.03122","reference_contexts":"Other works have investigated attention with multiple \"hops\", such as (Sukhbaatar et al., 2015) and (Gehring et al., 2017) .","reference_id":"MjcwNDkzMzA5NTYwMDM1Mzg4NA=="} +{"index":206,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1603.05027","reference_contexts":"First, we use the pre-activation residual block of (He et al., 2016) , defining a network of N layers in the following way:","reference_id":"MTA5MTE3NjM2MTkwMDc2ODY0NDE="} +{"index":207,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1606.08415","reference_contexts":"Our choice of f is the Gaussian Error Linear Unit (Hendrycks & Gimpel, 2016) , f (X) = X sigmoid(1.702 • X), as used in (Radford et al., 2018)","reference_id":"OTgxMDkxNjM1MDM5MjczMzUyMg=="} +{"index":208,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1809.04281","reference_contexts":"(Huang et al., 2018) apply Transformers to MIDI generation with an efficient relative attention.","reference_id":"NjkzOTU1MDEyMzc3OTk4ODc1OA=="} +{"index":209,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1602.02410","reference_contexts":"Recently, neural autoregressive models have achieved impressive results in this domain, achieving state-of-the-art in modeling natural language (Jozefowicz et al., 2016) (Radford et al., 2018) (Dai et al., 2018) , raw audio (Van Den Oord et al., 2016) (Mehri et al., 2016) , and images (Oord et al., 2016) (Menick & Kalchbrenner, 2018) (Salimans et al., 2017) (Reed et al., 2017) (Chen et al., 2017) .","reference_id":"ODA1NDUwODQ3ODgzNDI1NTgxNA=="} +{"index":210,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1402.3511","reference_contexts":"And for audio, in addition to (Van Den Oord et al., 2016) , (Mehri et al., 2016) used a hierarchical structure and RNNs of varying clock-rates to use long contexts during inference, similar to (Koutnik et al., 2014)","reference_id":"MTQ0MjA3ODgyMDkxMzYxMTQwNA=="} +{"index":211,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1801.10198","reference_contexts":"A simple and powerful choice for model θ is a Transformer (Vaswani et al., 2017) in decoder-only mode, as demonstrated by (Radford et al., 2018) and (Liu et al., 2018)","reference_id":"NzE2ODIxMzY3NDYxMTU3ODg2Mw=="} +{"index":212,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1612.07837","reference_contexts":"Recently, neural autoregressive models have achieved impressive results in this domain, achieving state-of-the-art in modeling natural language (Jozefowicz et al., 2016) (Radford et al., 2018) (Dai et al., 2018) , raw audio (Van Den Oord et al., 2016) (Mehri et al., 2016) , and images (Oord et al., 2016) (Menick & Kalchbrenner, 2018) (Salimans et al., 2017) (Reed et al., 2017) (Chen et al., 2017) .","reference_id":"ODA1NDUwODQ3ODgzNDI1NTgxNA=="} +{"index":213,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1612.07837","reference_contexts":"And for audio, in addition to (Van Den Oord et al., 2016) , (Mehri et al., 2016) used a hierarchical structure and RNNs of varying clock-rates to use long contexts during inference, similar to (Koutnik et al., 2014)","reference_id":"MTQ0MjA3ODgyMDkxMzYxMTQwNA=="} +{"index":214,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1812.01608","reference_contexts":"Recently, neural autoregressive models have achieved impressive results in this domain, achieving state-of-the-art in modeling natural language (Jozefowicz et al., 2016) (Radford et al., 2018) (Dai et al., 2018) , raw audio (Van Den Oord et al., 2016) (Mehri et al., 2016) , and images (Oord et al., 2016) (Menick & Kalchbrenner, 2018) (Salimans et al., 2017) (Reed et al., 2017) (Chen et al., 2017) .","reference_id":"ODA1NDUwODQ3ODgzNDI1NTgxNA=="} +{"index":215,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1812.01608","reference_contexts":"Our model achieves a loss of 3.44 bits per dim (3.437 across 1 run), in comparison to the previous 3.52 (Menick & Kalchbrenner, 2018) .","reference_id":"ODgzNTAzMTkwMzEzNDQ3NzE0NA=="} +{"index":216,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1812.01608","reference_contexts":"For images, (Reed et al., 2017) models conditional independence between the pixels in order to generate many locations in parallel, and (Menick & Kalchbrenner, 2018) imposes an ordering and multi-scale upsampling procedure to generate high fidelity samples","reference_id":"NjI0NjU2NjIzMzA1MjUyMTA2MQ=="} +{"index":217,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1812.01608","reference_contexts":"PixelCNN (Oord et al., 2016) 3.57 Parallel Multiscale (Reed et al., 2017) 3.7 Glow (Kingma & Dhariwal, 2018) 3.81 SPN 150M (Menick & Kalchbrenner, 2018) 3.52 Sparse Transformer 152M (strided) 3.44 Classical music, 5 seconds at 12 kHz Sparse Transformer 152M (strided) 1.97 the test set","reference_id":"MTM0NjI1Njc5ODk0OTIyMzk2MTE="} +{"index":218,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1710.03740","reference_contexts":"We store network weights in single-precision floating-point, but otherwise compute network activations and gradients in half-precision, as in (Micikevicius et al., 2017)","reference_id":"NjQzMjg2NzYxOTExNjk5MzMyMg=="} +{"index":219,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1601.06759","reference_contexts":"Recently, neural autoregressive models have achieved impressive results in this domain, achieving state-of-the-art in modeling natural language (Jozefowicz et al., 2016) (Radford et al., 2018) (Dai et al., 2018) , raw audio (Van Den Oord et al., 2016) (Mehri et al., 2016) , and images (Oord et al., 2016) (Menick & Kalchbrenner, 2018) (Salimans et al., 2017) (Reed et al., 2017) (Chen et al., 2017) .","reference_id":"ODA1NDUwODQ3ODgzNDI1NTgxNA=="} +{"index":220,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1601.06759","reference_contexts":"It is worth noting that the Gated Pixel CNN (Oord et al., 2016) and WaveNet (Van Den Oord et al., 2016) use multiplicative interactions in their networks, which are related to self-attention.","reference_id":"MTA5Njk4MTk3MjkyNzIyMTYzMzM="} +{"index":221,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1601.06759","reference_contexts":"Architectures based off CNNs (Oord et al., 2016) have made Figure 1","reference_id":"MTMwNzkyOTU2MDQxMjIwMTEzNzY="} +{"index":222,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1601.06759","reference_contexts":"PixelCNN (Oord et al., 2016) 3.57 Parallel Multiscale (Reed et al., 2017) 3.7 Glow (Kingma & Dhariwal, 2018) 3.81 SPN 150M (Menick & Kalchbrenner, 2018) 3.52 Sparse Transformer 152M (strided) 3.44 Classical music, 5 seconds at 12 kHz Sparse Transformer 152M (strided) 1.97 the test set","reference_id":"MTM0NjI1Njc5ODk0OTIyMzk2MTE="} +{"index":223,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1601.06759","reference_contexts":"In order to test the ability of the model to learn long range dependencies and scale to a large dataset, we train on the version of downsampled ImageNet released by (Oord et al., 2016) and evaluate on the validation set","reference_id":"MTEyMzM4NjQ4NzcxMjgzMDgzNjI="} +{"index":224,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1601.06759","reference_contexts":"PixelCNN (Oord et al., 2016) 3.03 PixelCNN++ (Salimans et al., 2017) 2.92 Image Transformer (Parmar et al., 2018) 2.90 PixelSNAIL (Chen et al., 2017) 2.85 Sparse Transformer 59M (strided) 2.80","reference_id":"MTA1MDQ0NTY3ODUxNDIyNzI5MjA="} +{"index":225,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1802.05751","reference_contexts":"In addition to the embedding of input symbols, positional embeddings are typically used in Transformers and other location-agnostic architectures to encode the spatial relationships of data (Gehring et al., 2017) , (Parmar et al., 2018)","reference_id":"Mjk1OTQ0MDc5NDI2MjU5OTUzNQ=="} +{"index":226,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1802.05751","reference_contexts":"PixelCNN (Oord et al., 2016) 3.03 PixelCNN++ (Salimans et al., 2017) 2.92 Image Transformer (Parmar et al., 2018) 2.90 PixelSNAIL (Chen et al., 2017) 2.85 Sparse Transformer 59M (strided) 2.80","reference_id":"MTA1MDQ0NTY3ODUxNDIyNzI5MjA="} +{"index":227,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1802.05751","reference_contexts":"(Parmar et al., 2018) uses blocks of local attention to apply Transformers to images","reference_id":"MTI3NzU1NjY1MTUxMTg0MjMwMDU="} +{"index":228,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1703.03664","reference_contexts":"Recently, neural autoregressive models have achieved impressive results in this domain, achieving state-of-the-art in modeling natural language (Jozefowicz et al., 2016) (Radford et al., 2018) (Dai et al., 2018) , raw audio (Van Den Oord et al., 2016) (Mehri et al., 2016) , and images (Oord et al., 2016) (Menick & Kalchbrenner, 2018) (Salimans et al., 2017) (Reed et al., 2017) (Chen et al., 2017) .","reference_id":"ODA1NDUwODQ3ODgzNDI1NTgxNA=="} +{"index":229,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1703.03664","reference_contexts":"For images, (Reed et al., 2017) models conditional independence between the pixels in order to generate many locations in parallel, and (Menick & Kalchbrenner, 2018) imposes an ordering and multi-scale upsampling procedure to generate high fidelity samples","reference_id":"NjI0NjU2NjIzMzA1MjUyMTA2MQ=="} +{"index":230,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1703.03664","reference_contexts":"PixelCNN (Oord et al., 2016) 3.57 Parallel Multiscale (Reed et al., 2017) 3.7 Glow (Kingma & Dhariwal, 2018) 3.81 SPN 150M (Menick & Kalchbrenner, 2018) 3.52 Sparse Transformer 152M (strided) 3.44 Classical music, 5 seconds at 12 kHz Sparse Transformer 152M (strided) 1.97 the test set","reference_id":"MTM0NjI1Njc5ODk0OTIyMzk2MTE="} +{"index":231,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1701.05517","reference_contexts":"Recently, neural autoregressive models have achieved impressive results in this domain, achieving state-of-the-art in modeling natural language (Jozefowicz et al., 2016) (Radford et al., 2018) (Dai et al., 2018) , raw audio (Van Den Oord et al., 2016) (Mehri et al., 2016) , and images (Oord et al., 2016) (Menick & Kalchbrenner, 2018) (Salimans et al., 2017) (Reed et al., 2017) (Chen et al., 2017) .","reference_id":"ODA1NDUwODQ3ODgzNDI1NTgxNA=="} +{"index":232,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1701.05517","reference_contexts":"PixelCNN (Oord et al., 2016) 3.03 PixelCNN++ (Salimans et al., 2017) 2.92 Image Transformer (Parmar et al., 2018) 2.90 PixelSNAIL (Chen et al., 2017) 2.85 Sparse Transformer 59M (strided) 2.80","reference_id":"MTA1MDQ0NTY3ODUxNDIyNzI5MjA="} +{"index":233,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1609.03499","reference_contexts":"And for audio, in addition to (Van Den Oord et al., 2016) , (Mehri et al., 2016) used a hierarchical structure and RNNs of varying clock-rates to use long contexts during inference, similar to (Koutnik et al., 2014)","reference_id":"MTQ0MjA3ODgyMDkxMzYxMTQwNA=="} +{"index":234,"citer_arxiv_id":"1904.10509","cited_arxiv_id":"1609.03499","reference_contexts":"It is worth noting that the Gated Pixel CNN (Oord et al., 2016) and WaveNet (Van Den Oord et al., 2016) use multiplicative interactions in their networks, which are related to self-attention.","reference_id":"MTA5Njk4MTk3MjkyNzIyMTYzMzM="} +{"index":235,"citer_arxiv_id":"1401.4082","cited_arxiv_id":"1401.1022","reference_contexts":"One popular approach is the REINFORCE algorithm (Williams, 1992) , since it is simple to implement and applicable to both discrete and continuous models, though control variate methods are becoming increasingly popular for variational inference problems (Hoffman et al., 2013; Blei et al., 2012; Ranganath et al., 2014; Salimans & Knowles, 2014)","reference_id":"NTYwNzA5NzIyNTM2NzQxNjM0OA=="} +{"index":236,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1603.04467","reference_contexts":"In the previous sections, we have applied the dual decomposition method to Problem (1) and devised the primal-dual algorithm, which is a single time-scale algorithm","reference_id":"NjAyNTAyNjQyOTk0NjY4NTM1OQ=="} +{"index":237,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1603.04467","reference_contexts":"The software framework for our experiments was Tensorflow 1.3 [1, 2] and Python 3.6","reference_id":"MTc5ODE3NDA4ODA5NzA3MTU4Mzk="} +{"index":238,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1701.07875","reference_contexts":"Generative adversarial networks (GANs) [18] have achieved outstanding results in generating realistic images [51, 36, 27, 3, 6] and producing text [23]","reference_id":"MTU2MzYxMDg2MjEzOTQzMTg4MDk="} +{"index":239,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1701.07875","reference_contexts":"[18] , its improved versions [20] , or recently proposed losses for GANs like the Wasserstein GAN [3]","reference_id":"MTI5OTM5NTY1NDA1NDg0MjkwMzA="} +{"index":240,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1701.07875","reference_contexts":"The Wasserstein GAN algorithm uses more update steps for the discriminator than for the generator [3]","reference_id":"NzI5Mzc3NTcyNDg0MTEzNTA0NQ=="} +{"index":241,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1703.10717","reference_contexts":"Generative adversarial networks (GANs) [18] have achieved outstanding results in generating realistic images [51, 36, 27, 3, 6] and producing text [23]","reference_id":"MTU2MzYxMDg2MjEzOTQzMTg4MDk="} +{"index":242,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1703.10717","reference_contexts":"The Boundary Equilibrium GAN (BEGAN) [6] maintains an equilibrium between the discriminator and generator loss (cf","reference_id":"MzAxNzIwOTIyMDI1NTkyMDc0NQ=="} +{"index":243,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1612.02136","reference_contexts":"Different methods have been proposed to avoid mode collapsing [11, 43]","reference_id":"MTgwMjE3NzIwMjM2OTU0NzUzNTk="} +{"index":244,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1312.3005","reference_contexts":"Finally the One Billion Word Benchmark [12] serves to evaluate TTUR on WGAN-GP","reference_id":"NzAyMDY0Nzg5OTEzNzc5ODk2Mw=="} +{"index":245,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1312.3005","reference_contexts":"We used the following datasets to evaluate GANs: The Large-scale CelebFaces Attributes (CelebA) dataset, aligned and cropped [41] , the training dataset of the bedrooms category of the large scale image database (LSUN) [60] , the CIFAR-10 training dataset [34] , the Street View House Numbers training dataset (SVHN) [48] , and the One Billion Word Benchmark [12] .","reference_id":"MTI4MzYzMzE5MDIyMjAyNTc1MjY="} +{"index":246,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1511.07289","reference_contexts":"Consequently, networks with Lipschitz smooth activation functions like ELUs (α = 1) [13] fulfill the assumption but not ReLU networks.","reference_id":"MTcwMTk4MjM5ODQ4NjY1OTI1OTY="} +{"index":247,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1412.6515","reference_contexts":"To characterize the convergence properties of training general GANs is still an open challenge [19, 20]","reference_id":"MTgyODYxMDY3ODk1MjcxMjcwNjI="} +{"index":248,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1701.00160","reference_contexts":"To characterize the convergence properties of training general GANs is still an open challenge [19, 20]","reference_id":"MTgyODYxMDY3ODk1MjcxMjcwNjI="} +{"index":249,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1701.00160","reference_contexts":"[18] , its improved versions [20] , or recently proposed losses for GANs like the Wasserstein GAN [3]","reference_id":"MTI5OTM5NTY1NDA1NDg0MjkwMzA="} +{"index":250,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1701.00160","reference_contexts":"Since training GANs is a game and its solution is a Nash equilibrium, gradient descent may fail to converge [53, 18, 20]","reference_id":"MTkxNjM4NDkzODA3MTE0MjIxMQ=="} +{"index":251,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1706.03269","reference_contexts":"For special GAN variants, convergence can be proved under certain assumptions [39, 22, 56]","reference_id":"MTAwNzU1MjI1MzY0NTc1NTk3MTQ="} +{"index":252,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1704.00028","reference_contexts":"We used the WGAN-GP image model [23] to test TTUR with the CIFAR-10 and LSUN Bedrooms datasets","reference_id":"MTQ5ODE5MzUwMDAwNDkyOTUxNTM="} +{"index":253,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1704.00028","reference_contexts":"Generative adversarial networks (GANs) [18] have achieved outstanding results in generating realistic images [51, 36, 27, 3, 6] and producing text [23]","reference_id":"MTU2MzYxMDg2MjEzOTQzMTg4MDk="} +{"index":254,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1704.00028","reference_contexts":"Since the FID criterium only works for images, we measured the performance by the Jensen-Shannon-divergence (JSD) between the model and the real world distribution as has been done previously [23]","reference_id":"MTMyNDI5NDIyMjQ1NDk4Nzc4MDI="} +{"index":255,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1702.08431","reference_contexts":"Recent convergence proofs for GANs hold for expectations over training samples or for the number of examples going to infinity [37, 45, 40, 4] , thus do not consider mini-batch learning which leads to a stochastic gradient [57, 25, 42, 38] .","reference_id":"ODc3OTM5ODE5NjIyMjE1NDU2OQ=="} +{"index":256,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1611.07004","reference_contexts":"Generative adversarial networks (GANs) [18] have achieved outstanding results in generating realistic images [51, 36, 27, 3, 6] and producing text [23]","reference_id":"MTU2MzYxMDg2MjEzOTQzMTg4MDk="} +{"index":257,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1611.07004","reference_contexts":"(27) holds:","reference_id":"Mjk0NzUwMzc3NDQ5OTYxODQ4Mw=="} +{"index":258,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1412.6980","reference_contexts":"We obviate mode collapsing by using Adam stochastic approximation [29]","reference_id":"NzM3OTk3OTEwMjU2NTcxOTM3"} +{"index":259,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1412.6980","reference_contexts":"( 166 ) is the first moment update of ADAM [29] .","reference_id":"MTA5NjI1ODIxNDE4MTI3NjIyMTM="} +{"index":260,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1412.6980","reference_contexts":"These update rules are the first moment update rules of Adam [29]","reference_id":"MTE3ODYzNzI2MjY3NDA2ODg3NzU="} +{"index":261,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1412.6980","reference_contexts":"[29] state in Theorem 4.1 convergence of ADAM while assuming that β 1 , the first moment running average coefficient, decays exponentially","reference_id":"MTY2NDIxNTYwMDI0NTg3ODk3MDU="} +{"index":262,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1609.04802","reference_contexts":"Generative adversarial networks (GANs) [18] have achieved outstanding results in generating realistic images [51, 36, 27, 3, 6] and producing text [23]","reference_id":"MTU2MzYxMDg2MjEzOTQzMTg4MDk="} +{"index":263,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1705.08584","reference_contexts":"Recent convergence proofs for GANs hold for expectations over training samples or for the number of examples going to infinity [37, 45, 40, 4] , thus do not consider mini-batch learning which leads to a stochastic gradient [57, 25, 42, 38] .","reference_id":"ODc3OTM5ODE5NjIyMjE1NDU2OQ=="} +{"index":264,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1706.09884","reference_contexts":"Recent convergence proofs for GANs hold for expectations over training samples or for the number of examples going to infinity [37, 45, 40, 4] , thus do not consider mini-batch learning which leads to a stochastic gradient [57, 25, 42, 38] .","reference_id":"ODc3OTM5ODE5NjIyMjE1NDU2OQ=="} +{"index":265,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1705.02894","reference_contexts":"For special GAN variants, convergence can be proved under certain assumptions [39, 22, 56]","reference_id":"MTAwNzU1MjI1MzY0NTc1NTk3MTQ="} +{"index":266,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1705.08991","reference_contexts":"Recent convergence proofs for GANs hold for expectations over training samples or for the number of examples going to infinity [37, 45, 40, 4] , thus do not consider mini-batch learning which leads to a stochastic gradient [57, 25, 42, 38] .","reference_id":"ODc3OTM5ODE5NjIyMjE1NDU2OQ=="} +{"index":267,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1611.02163","reference_contexts":"Different methods have been proposed to avoid mode collapsing [11, 43]","reference_id":"MTgwMjE3NzIwMjM2OTU0NzUzNTk="} +{"index":268,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1705.09675","reference_contexts":"Recent convergence proofs for GANs hold for expectations over training samples or for the number of examples going to infinity [37, 45, 40, 4] , thus do not consider mini-batch learning which leads to a stochastic gradient [57, 25, 42, 38] .","reference_id":"ODc3OTM5ODE5NjIyMjE1NDU2OQ=="} +{"index":269,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1706.04156","reference_contexts":"Futhermore, the Appendix presents work on the rate of convergence for both linear and nonlinear update rules using similar techniques as the local stability analysis of Nagarajan and Kolter [46]","reference_id":"NjgyODQwOTgwMjQ2MDA1Mzc0OA=="} +{"index":270,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1706.04156","reference_contexts":"A prerequisit for many convergence proofs is local stability [35] which was shown for GANs by Nagarajan and Kolter [46] for a min-max GAN setting","reference_id":"MTIxNzI2ODIxOTMzMzI2NDI4Mzk="} +{"index":271,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1511.06434","reference_contexts":"Generative adversarial networks (GANs) [18] have achieved outstanding results in generating realistic images [51, 36, 27, 3, 6] and producing text [23]","reference_id":"MTU2MzYxMDg2MjEzOTQzMTg4MDk="} +{"index":272,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1511.01844","reference_contexts":"However, defining appropriate performance measures for generative models is difficult [55]","reference_id":"NzUyODc1ODczMDg0NTk5NjQz"} +{"index":273,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1701.02386","reference_contexts":"For special GAN variants, convergence can be proved under certain assumptions [39, 22, 56]","reference_id":"MTAwNzU1MjI1MzY0NTc1NTk3MTQ="} +{"index":274,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1611.04273","reference_contexts":"The best known measure is the likelihood, which can be estimated by annealed importance sampling [59]","reference_id":"MTA1MDcyMDczNDg5NDA5MDY3ODE="} +{"index":275,"citer_arxiv_id":"1706.08500","cited_arxiv_id":"1506.03365","reference_contexts":"We used the following datasets to evaluate GANs: The Large-scale CelebFaces Attributes (CelebA) dataset, aligned and cropped [41] , the training dataset of the bedrooms category of the large scale image database (LSUN) [60] , the CIFAR-10 training dataset [34] , the Street View House Numbers training dataset (SVHN) [48] , and the One Billion Word Benchmark [12] .","reference_id":"MTI4MzYzMzE5MDIyMjAyNTc1MjY="} +{"index":276,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1706.01350","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":277,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1809.11096","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":278,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1810.01365","reference_contexts":"[6] , it is becoming clear that the traditional GAN generator architecture is in every way inferior to a style-based design","reference_id":"NDA0MDA1ODIzNzQ2MjE0NzE0OA=="} +{"index":279,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1810.01365","reference_contexts":"[6] \"self modulate\" the generator using AdaINs, similarly to our work, but do not consider an intermediate latent space or noise inputs.","reference_id":"NDkzNzc3MTE1MTEzNTY1NDg2OQ=="} +{"index":280,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1810.01365","reference_contexts":"We observe that the style-based generator (E) improves FIDs quite significantly over the traditional generator (B), almost 20%, corroborating the large-scale ImageNet measurements made in parallel work [6, 5]","reference_id":"MjIyNjI3NDg0MzYzNTU1MjA3"} +{"index":281,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1802.04942","reference_contexts":"As such, we expect the training to yield a less entangled W in an unsupervised setting, i.e., when the factors of variation are not known in advance [10, 35, 49, 8, 26, 32, 7] .","reference_id":"MTYxMzY4OTI0MDQ0MTg3NzExMzQ="} +{"index":282,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1802.04942","reference_contexts":"Unfortunately the metrics recently proposed for quantifying disentanglement [26, 32, 7, 19 ] require an encoder network that maps input images to latent codes","reference_id":"MjgxNDYyODQ0MDgwNTE1OTIxNw=="} +{"index":283,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1606.03657","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":284,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1506.05751","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":285,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1605.09782","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":286,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1611.01673","reference_contexts":"Much of the work on GAN architectures has focused on improving the discriminator by, e.g., using multiple discriminators [18, 47, 11] , multiresolution discrimination [60, 55] , or self-attention [63]","reference_id":"MTQ3MzQ5MTI5NTYwNzkyMjE1NjA="} +{"index":287,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1704.00028","reference_contexts":"We evaluate our methods using two different loss functions: for CELEBA-HQ we rely on WGAN-GP [24] , Figure 2","reference_id":"MzU0MDc4NTM1MDUyMzIxMzU0OQ=="} +{"index":288,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1704.00028","reference_contexts":"We do not modify the discriminator or the loss function in any way, and our work is thus orthogonal to the ongoing discussion about GAN loss functions, regularization, and hyperparameters [24, 45, 5, 40, 44, 36] .","reference_id":"OTY2Mzg4NTcwMTY4NzAwMzk0Mg=="} +{"index":289,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1703.06868","reference_contexts":"We choose to reuse the word \"style\" for y because similar network architectures are already used for feedforward style transfer [27] , unsupervised image-toimage translation [28] , and domain mixtures [23]","reference_id":"MTUyMTM5MDI0NTcyMDU2Mjg4MTQ="} +{"index":290,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1703.06868","reference_contexts":"Motivated by style transfer literature [27] , we re-design the generator architecture in a way that exposes novel ways to control the image synthesis process","reference_id":"ODMwNDM0NDYxMzYzNDQxNjM2Ng=="} +{"index":291,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1703.06868","reference_contexts":"Learned affine transformations then specialize w to styles y = (y s , y b ) that control adaptive instance normalization (AdaIN) [27, 17, 21, 16] operations after each convolution layer of the synthesis network g","reference_id":"ODM3MjA5NTMxODQ0OTcwMzc1Nw=="} +{"index":292,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"The resolution and quality of images produced by generative methods -especially generative adversarial networks (GAN) [22] -have seen rapid improvement recently [30, 45, 5 ]","reference_id":"MTI5NTc4MDkwNzkxMjU4OTE0MDA="} +{"index":293,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"While a traditional generator [30] feeds the latent code though the input layer only, we first map the input to an intermediate latent space W, which then controls the generator through adaptive instance normalization (AdaIN) at each convolution layer","reference_id":"MTcyNTIxMzA1MDE1NzU4NTA1NjA="} +{"index":294,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"[30] , but we start from 8 2 images instead of 4 2","reference_id":"NDE0OTEzNDMyNDE3OTMzMjQ4MA=="} +{"index":295,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"The classifiers used by our separability metric (Section 4.2) have the same architecture as our discriminator except that minibatch standard deviation [30] is disabled","reference_id":"ODkzNjMwNDkzMzEyMjA0MDk4OQ=="} +{"index":296,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"Table 1 gives Fréchet inception distances (FID) [25] for various generator architectures in CELEBA-HQ [30] and our new FFHQ dataset (Appendix A)","reference_id":"MTA4NTY5MjY1NjA1OTUyNjk0OTY="} +{"index":297,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"[30] , from which we inherit most of the training details","reference_id":"MTMwNjA5ODg2Njg4MjQyNTc4MDQ="} +{"index":298,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"In our tests, the classifiers had the same architecture as the discriminator we use (i.e., same as in [30] ), and were trained using the CELEBA-HQ dataset that retains the 40 attributes available in the original CelebA dataset","reference_id":"OTAzNjUyNzk3ODYwNzM4OTYx"} +{"index":299,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"The dataset includes vastly more variation than CELEBA-HQ [30] in terms of age, ethnicity and image background, and also has much better coverage of accessories such as eyeglasses, sunglasses, hats, etc","reference_id":"MTM1MTU4MTQwNDMyNTU3MDExMTI="} +{"index":300,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1710.10196","reference_contexts":"For our style-based generator (F in Table 1 ), we use leaky ReLU [41] with α = 0.2 and equalized learning rate [30] for all layers","reference_id":"MTk3ODAzNDc4NzEwMjM2OTQ1Ng=="} +{"index":301,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1807.03039","reference_contexts":"For this figure, we avoided sampling from the extreme regions of W using the so-called truncation trick [42, 5, 34] -Appendix B details how the trick can be performed in W instead of Z","reference_id":"MTE0NDgwNjY4NDU1NDgwOTAzNDA="} +{"index":302,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1807.03039","reference_contexts":"Here we used a variation of the truncation trick [42, 5, 34] with ψ = 0.7 for resolutions 4 2 -32 2","reference_id":"MTE3MDgzNTUxODA1MzQxNzMxNTA="} +{"index":303,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1807.03039","reference_contexts":"However, it is known that drawing latent vectors from a truncated [42, 5] or otherwise shrunk [34] sampling space tends to improve average image quality, although some amount of variation is lost.","reference_id":"MzU5MjEzMzU4NzM3MTAwNzYzOA=="} +{"index":304,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1807.04720","reference_contexts":"We do not modify the discriminator or the loss function in any way, and our work is thus orthogonal to the ongoing discussion about GAN loss functions, regularization, and hyperparameters [24, 45, 5, 40, 44, 36] .","reference_id":"OTY2Mzg4NTcwMTY4NzAwMzk0Mg=="} +{"index":305,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1701.01036","reference_contexts":"This observation is in line with style transfer literature, where it has been established that spatially invariant statistics (Gram matrix, channel-wise mean, variance, etc.) reliably encode the style of an image [20, 39] while spatially varying features encode a specific instance.","reference_id":"MTcyNjY4ODI3NDE5NTE3MjY4Mjc="} +{"index":306,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1706.00082","reference_contexts":"For this figure, we avoided sampling from the extreme regions of W using the so-called truncation trick [42, 5, 34] -Appendix B details how the trick can be performed in W instead of Z","reference_id":"MTE0NDgwNjY4NDU1NDgwOTAzNDA="} +{"index":307,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1706.00082","reference_contexts":"Here we used a variation of the truncation trick [42, 5, 34] with ψ = 0.7 for resolutions 4 2 -32 2","reference_id":"MTE3MDgzNTUxODA1MzQxNzMxNTA="} +{"index":308,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1706.00082","reference_contexts":"However, it is known that drawing latent vectors from a truncated [42, 5] or otherwise shrunk [34] sampling space tends to improve average image quality, although some amount of variation is lost.","reference_id":"MzU5MjEzMzU4NzM3MTAwNzYzOA=="} +{"index":309,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1801.04406","reference_contexts":"For the FFHQ dataset, we switch from WGAN-GP to the non-saturating loss [22] with R 1 regularization [44] using γ = 10","reference_id":"MTcxNjgzMjY4OTc3MDE3ODQ1OTI="} +{"index":310,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1801.04406","reference_contexts":"while FFHQ uses WGAN-GP for configuration A and nonsaturating loss [22] with R 1 regularization [44, 51, 14] for configurations B-F","reference_id":"ODcxMzM4ODkzNzA1Njg3OTkyMQ=="} +{"index":311,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1801.04406","reference_contexts":"We do not modify the discriminator or the loss function in any way, and our work is thus orthogonal to the ongoing discussion about GAN loss functions, regularization, and hyperparameters [24, 45, 5, 40, 44, 36] .","reference_id":"OTY2Mzg4NTcwMTY4NzAwMzk0Mg=="} +{"index":312,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1802.05957","reference_contexts":"We do not use batch normalization [29], spectral normalization [45] , attention mechanisms [63] , dropout [59] , or pixelwise feature vector normalization [30] in our networks.","reference_id":"MzE1MTgwMzgyMzAwNzc3NjA4NA=="} +{"index":313,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1802.05957","reference_contexts":"The resolution and quality of images produced by generative methods -especially generative adversarial networks (GAN) [22] -have seen rapid improvement recently [30, 45, 5 ]","reference_id":"MTI5NTc4MDkwNzkxMjU4OTE0MDA="} +{"index":314,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1802.05957","reference_contexts":"We do not modify the discriminator or the loss function in any way, and our work is thus orthogonal to the ongoing discussion about GAN loss functions, regularization, and hyperparameters [24, 45, 5, 40, 44, 36] .","reference_id":"OTY2Mzg4NTcwMTY4NzAwMzk0Mg=="} +{"index":315,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1802.05637","reference_contexts":"Recent conditional generators feed the class identifier through a separate embedding network to a large number of layers in the generator [46] , while the latent is still provided though the input layer","reference_id":"MTA2ODI0NDMyMjA1Mzg3MjA2NTI="} +{"index":316,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1807.11346","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":317,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1809.03627","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":318,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1612.05299","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":319,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1711.11585","reference_contexts":"Much of the work on GAN architectures has focused on improving the discriminator by, e.g., using multiple discriminators [18, 47, 11] , multiresolution discrimination [60, 55] , or self-attention [63]","reference_id":"MTQ3MzQ5MTI5NTYwNzkyMjE1NjA="} +{"index":320,"citer_arxiv_id":"1812.04948","cited_arxiv_id":"1805.08318","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":321,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1603.04467","reference_contexts":"We used 200K synchronous updates over 32 GPUs in TensorFlow [1]","reference_id":"MTQ2ODc4MTQyMjQ1NzI2ODY4NzI="} +{"index":322,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1606.01868","reference_contexts":"As well as providing excellent samples, this network has the advantage of returning explicit probability densities (unlike alternatives such as generative adversarial networks [6, 3, 19] ), making it straightforward to apply in domains such as compression [32] and probabilistic planning and exploration [2]","reference_id":"MTcyMDE4NzUxMzA5MzUzOTc0OTM="} +{"index":323,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1410.8516","reference_contexts":"Uniform Distribution: [30] 8.00 Multivariate Gaussian: [30] 4.70 NICE: [4] 4.48 Deep Diffusion: [24] 4.20 DRAW: [9] 4.13 Deep GMMs: [31, 29] 4.00 Conv DRAW: [8] 3.58 (3.57) RIDE: [26, 30] 3.47 PixelCNN: [30] 3.14 (3.08) PixelRNN: [30] 3.00 (2.93)","reference_id":"Mzc1NzI4NDA1OTU2Nzc5NzMwOA=="} +{"index":324,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1508.06576","reference_contexts":"Neural artwork [18, 5] and content generation represent potential future uses for conditional generation","reference_id":"MTg0MDU0ODg0MjA2Nzc3NzAwMDQ="} +{"index":325,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"Uniform Distribution: [30] 8.00 Multivariate Gaussian: [30] 4.70 NICE: [4] 4.48 Deep Diffusion: [24] 4.20 DRAW: [9] 4.13 Deep GMMs: [31, 29] 4.00 Conv DRAW: [8] 3.58 (3.57) RIDE: [26, 30] 3.47 PixelCNN: [30] 3.14 (3.08) PixelRNN: [30] 3.00 (2.93)","reference_id":"Mzc1NzI4NDA1OTU2Nzc5NzMwOA=="} +{"index":326,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1512.03385","reference_contexts":"As proposed in [30] we also use a residual connection [11] in the horizontal stack","reference_id":"MTEzNzY5NjU4NzcxMTI3OTU0NjE="} +{"index":327,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1511.08228","reference_contexts":"Feed-forward neural networks with gates have been explored in previous works, such as highway networks [25] , grid LSTM [13] and neural GPUs [12] , and have generally proved beneficial to performance.","reference_id":"MzUyNTIxMzI1MjM1MDQ2Njcx"} +{"index":328,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1507.01526","reference_contexts":"Feed-forward neural networks with gates have been explored in previous works, such as highway networks [25] , grid LSTM [13] and neural GPUs [12] , and have generally proved beneficial to performance.","reference_id":"MzUyNTIxMzI1MjM1MDQ2Njcx"} +{"index":329,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1511.02793","reference_contexts":"[15]","reference_id":"OTIzOTI5MDgzNjQ2NzIxNjA0Ng=="} +{"index":330,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1605.05396","reference_contexts":"As well as providing excellent samples, this network has the advantage of returning explicit probability densities (unlike alternatives such as generative adversarial networks [6, 3, 19] ), making it straightforward to apply in domains such as compression [32] and probabilistic planning and exploration [2]","reference_id":"MTcyMDE4NzUxMzA5MzUzOTc0OTM="} +{"index":331,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1603.05106","reference_contexts":"In the future it might be interesting to try and generate new images with a certain animal or object solely from a single example image [21, 22]","reference_id":"MzU3MjAxMjI0NzkzMDMxNzY3OQ=="} +{"index":332,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1511.01844","reference_contexts":"On the other hand, as noted in [27] , we observed great improvements in the visual quality of the generated samples.","reference_id":"MTA2NDI3ODQyMDUwNjc3MDM1OA=="} +{"index":333,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1605.02226","reference_contexts":"Recent advances in image modelling with neural networks [30, 26, 20, 10, 9, 28, 6] have made it feasible to generate diverse natural images that capture the high-level structure of the training data","reference_id":"MTM3NjUwOTgxOTUxNjM3NjA1Nzc="} +{"index":334,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"PixelCNNs (and PixelRNNs) [30] model the joint distribution of pixels over an image x as the following product of conditional distributions, where x i is a single pixel:","reference_id":"MTUwOTE5NTIwMzU3NjMzNDkxMDE="} +{"index":335,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"As proposed in [30] we also use a residual connection [11] in the horizontal stack","reference_id":"MTEzNzY5NjU4NzcxMTI3OTU0NjE="} +{"index":336,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"We were able to achieve similar performance to the PixelRNN (Row LSTM [30] ) in less than half the training time (60 hours using 32 GPUs)","reference_id":"MTEyOTIxMDE0MjY3MDAyMjEzMzQ="} +{"index":337,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"Uniform Distribution: [30] 8.00 Multivariate Gaussian: [30] 4.70 NICE: [4] 4.48 Deep Diffusion: [24] 4.20 DRAW: [9] 4.13 Deep GMMs: [31, 29] 4.00 Conv DRAW: [8] 3.58 (3.57) RIDE: [26, 30] 3.47 PixelCNN: [30] 3.14 (3.08) PixelRNN: [30] 3.00 (2.93)","reference_id":"Mzc1NzI4NDA1OTU2Nzc5NzMwOA=="} +{"index":338,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"This work introduced the Gated PixelCNN, an improvement over the original PixelCNN that is able to match or outperform PixelRNN [30] , and is computationally more efficient","reference_id":"NzEyNzc3NzY5MjY3ODY5MDY2Mw=="} +{"index":339,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"Recent advances in image modelling with neural networks [30, 26, 20, 10, 9, 28, 6] have made it feasible to generate diverse natural images that capture the high-level structure of the training data","reference_id":"MTM3NjUwOTgxOTUxNjM3NjA1Nzc="} +{"index":340,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"PixelRNNs, which use spatial LSTM layers instead of convolutional stacks, have previously been shown to outperform PixelCNNs as generative models [30]","reference_id":"MTE1MTgzNzgwMDgyNzk0MTM5MzY="} +{"index":341,"citer_arxiv_id":"1606.05328","cited_arxiv_id":"1601.06759","reference_contexts":"This paper explores the potential for conditional image modelling by adapting and improving a convolutional variant of the PixelRNN architecture [30]","reference_id":"Mjg0Nzc3NTQzNDUxMjQyMjg0NA=="} +{"index":342,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1302.4389","reference_contexts":"Strong regularization such as maxout [10] or dropout [14] is applied to obtain the best results ( [10, 25, 24, 35] ) on this dataset","reference_id":"NDI2OTQwNDg0OTU0NjMyNDA2Mg=="} +{"index":343,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1302.4389","reference_contexts":"On this dataset we use identity shortcuts in all cases (i.e., option A), method error (%) Maxout [10] 9.38 NIN [25] 8.81 DSN [24] 8.22 # layers # params FitNet [ so our residual models have exactly the same depth, width, and number of parameters as the plain counterparts","reference_id":"NTA1MjYyMzk1NzI5Mjg2MDUz"} +{"index":344,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1207.0580","reference_contexts":"Strong regularization such as maxout [10] or dropout [14] is applied to obtain the best results ( [10, 25, 24, 35] ) on this dataset","reference_id":"NDI2OTQwNDg0OTU0NjMyNDA2Mg=="} +{"index":345,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1207.0580","reference_contexts":"We do not use dropout [14] , following the practice in [16] .","reference_id":"MTUzOTk0NzM1MTg1MzE4ODgxNg=="} +{"index":346,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1408.5093","reference_contexts":"The entire network can still be trained end-to-end by SGD with backpropagation, and can be easily implemented using common libraries (e.g., Caffe [19] ) without modifying the solvers.","reference_id":"MTM4ODE3MTExODkxNzg0NTk2MzA="} +{"index":347,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.5185","reference_contexts":"Strong regularization such as maxout [10] or dropout [14] is applied to obtain the best results ( [10, 25, 24, 35] ) on this dataset","reference_id":"NDI2OTQwNDg0OTU0NjMyNDA2Mg=="} +{"index":348,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.5185","reference_contexts":"On this dataset we use identity shortcuts in all cases (i.e., option A), method error (%) Maxout [10] 9.38 NIN [25] 8.81 DSN [24] 8.22 # layers # params FitNet [ so our residual models have exactly the same depth, width, and number of parameters as the plain counterparts","reference_id":"NTA1MjYyMzk1NzI5Mjg2MDUz"} +{"index":349,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.5185","reference_contexts":"In [44, 24] , a few intermediate layers are directly connected to auxiliary classifiers for addressing vanishing/exploding gradients","reference_id":"MTIzMTAxNDcwMzc4MDE3MDA5NTA="} +{"index":350,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1312.4400","reference_contexts":"Strong regularization such as maxout [10] or dropout [14] is applied to obtain the best results ( [10, 25, 24, 35] ) on this dataset","reference_id":"NDI2OTQwNDg0OTU0NjMyNDA2Mg=="} +{"index":351,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1312.4400","reference_contexts":"On this dataset we use identity shortcuts in all cases (i.e., option A), method error (%) Maxout [10] 9.38 NIN [25] 8.81 DSN [24] 8.22 # layers # params FitNet [ so our residual models have exactly the same depth, width, and number of parameters as the plain counterparts","reference_id":"NTA1MjYyMzk1NzI5Mjg2MDUz"} +{"index":352,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1504.06066","reference_contexts":"We select two adjacent scales from the pyramid following [33]","reference_id":"MTc3MTEzMjAxNDM0NTA5MDIyMTg="} +{"index":353,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1504.06066","reference_contexts":"We adopt the idea of \"Networks on Conv feature maps\" (NoC) [33] to address this issue","reference_id":"NzIxNjU2NDYzNDIyMjY3MDU2"} +{"index":354,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1504.06066","reference_contexts":"Multi-scale training/testing has been developed in [12, 7] by selecting a scale from a feature pyramid, and in [33] by using maxout layers","reference_id":"MTE3MDYxMjg0NDA5MDE2NjU4ODM="} +{"index":355,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.0575","reference_contexts":"On the ImageNet classification dataset [36] , we obtain excellent results by extremely deep residual nets","reference_id":"MTM4MDMxMjYzMDQyMDE0ODI1Njk="} +{"index":356,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.0575","reference_contexts":"The ImageNet Localization (LOC) task [36] requires to classify and localize the objects","reference_id":"Njg2MTI4MzE5NTA0NzYxNDg2Nw=="} +{"index":357,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.0575","reference_contexts":"We evaluate our method on the ImageNet 2012 classification dataset [36] that consists of 1000 classes","reference_id":"MjgyNzI1OTc3NDY4ODExNzg1Nw=="} +{"index":358,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.0575","reference_contexts":"Recent evidence [41, 44] reveals that network depth is of crucial importance, and the leading results [41, 44, 13, 16] on the challenging ImageNet dataset [36] all exploit \"very deep\" [41] models, with a depth of sixteen [41] to thirty [16]","reference_id":"MTIzMTg1ODU1MjU3MjkzOTU4OA=="} +{"index":359,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1409.0575","reference_contexts":"We present comprehensive experiments on ImageNet [36] to show the degradation problem and evaluate our method","reference_id":"MTI1ODA5MTg4OTQyNzM0NDgwNDQ="} +{"index":360,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1312.6120","reference_contexts":"This problem, however, has been largely addressed by normalized initialization [23, 9, 37, 13] and intermediate normalization layers [16] , which enable networks with tens of layers to start converging for stochastic gradient descent (SGD) with backpropagation [22] .","reference_id":"OTQ0Mjg2MTc1NjM3NTU2OTA5NQ=="} +{"index":361,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1505.00387","reference_contexts":"Unexpectedly, such degradation is not caused by overfitting, and adding more layers to a suitably deep model leads to higher training error, as reported in [11, 42] and thoroughly verified by our experiments","reference_id":"MTgxMzgzOTE5MjMzNzcwMTQxMTA="} +{"index":362,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1505.00387","reference_contexts":"4 , left) and on MNIST (see [42] ), suggesting that such an optimization difficulty is a fundamental problem.","reference_id":"NDM2NjQ5NzYxNzgyODY1NTk0Mg=="} +{"index":363,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1505.00387","reference_contexts":"It has fewer parameters than other deep and thin networks such as FitNet [35] and Highway [42] (Table 6 ), yet is among the state-of-the-art results (6.43%, Table 6 ).","reference_id":"MjQ3Mzc0MjI5MDMwMTkyMzI5Ng=="} +{"index":364,"citer_arxiv_id":"1512.03385","cited_arxiv_id":"1505.00387","reference_contexts":"Concurrent with our work, \"highway networks\" [42, 43] present shortcut connections with gating functions [15]","reference_id":"MTgyODY1OTI3Mjk3MzgyODgx"} +{"index":365,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1502.01852","reference_contexts":"The computational cost of Inception is also much lower than VGGNet or its higher performing successors [6]","reference_id":"MTc1NjE2MDM0MDE4ODM4Nzc1ODY="} +{"index":366,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1502.01852","reference_contexts":"Still our solution uses much less computation than the best published results based on denser networks: our model outperforms the results of He et al [6] -cutting the top-5 (top-1) error by 25% (14%) relative, respectively -while being six times cheaper computationally and using at least five times less parameters (estimated)","reference_id":"MTIxMjkwMjgxNjk2NDgwOTQyOTQ="} +{"index":367,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1509.09308","reference_contexts":"It is certainly possible to mitigate parts of these issues by applying specialized solutions to target memory use [2] , [15] or by optimizing the execution of certain operations via computational tricks [10]","reference_id":"MTU0Nzg2ODE1ODM1NTAyMTM4MzQ="} +{"index":368,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1409.5185","reference_contexts":"Also Lee et al [11] argues that auxiliary classifiers promote more stable learning and better convergence","reference_id":"ODYxMjY1Mzc0MDk1Mzc4NTkwNA=="} +{"index":369,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1211.5063","reference_contexts":"In addition, gradient clipping [14] with threshold 2.0 was found to be useful to stabilize the training","reference_id":"MTY0NTUzMzY3MzY1NzU4NTMwOA=="} +{"index":370,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1503.03832","reference_contexts":"This has made it feasible to utilize Inception networks in big-data scenarios [17] , [13] , where huge amount of data needed to be processed at reasonable cost or scenarios where memory or computational capacity is inherently limited, for example in mobile vision settings","reference_id":"Njc1NjQwNzA0NTY5MTE4ODIxOQ=="} +{"index":371,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1409.1556","reference_contexts":"VGGNet [18] and GoogLeNet [20] yielded simi-larly high performance in the 2014 ILSVRC [16] classification challenge","reference_id":"MTI2MzYyODUwNzExNzUwMjgwOQ=="} +{"index":372,"citer_arxiv_id":"1512.00567","cited_arxiv_id":"1409.1556","reference_contexts":"Although VGGNet [18] has the compelling feature of architectural simplicity, this comes at a high cost: evaluating the network requires a lot of computation","reference_id":"MTczMjk3NDQ5NDEyOTQ3NTEyNTQ="} +{"index":373,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1508.00657","reference_contexts":"There is an increased interest in incorporating characterlevel inputs to build word embeddings for various NLP problems, including part-of-speech tagging, parsing and language modeling (Ling et al., 2015; Kim et al., 2015; Ballesteros et al., 2015)","reference_id":"ODIyMTQzOTg2MzQwNDAyMDg1NA=="} +{"index":374,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1312.3005","reference_contexts":"Much work has been done on small and large scale RNN-based LMs (Mikolov et al., 2010; Mikolov, 2012; Chelba et al., 2013; Zaremba et al., 2014; Williams et al., 2015; Ji et al., 2015a; Wang & Cho, 2015; Ji et al., 2015b)","reference_id":"NDI3MjQ0MDY4NzM1NjM3MzIzMA=="} +{"index":375,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1312.3005","reference_contexts":"Indeed, most recent work on large scale LM has shown that RNNs are great in combination with N-grams, as they may have different strengths that complement N-gram models, but worse when considered in isolation (Mikolov et al., 2011; Mikolov, 2012; Chelba et al., 2013; Williams et al., 2015; Ji et al., 2015a; Shazeer et al., 2015) .","reference_id":"NDY2MTkwODA3MjE4MzkzNTk5MA=="} +{"index":376,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1312.3005","reference_contexts":"We focused on a well known, large scale LM benchmark: the One Billion Word Benchmark data set (Chelba et al., 2013)","reference_id":"MTY0NjY0MzQyODE5NTM0MTk0MDg="} +{"index":377,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1312.3005","reference_contexts":"The experiments are performed on the 1B Word Benchmark data set introduced by (Chelba et al., 2013) , which is a publicly available benchmark for measuring progress of statistical language modeling","reference_id":"OTExNzc3ODgyNzkyNDM3MTU5NQ=="} +{"index":378,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1406.1078","reference_contexts":"Lastly, recently proposed sequence-to-sequence models employ conditional language models (Mikolov & Zweig, 2012) as their key component to solve diverse tasks like machine translation (Sutskever et al., 2014; Cho et al., 2014; Kalchbrenner et al., 2014) or video generation (Srivastava et al., 2015a) .","reference_id":"NDI4NjY4NjMzMDM4OTg3NjY3OQ=="} +{"index":379,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1512.00103","reference_contexts":"This has shown to help when using byte-level input embeddings for named entity recognition (Gillick et al., 2015) , and we hope it will enable similar gains when used to map onto words.","reference_id":"NzQzOTA1MDc2MzE5ODYzMTgzNA=="} +{"index":380,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1308.0850","reference_contexts":"A class of models that solve this problem more efficiently are character-level LSTMs (Sutskever et al., 2011; Graves, 2013)","reference_id":"MTQxMzQ5MjMzNDQ4NzMwODI3NjY="} +{"index":381,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1308.0850","reference_contexts":"On the other hand, these models are more difficult to train and seem to perform worse even in small tasks like PTB (Graves, 2013) .","reference_id":"MTMzMTYzNDYyOTA0NzIwODcyOTA="} +{"index":382,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1511.06909","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":383,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1511.03962","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":384,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1404.2188","reference_contexts":"Lastly, recently proposed sequence-to-sequence models employ conditional language models (Mikolov & Zweig, 2012) as their key component to solve diverse tasks like machine translation (Sutskever et al., 2014; Cho et al., 2014; Kalchbrenner et al., 2014) or video generation (Srivastava et al., 2015a) .","reference_id":"NDI4NjY4NjMzMDM4OTg3NjY3OQ=="} +{"index":385,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1508.06615","reference_contexts":"Recent efforts on small scale language modeling have used CNN character embeddings for the input embeddings (Kim et al., 2015)","reference_id":"MTg1MTM1NTUyODMwMzgxODI1NQ=="} +{"index":386,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1508.06615","reference_contexts":"For any model using character embedding CNNs, we closely follow the architecture from (Kim et al., 2015)","reference_id":"NzEzNjMxNzkwODA2OTQ4MTM3Nw=="} +{"index":387,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1508.06615","reference_contexts":"There is an increased interest in incorporating characterlevel inputs to build word embeddings for various NLP problems, including part-of-speech tagging, parsing and language modeling (Ling et al., 2015; Kim et al., 2015; Ballesteros et al., 2015)","reference_id":"ODIyMTQzOTg2MzQwNDAyMDg1NA=="} +{"index":388,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1508.06615","reference_contexts":"In (Kim et al., 2015) , the words characters are processed by a 1-d CNN (Le Cun et al., 1990) with max-pooling across the sequence for each convolutional feature","reference_id":"Njc1NzM5MTM1OTMwMjY1MTAyOA=="} +{"index":389,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1508.02096","reference_contexts":"The approach proposed in (Ling et al., 2015) builds word embeddings using bidirectional LSTMs (Schuster & Paliwal, 1997; Graves & Schmidhuber, 2005) over the characters","reference_id":"MTE3MjU4NzA2Njk3NTM5MzMxMTk="} +{"index":390,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1508.02096","reference_contexts":"There is an increased interest in incorporating characterlevel inputs to build word embeddings for various NLP problems, including part-of-speech tagging, parsing and language modeling (Ling et al., 2015; Kim et al., 2015; Ballesteros et al., 2015)","reference_id":"ODIyMTQzOTg2MzQwNDAyMDg1NA=="} +{"index":391,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1410.8206","reference_contexts":"This may be useful for other problems such as Machine Translation where handling out-of-vocabulary words is very important (Luong et al., 2014)","reference_id":"MTM0MDkzNzQyOTA2MTk1MzExMTE="} +{"index":392,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1211.5063","reference_contexts":"We clip the gradients of the LSTM weights such that their norm is bounded by 1.0 (Pascanu et al., 2012) .","reference_id":"MTM5NzgwMTQ3OTY0Mzg5NTQxNjY="} +{"index":393,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1509.00685","reference_contexts":"LMs have played a key role in traditional NLP tasks such as speech recognition (Mikolov et al., 2010; Arisoy et al., 2012) , machine translation (Schwenk et al., 2012; Vaswani et al.) , or text summarization (Rush et al., 2015; Filippova et al., 2015)","reference_id":"MTA0NjgxNDgzNjkwNTI2MTk5Nzc="} +{"index":394,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1507.04808","reference_contexts":"For example, when trained on movie subtitles (Serban et al., 2015; Vinyals & Le, 2015) , these language models are able to generate basic answers to questions about object colors, facts about people, etc","reference_id":"MTgzMzExMjMwMDkxOTg0MjQ5Mzg="} +{"index":395,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1502.04681","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":396,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1506.05869","reference_contexts":"For example, when trained on movie subtitles (Serban et al., 2015; Vinyals & Le, 2015) , these language models are able to generate basic answers to questions about object colors, facts about people, etc","reference_id":"MTgzMzExMjMwMDkxOTg0MjQ5Mzg="} +{"index":397,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1511.03729","reference_contexts":"Much work has been done on small and large scale RNN-based LMs (Mikolov et al., 2010; Mikolov, 2012; Chelba et al., 2013; Zaremba et al., 2014; Williams et al., 2015; Ji et al., 2015a; Wang & Cho, 2015; Ji et al., 2015b)","reference_id":"NDI3MjQ0MDY4NzM1NjM3MzIzMA=="} +{"index":398,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1409.2329","reference_contexts":"Much work has been done on small and large scale RNN-based LMs (Mikolov et al., 2010; Mikolov, 2012; Chelba et al., 2013; Zaremba et al., 2014; Williams et al., 2015; Ji et al., 2015a; Wang & Cho, 2015; Ji et al., 2015b)","reference_id":"NDI3MjQ0MDY4NzM1NjM3MzIzMA=="} +{"index":399,"citer_arxiv_id":"1602.02410","cited_arxiv_id":"1409.2329","reference_contexts":"Following (Zaremba et al., 2014) we use dropout (Srivastava, 2013) before and after every LSTM layer","reference_id":"NjQyMDc1MzA3MDc0NTIwNzgzMQ=="} +{"index":400,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2112.05219","reference_contexts":"Recently, text-driven image manipulation has achieved significant progress using GANs [15, 8, [19] [20] [21] , which are known for their high-quality generation, in tandem with CLIP [32] , which consists of a semantically rich joint image-text representation, trained over millions of text-image pairs","reference_id":"MTUyODI0MzgyODAyMjU0MjQ4ODU="} +{"index":401,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2206.02779","reference_contexts":"To overcome this, several works [28, 4] assume that the user provides a mask to restrict the area in which the changes are applied.","reference_id":"Mjc2MTQ5MzMzMzEzMzk4ODc4Ng=="} +{"index":402,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2206.02779","reference_contexts":"To circumvent this, LLI-based methods [28, 4, 33] require the user to explicitly mask a part of the image to be inpainted, and drive the edited image to change in the masked area only, while matching the background of the original image","reference_id":"Njk3ODc0NTY0MTUwMzU4MzU1MQ=="} +{"index":403,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2204.02491","reference_contexts":"[6] proposed a text-based localized editing technique without using any mask, showing impressive results","reference_id":"MTYzNDUzOTU0NDc4OTgwOTEzODg="} +{"index":404,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"1809.11096","reference_contexts":"This model is conditioned on text embeddings via both cross-attention layers at resolutions [16, 8] and hybrid-attention layers at resolutions [32, 16, 8] of the downsampling and upsampling within the U-Net.","reference_id":"NTY5MDgxMDI4MTQyMzE0NjYxNg=="} +{"index":405,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"1809.11096","reference_contexts":"Recently, text-driven image manipulation has achieved significant progress using GANs [15, 8, [19] [20] [21] , which are known for their high-quality generation, in tandem with CLIP [32] , which consists of a semantically rich joint image-text representation, trained over millions of text-image pairs","reference_id":"MTUyODI0MzgyODAyMjU0MjQ4ODU="} +{"index":406,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2204.08583","reference_contexts":"[9] use VQ-GAN [12] , trained over diverse data, as a backbone","reference_id":"MTAzNjkyMDQzMTY1MjM4ODE3ODc="} +{"index":407,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2204.08583","reference_contexts":"Numerous works [11, 16, 42, 25, 26, 30, 31, 34, 49, 9, 13, 36] significantly advanced the generation of images conditioned on plain text, known as text-to-image synthesis","reference_id":"MzM5NDY0NzIzODc1NzE2NjIxNA=="} +{"index":408,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2204.08583","reference_contexts":"While most works that require only text (i.e., no masks) are limited to global editing [9, 23] , Bar-Tal et al","reference_id":"OTgwNDkwOTkyNzg3MzUzMjIwNQ=="} +{"index":409,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2203.13131","reference_contexts":"Numerous works [11, 16, 42, 25, 26, 30, 31, 34, 49, 9, 13, 36] significantly advanced the generation of images conditioned on plain text, known as text-to-image synthesis","reference_id":"MzM5NDY0NzIzODc1NzE2NjIxNA=="} +{"index":410,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2108.00946","reference_contexts":"Seminal works [29, 14, 46, 2] which combined these components were revolutionary, since they did not require extra manual labor, and produced \"lemon cake.\" \"cheese cake.\" \"apple cake.\" \"chocolate cake.\" \"pistachio cake.\" \"jello cake.\" \"matcha cake.\" \"pumpkin cake.\" \"pepperoni cake.\" \"fish cake.\" \"rice cake.\" \"pasta cake.\" \"beet cake.\" \"lego cake.\" \"brick cake.\" \"monster cake.\" \"lego cake.\" \"brick cake.\" \"monster cake.\" \"lemon cake.\" \"cheese cake.\" \"apple cake.\" \"chocolate cake.\" \"pistachio cake.\" \"jello cake.\" \"matcha cake.\" \"pumpkin cake.\" \"pepperoni cake.\" \"fish cake.\" \"rice cake.\" \"pasta cake.\" \"beet cake.\"","reference_id":"MTUzMDkxMTYwOTk0MjEyNzg2NTU="} +{"index":411,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2112.00374","reference_contexts":"While most works that require only text (i.e., no masks) are limited to global editing [9, 23] , Bar-Tal et al","reference_id":"OTgwNDkwOTkyNzg3MzUzMjIwNQ=="} +{"index":412,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2112.10741","reference_contexts":"To overcome this, several works [28, 4] assume that the user provides a mask to restrict the area in which the changes are applied.","reference_id":"Mjc2MTQ5MzMzMzEzMzk4ODc4Ng=="} +{"index":413,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2112.10741","reference_contexts":"To circumvent this, LLI-based methods [28, 4, 33] require the user to explicitly mask a part of the image to be inpainted, and drive the edited image to change in the masked area only, while matching the background of the original image","reference_id":"Njk3ODc0NTY0MTUwMzU4MzU1MQ=="} +{"index":414,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2112.10741","reference_contexts":"Imagen [38] , similar to GLIDE [28] , conditions on the text prompt in the noise prediction of each diffusion step (see appendix A.2) through two types of attention layers: i) cross-attention layers","reference_id":"Nzk3MDYzNjM1NDI4Mzk5Mjc4OA=="} +{"index":415,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2103.17249","reference_contexts":"Seminal works [29, 14, 46, 2] which combined these components were revolutionary, since they did not require extra manual labor, and produced \"lemon cake.\" \"cheese cake.\" \"apple cake.\" \"chocolate cake.\" \"pistachio cake.\" \"jello cake.\" \"matcha cake.\" \"pumpkin cake.\" \"pepperoni cake.\" \"fish cake.\" \"rice cake.\" \"pasta cake.\" \"beet cake.\" \"lego cake.\" \"brick cake.\" \"monster cake.\" \"lego cake.\" \"brick cake.\" \"monster cake.\" \"lemon cake.\" \"cheese cake.\" \"apple cake.\" \"chocolate cake.\" \"pistachio cake.\" \"jello cake.\" \"matcha cake.\" \"pumpkin cake.\" \"pepperoni cake.\" \"fish cake.\" \"rice cake.\" \"pasta cake.\" \"beet cake.\"","reference_id":"MTUzMDkxMTYwOTk0MjEyNzg2NTU="} +{"index":416,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2103.00020","reference_contexts":"This model is conditioned on text embeddings via both cross-attention layers at resolutions [16, 8] and hybrid-attention layers at resolutions [32, 16, 8] of the downsampling and upsampling within the U-Net.","reference_id":"NTY5MDgxMDI4MTQyMzE0NjYxNg=="} +{"index":417,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2103.00020","reference_contexts":"Recently, text-driven image manipulation has achieved significant progress using GANs [15, 8, [19] [20] [21] , which are known for their high-quality generation, in tandem with CLIP [32] , which consists of a semantically rich joint image-text representation, trained over millions of text-image pairs","reference_id":"MTUyODI0MzgyODAyMjU0MjQ4ODU="} +{"index":418,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2204.06125","reference_contexts":"Several large-scale text-image models have recently emerged, such as Imagen [38] , DALL-E2 [33] , and Parti [48] , demonstrating unprecedented semantic generation","reference_id":"NTg3OTc4MjI4NzI5NzUzNTc2NA=="} +{"index":419,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2204.06125","reference_contexts":"To circumvent this, LLI-based methods [28, 4, 33] require the user to explicitly mask a part of the image to be inpainted, and drive the edited image to change in the masked area only, while matching the background of the original image","reference_id":"Njk3ODc0NTY0MTUwMzU4MzU1MQ=="} +{"index":420,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2204.06125","reference_contexts":"Recently, large-scale language-image (LLI) models, such as Imagen [38] , DALL•E 2 [33] and Parti [48] , have shown phenomenal generative semantic and compositional power, and gained unprecedented attention from the research community and the public eye","reference_id":"NzA5OTYwMTkyMTk0NDk4OTg4Mg=="} +{"index":421,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2205.11487","reference_contexts":"We thank Noa Glaser , Adi Zicher , Yaron Brodsky and Shlomi Fruchter for their valuable inputs that helped improve this work, and to Mohammad Norouzi , Chitwan Saharia and William Chan for providing us with their support and the pretrained models of Imagen [38]","reference_id":"MTQ2MjU5OTY5ODczMzUwMTU4ODE="} +{"index":422,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2205.11487","reference_contexts":"Imagen [38] , similar to GLIDE [28] , conditions on the text prompt in the noise prediction of each diffusion step (see appendix A.2) through two types of attention layers: i) cross-attention layers","reference_id":"Nzk3MDYzNjM1NDI4Mzk5Mjc4OA=="} +{"index":423,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2205.11487","reference_contexts":"Let I be an image which was generated by a text-guided diffusion model [38] using the text prompt P and a random seed s","reference_id":"MTgwMTkzMzQxMDI4MDc2MTM2NjE="} +{"index":424,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2205.11487","reference_contexts":"Imagen [38] consists of three text-conditioned diffusion models: A text-to-image 64 × 64 model, and two super-resolution models -64 × 64 → 256 × 256 and 256 × 256 → 1024 × 1024","reference_id":"MTQ1ODU0NTAwNTg2OTU0NzExMg=="} +{"index":425,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2205.11487","reference_contexts":"Recently, large-scale language-image (LLI) models, such as Imagen [38] , DALL•E 2 [33] and Parti [48] , have shown phenomenal generative semantic and compositional power, and gained unprecedented attention from the research community and the public eye","reference_id":"NzA5OTYwMTkyMTk0NDk4OTg4Mg=="} +{"index":426,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2205.11487","reference_contexts":"Several large-scale text-image models have recently emerged, such as Imagen [38] , DALL-E2 [33] , and Parti [48] , demonstrating unprecedented semantic generation","reference_id":"NTg3OTc4MjI4NzI5NzUzNTc2NA=="} +{"index":427,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2205.11487","reference_contexts":"We use the Imagen [38] text-guided synthesis model as a backbone","reference_id":"Mjc3MjM0OTg0MjUxNzI0NDYyNg=="} +{"index":428,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2008.05865","reference_contexts":"Numerous works [11, 16, 42, 25, 26, 30, 31, 34, 49, 9, 13, 36] significantly advanced the generation of images conditioned on plain text, known as text-to-image synthesis","reference_id":"MzM5NDY0NzIzODc1NzE2NjIxNA=="} +{"index":429,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2102.02766","reference_contexts":"This process, known as inversion, has recently drawn considerable attention for GANs, e.g., [51, 1, 3, 35, 50, 43, 45, 47] , but has not yet been fully addressed for text-guided diffusion models.","reference_id":"MTE3NDE3MTQxMDA2ODYwNDk4MDU="} +{"index":430,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2102.02766","reference_contexts":"This is partially due to a distortion-editability tradeoff [43] , where we recognize that reducing the classifier-free guidance [18] parameter (i.e., reducing the prompt influence) improves reconstruction but constrains our ability to perform significant manipulations.","reference_id":"MTM2NzMxOTU2MzczMTcwMjI3MTg="} +{"index":431,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2109.06590","reference_contexts":"This process, known as inversion, has recently drawn considerable attention for GANs, e.g., [51, 1, 3, 35, 50, 43, 45, 47] , but has not yet been fully addressed for text-guided diffusion models.","reference_id":"MTE3NDE3MTQxMDA2ODYwNDk4MDU="} +{"index":432,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2206.10789","reference_contexts":"Several large-scale text-image models have recently emerged, such as Imagen [38] , DALL-E2 [33] , and Parti [48] , demonstrating unprecedented semantic generation","reference_id":"NTg3OTc4MjI4NzI5NzUzNTc2NA=="} +{"index":433,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2206.10789","reference_contexts":"Recently, large-scale language-image (LLI) models, such as Imagen [38] , DALL•E 2 [33] and Parti [48] , have shown phenomenal generative semantic and compositional power, and gained unprecedented attention from the research community and the public eye","reference_id":"NzA5OTYwMTkyMTk0NDk4OTg4Mg=="} +{"index":434,"citer_arxiv_id":"2208.01626","cited_arxiv_id":"2004.00049","reference_contexts":"This process, known as inversion, has recently drawn considerable attention for GANs, e.g., [51, 1, 3, 35, 50, 43, 45, 47] , but has not yet been fully addressed for text-guided diffusion models.","reference_id":"MTE3NDE3MTQxMDA2ODYwNDk4MDU="} +{"index":435,"citer_arxiv_id":"2207.12598","cited_arxiv_id":"2105.05233","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":436,"citer_arxiv_id":"2207.12598","cited_arxiv_id":"2106.15282","reference_contexts":"BigGAN-deep, max IS (Brock et al., 2019) 25 253 BigGAN-deep (Brock et al., 2019) 5.7 124.5 CDM (Ho et al., 2021) 3.52 128.8 LOGAN (Wu et al., 2019) 3.36 148.2 ADM-G (Dhariwal & Nichol, 2021) 2.97 -Ours T = 128/256/1024 w = 0.0 8.11 / 7.27 / 7.22 81.46 / 82.45 / 81.54 w = 0.1 5.31 / 4.53 / 4.5 105.01 / 106.12 / 104.67 w = 0.2 3.7 / 3.03 / 3 130.79 / 132.54 / 130.09 w = 0.3 3.04 / 2.43 / 2.43 156.09 / 158.47 / 156 w = 0.4 3.02 / 2.49 / 2.48 183.01 / 183.41 / 180.88 w = 0.5 3.43 / 2.98 / 2.96 206.94 / 207.98 / 204.31 w = 0.6 4.09 / 3.76 / 3.73 227.72 / 228.83 / 226.76 w = 0.7 4.96 / 4.67 / 4.69 247.92 / 249.25 / 247.89 w = 0.8 5.93 / 5.74 / 5.71 265.54 / 267.99 / 265.52 w = 0.9 6.89 / 6.8 / 6.81 280.19 / 283.41 / 281.14 w = 1.0 7.88 / 7.86 / 7.8 295.29 / 297.98 / 294.56 w = 2.0 15.9 / 15.93 / 15.75 378.56 / 377.37 / 373.18 w = 3.0 19.77 / 19.77 / 19.56 409.16 / 407.44 / 405.68 w = 4.0 21.55 / 21.53 / 21.45 422.29 / 421.03 / 419.06 Table 2: ImageNet 128x128 results (w = 0.0 refers to non-guided models)","reference_id":"NTc2MDAxNDY2NzEzMzM3NTc4NQ=="} +{"index":437,"citer_arxiv_id":"2207.12598","cited_arxiv_id":"2106.15282","reference_contexts":"These models have delivered audio synthesis performance rivaling the quality of autoregressive models with substantially fewer inference steps (Chen et al., 2021; Kong et al., 2021) , and they have delivered ImageNet generation results outperforming BigGAN-deep (Brock et al., 2019) and VQ-VAE-2 (Razavi et al., 2019) in terms of FID score and classification accuracy score (Ho et al., 2021; Dhariwal & Nichol, 2021)","reference_id":"NDEwNDc5OTE3MzY2NDE5ODk2OQ=="} +{"index":438,"citer_arxiv_id":"2207.12598","cited_arxiv_id":"2106.15282","reference_contexts":"ADM (Dhariwal & Nichol, 2021) 2.07 -CDM (Ho et al., 2021) 1","reference_id":"MTgwNTQyMjc4NjU4NjAzMDI4MjA="} +{"index":439,"citer_arxiv_id":"2207.12598","cited_arxiv_id":"2107.00630","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":440,"citer_arxiv_id":"2207.12598","cited_arxiv_id":"1912.00953","reference_contexts":"BigGAN-deep, max IS (Brock et al., 2019) 25 253 BigGAN-deep (Brock et al., 2019) 5.7 124.5 CDM (Ho et al., 2021) 3.52 128.8 LOGAN (Wu et al., 2019) 3.36 148.2 ADM-G (Dhariwal & Nichol, 2021) 2.97 -Ours T = 128/256/1024 w = 0.0 8.11 / 7.27 / 7.22 81.46 / 82.45 / 81.54 w = 0.1 5.31 / 4.53 / 4.5 105.01 / 106.12 / 104.67 w = 0.2 3.7 / 3.03 / 3 130.79 / 132.54 / 130.09 w = 0.3 3.04 / 2.43 / 2.43 156.09 / 158.47 / 156 w = 0.4 3.02 / 2.49 / 2.48 183.01 / 183.41 / 180.88 w = 0.5 3.43 / 2.98 / 2.96 206.94 / 207.98 / 204.31 w = 0.6 4.09 / 3.76 / 3.73 227.72 / 228.83 / 226.76 w = 0.7 4.96 / 4.67 / 4.69 247.92 / 249.25 / 247.89 w = 0.8 5.93 / 5.74 / 5.71 265.54 / 267.99 / 265.52 w = 0.9 6.89 / 6.8 / 6.81 280.19 / 283.41 / 281.14 w = 1.0 7.88 / 7.86 / 7.8 295.29 / 297.98 / 294.56 w = 2.0 15.9 / 15.93 / 15.75 378.56 / 377.37 / 373.18 w = 3.0 19.77 / 19.77 / 19.56 409.16 / 407.44 / 405.68 w = 4.0 21.55 / 21.53 / 21.45 422.29 / 421.03 / 419.06 Table 2: ImageNet 128x128 results (w = 0.0 refers to non-guided models)","reference_id":"NTc2MDAxNDY2NzEzMzM3NTc4NQ=="} +{"index":441,"citer_arxiv_id":"1609.04802","cited_arxiv_id":"1605.02688","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":442,"citer_arxiv_id":"1609.04802","cited_arxiv_id":"1608.05148","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":443,"citer_arxiv_id":"1609.04802","cited_arxiv_id":"1607.07539","reference_contexts":"[64] for inpainting","reference_id":"ODM5MDU5Njk1MDE2NjU1NDMyNQ=="} +{"index":444,"citer_arxiv_id":"1805.08306","cited_arxiv_id":"1412.6980","reference_contexts":"All experiments presented so far were implemented in Tensorflow (Abadi et al., 2016) using Adam (Kingma & Ba, 2014) as the optimizer with the learning rate of 0.001.","reference_id":"NjEwOTY1OTI5OTQxNDIwMjQ4MA=="} +{"index":445,"citer_arxiv_id":"1805.08306","cited_arxiv_id":"1312.6114","reference_contexts":"Variational autoencoders were designed to tackle this problem with a variational objective (Kingma & Welling, 2013)","reference_id":"MTU3MjQ3ODk3NDAwNTA3NjM3Nzk="} +{"index":446,"citer_arxiv_id":"1805.08306","cited_arxiv_id":"1711.09268","reference_contexts":"But there are important new developments on the HMC front (Girolami & Calderhead, 2011; Levy et al., 2017) , which DEEN can build on","reference_id":"NjQxNjE3MDU2MzUyODU4NTAxNg=="} +{"index":447,"citer_arxiv_id":"1805.08306","cited_arxiv_id":"1707.00797","reference_contexts":"Stein contrastive divergence-In preparing this submission, we came across the work by Liu & Wang (2017) and the Stein contrastive divergence introduced there for learning energy models.","reference_id":"MTE1OTQ0MjE1NTkzNTcxNTMzNzI="} +{"index":448,"citer_arxiv_id":"1805.08306","cited_arxiv_id":"1206.6464","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":449,"citer_arxiv_id":"1805.08306","cited_arxiv_id":"1511.01844","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":450,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1911.11544","reference_contexts":"[1] focus on inverting StyleGAN [19] models by exploiting the layer-wise noises","reference_id":"MTQ4MTI3MTk2ODIxNTg1ODc2MzM="} +{"index":451,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1911.11544","reference_contexts":"Here, the semantics refer to the emergent knowledge that GAN has learned from the observed data [11, 16, 30, 32]","reference_id":"MzQ0MjczNzc3MDY4MDQ0MDMyNg=="} +{"index":452,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1703.10717","reference_contexts":"Many variants of GANs are proposed to improve the synthesis quality [25, 34, 18, 7, 19] and training stability [2, 14, 6]","reference_id":"MjM3NDY2NjU1Nzg1NTEwNDE4Ng=="} +{"index":453,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1907.07171","reference_contexts":"Here, the semantics refer to the emergent knowledge that GAN has learned from the observed data [11, 16, 30, 32]","reference_id":"MzQ0MjczNzc3MDY4MDQ0MDMyNg=="} +{"index":454,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1907.07171","reference_contexts":"Recent work [11, 16, 30] has shown that GANs spontaneously learn to encode rich semantics inside the latent space and that varying the latent code leads to the manipulation of the corresponding attributes occurring in the output image","reference_id":"MTcxMzYzNzcwODg0MzA3ODk0MjI="} +{"index":455,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1907.07171","reference_contexts":"[16] achieved camera movements and color changes by shifting the latent distribution, Shen et al","reference_id":"MTIyNzE4MTgyMDU5NDk3MTcyMTA="} +{"index":456,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1907.07171","reference_contexts":"Prior work has shown that a well-trained GAN model is able to encode interpretable semantics inside the latent space [30, 16, 32]","reference_id":"MTMwNzM5MzIyODYwNzIzMDE5OA=="} +{"index":457,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1912.04958","reference_contexts":"[20] and Abdal et al","reference_id":"MTY1MzA1NjgzODY3OTYxNDI2MjU="} +{"index":458,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1912.04958","reference_contexts":"As a result, the disentangled space W is widely used for the GAN inversion task [29, 35, 1, 20]","reference_id":"MTEzNjUyMzgyMzY1NTE5MzYyMzM="} +{"index":459,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"2003.13659","reference_contexts":"[26] optimizes the parameters of the generator together with the latent code, Karras et al","reference_id":"Mzc3ODU1MzAzMzQ4MDk5MjAxNA=="} +{"index":460,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1911.09267","reference_contexts":"Here, the semantics refer to the emergent knowledge that GAN has learned from the observed data [11, 16, 30, 32]","reference_id":"MzQ0MjczNzc3MDY4MDQ0MDMyNg=="} +{"index":461,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1911.09267","reference_contexts":"As pointed out by prior work [30, 32] , GANs can learn rich semantics in the latent space, enabling image manipulation by linearly transforming the latent representation","reference_id":"MTMzNTgyMDMyNzEzNDcyNDc1MjI="} +{"index":462,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1911.09267","reference_contexts":"[32] observed that hierarchical semantics emerge from the layer-wise latent codes of GANs for scene synthesis","reference_id":"ODc3ODc0ODgxMDU2MzA4MDM3MA=="} +{"index":463,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1911.09267","reference_contexts":"Prior work has shown that a well-trained GAN model is able to encode interpretable semantics inside the latent space [30, 16, 32]","reference_id":"MTMwNzM5MzIyODYwNzIzMDE5OA=="} +{"index":464,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1506.03365","reference_contexts":"We conduct experiments on FFHQ dataset [19] , which contains 70,000 highquality face images, and LSUN dataset [33] , which consists of images from 10 different scene categories","reference_id":"MTUzNzI3OTU0MDE2NDY4MDczNQ=="} +{"index":465,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1506.03365","reference_contexts":"This appendix is organized as follows: In Sec.B, we show image reconstruction results using the model trained on LSUN bedroom dataset [33]","reference_id":"MjUwMTI1MjI2NDk5OTU1NjEzOQ=="} +{"index":466,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1906.08090","reference_contexts":"As a result, the disentangled space W is widely used for the GAN inversion task [29, 35, 1, 20]","reference_id":"MTEzNjUyMzgyMzY1NTE5MzYyMzM="} +{"index":467,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1906.08090","reference_contexts":"than Z space [35]","reference_id":"MTM1MzE2MTUzNDc3NDI5MzYyMDA="} +{"index":468,"citer_arxiv_id":"2004.00049","cited_arxiv_id":"1906.08090","reference_contexts":"There are also some models that take invertibility into account at the training stage by designing new architectures [10, 9, 35, 21]","reference_id":"MTUzMDcxMDA3NjI1MzQ4MzQ5MDM="} +{"index":469,"citer_arxiv_id":"1905.09883","cited_arxiv_id":"1905.11065","reference_contexts":"In particular, the use of SDEs to enhance the expressive power of continuous-time neural nets was proposed in a concurrent work of Peluchetti and Favaro (2019)","reference_id":"MTI4ODk3NjIzODMwOTg3MDI4MTU="} +{"index":470,"citer_arxiv_id":"1905.09883","cited_arxiv_id":"1902.02376","reference_contexts":"The code for all experiments was written in Julia using the DiffEqFlux library (Rackauckas et al., 2019) and executed on a CPU","reference_id":"MzYwNDgwMDE0NDIxMTU0NzQ4Mw=="} +{"index":471,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1603.06147","reference_contexts":"Since then, many novel techniques have been proposed to further improve NMT: using an attention mechanism to deal with rare words [37] , a mechanism to model translation coverage [42] , multi-task and semi-supervised training to incorporate more data [14, 29] , a character decoder [9] , a character encoder [11] , subword units [38] also to deal with rare word outputs, different kinds of attention mechanisms [30] , and sentence-level loss minimization [39, 34]","reference_id":"MTYyNTg0NjEwOTM3MDcyOTQ1ODU="} +{"index":472,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1603.06147","reference_contexts":"Another broad category of approaches is to use sub-word units, e.g., chararacters [10] , mixed word/characters [28] , or more intelligent sub-words [38] .","reference_id":"ODc3MDg2OTk1MzY2MjUyNzkxOA=="} +{"index":473,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1603.00810","reference_contexts":"Since then, many novel techniques have been proposed to further improve NMT: using an attention mechanism to deal with rare words [37] , a mechanism to model translation coverage [42] , multi-task and semi-supervised training to incorporate more data [14, 29] , a character decoder [9] , a character encoder [11] , subword units [38] also to deal with rare word outputs, different kinds of attention mechanisms [30] , and sentence-level loss minimization [39, 34]","reference_id":"MTYyNTg0NjEwOTM3MDcyOTQ1ODU="} +{"index":474,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1603.08148","reference_contexts":"One approach is to simply copy rare words from source to target (as most rare words are names or numbers where the correct translation is just a copy), either based on the attention model [37] , using an external alignment model [31] , or even using a more complicated special purpose pointing network [18]","reference_id":"MTQ2OTYzODQ5MzI2NTExNDAxODU="} +{"index":475,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1502.02551","reference_contexts":"Many of those previous studies [19, 20, 43, 27] however mostly focus on CNN models with relatively few layers","reference_id":"MTU2NTI1NzU4OTcyODYxNTMwOTg="} +{"index":476,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1510.00149","reference_contexts":"Many of those previous studies [19, 20, 43, 27] however mostly focus on CNN models with relatively few layers","reference_id":"MTU2NTI1NzU4OTcyODYxNTMwOTg="} +{"index":477,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1412.6980","reference_contexts":"For the first stage of maximum likelihood training (that is, to optimize for objective function 7), we use a combination of Adam [25] and simple SGD learning algorithms provided by the TensorFlow runtime system","reference_id":"MTM1MjE5OTI1MTU5MzcwOTE3OTQ="} +{"index":478,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1412.6980","reference_contexts":"The n replicas all share one copy of model parameters, with each replica asynchronously updating the parameters using a combination of Adam [25] and SGD algorithms","reference_id":"MTc3MTI3NjgzMDQ3MDAxNzQzMjA="} +{"index":479,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1605.04711","reference_contexts":"In [27] , it is demonstrated that neural network model weights can be quantized to only three states, -1, 0, and +1.","reference_id":"MTE3OTQ3NDMyODUyNTc5NjY4NTE="} +{"index":480,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1605.04711","reference_contexts":"Many of those previous studies [19, 20, 43, 27] however mostly focus on CNN models with relatively few layers","reference_id":"MTU2NTI1NzU4OTcyODYxNTMwOTg="} +{"index":481,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1604.00788","reference_contexts":"Another broad category of approaches is to use sub-word units, e.g., chararacters [10] , mixed word/characters [28] , or more intelligent sub-words [38] .","reference_id":"ODc3MDg2OTk1MzY2MjUyNzkxOA=="} +{"index":482,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1211.5063","reference_contexts":"However, simply stacking more layers of LSTM works only to a certain number of layers, beyond which the network becomes too slow and difficult to train, likely due to exploding and vanishing gradient problems [33, 22]","reference_id":"MjQ2NTk5MzM2MzMwMDU4MjkyMQ=="} +{"index":483,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1505.00387","reference_contexts":"Motivated by the idea of modeling differences between an intermediate layer's output and the targets, which has shown to work well for many projects in the past [16, 21, 40] , we introduce residual connections among the LSTM layers in a stack (see Figure 2 )","reference_id":"Mzc0MDE0NDM5MDgzOTIxNjEyNQ=="} +{"index":484,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1512.06473","reference_contexts":"Many of those previous studies [19, 20, 43, 27] however mostly focus on CNN models with relatively few layers","reference_id":"MTU2NTI1NzU4OTcyODYxNTMwOTg="} +{"index":485,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1512.06473","reference_contexts":"For example, in [43] , it is demonstrated that a convolutional neural network model can be sped up by a factor of 4-6 with minimal loss on classification accuracy on the ILSVRC-12 benchmark","reference_id":"NzU5MDIxNDk3Mzg2Mzg0MDA5NA=="} +{"index":486,"citer_arxiv_id":"1609.08144","cited_arxiv_id":"1606.04199","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":487,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1311.2524","reference_contexts":"Previous works [6, 13, 7] analyzed the semantic meaning of various units by finding the set of inputs that maximally activate a given unit","reference_id":"OTgyNDY1ODA4Mjc2MDM1MDkwNA=="} +{"index":488,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1311.2524","reference_contexts":"They look for input images which maximize the activation value of this single feature [6, 13, 7, 4] .","reference_id":"MTY0NTY5NTIwOTE1MTMxNzMyMjA="} +{"index":489,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1311.2524","reference_contexts":"So far, unit-level inspection methods had relatively little utility beyond confirming certain intuitions regarding the complexity of the representations learned by a deep neural network [6, 13, 7, 4]","reference_id":"MTQ2MjY3NDI3MjI3ODU5OTAzOTU="} +{"index":490,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1112.6209","reference_contexts":"For all the networks we studied (MNIST, QuocNet [10] , AlexNet [9] ), for each sample, we have always managed to generate very close, visually hard to distinguish, adversarial examples that are misclassified by the original network (see figure 5 and http://goo.gl/huaGPb for examples)","reference_id":"MTU2MzYzNjY4MjE0MTIxMjUxMTU="} +{"index":491,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1112.6209","reference_contexts":"• ∼ 10M image samples from Youtube (see [10] ) -Unsupervised trained network with ∼ 1 billion learnable parameters","reference_id":"MTcyMjE3NzkzNjk2ODg0MTUxNzU="} +{"index":492,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1301.3781","reference_contexts":"[12] for word representations, where the various directions in the vector space representing the words are shown to give rise to a surprisingly rich semantic encoding of relations and analogies","reference_id":"MTA3NzM0NDY1NzE2MjYzMzIwMzM="} +{"index":493,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1311.2901","reference_contexts":"Previous works [6, 13, 7] analyzed the semantic meaning of various units by finding the set of inputs that maximally activate a given unit","reference_id":"OTgyNDY1ODA4Mjc2MDM1MDkwNA=="} +{"index":494,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1311.2901","reference_contexts":"Already, a variety of recent state of the art computer vision models employ input deformations during training for increasing the robustness and convergence speed of the models [9, 13]","reference_id":"MTU4MTQ1NzEwNTAzMzQ5ODI4OTE="} +{"index":495,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1311.2901","reference_contexts":"They look for input images which maximize the activation value of this single feature [6, 13, 7, 4] .","reference_id":"MTY0NTY5NTIwOTE1MTMxNzMyMjA="} +{"index":496,"citer_arxiv_id":"1312.6199","cited_arxiv_id":"1311.2901","reference_contexts":"So far, unit-level inspection methods had relatively little utility beyond confirming certain intuitions regarding the complexity of the representations learned by a deep neural network [6, 13, 7, 4]","reference_id":"MTQ2MjY3NDI3MjI3ODU5OTAzOTU="} +{"index":497,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2201.07520","reference_contexts":"Generative transformer methods usually train autoregressive [4, 10, 11, 26, 37, 38] or non-autoregressive (masked) [9] transformers in discrete token space to model the generation process","reference_id":"ODg1NTE4Mjg4MzUyNzQ5OTg3NQ=="} +{"index":498,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2211.01324","reference_contexts":"The denoising process can occur either in pixel space or latent space, resulting in pixel diffusion models [5, 25, 30] or latent diffusion models [21, 27] , which feature higher efficiency by reducing the size of spatial features","reference_id":"OTUwOTU1NjA1NDYyMjQ3MjAzMA=="} +{"index":499,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2108.07258","reference_contexts":"First, before instruction-tuning, language models are more prone to generating low-quality text, which may be inconsistent in tone, overly verbose or concise, or simply unhelpful or even toxic [6, 7, 13, 20, 35] ; ChatGPT-level [2] performance is achieved with effective instruction-tuning [20]","reference_id":"NzMzNzI4MTkzMTMwNzUyMDA4"} +{"index":500,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2301.00704","reference_contexts":"Generative transformer methods usually train autoregressive [4, 10, 11, 26, 37, 38] or non-autoregressive (masked) [9] transformers in discrete token space to model the generation process","reference_id":"ODg1NTE4Mjg4MzUyNzQ5OTg3NQ=="} +{"index":501,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2208.01618","reference_contexts":"A number of techniques have been developed to personalize or adapt text-to-image models to a new subject or style [12, 15, 28]","reference_id":"NDEyMzA3NjY2NzYwNTkwMzI1NA=="} +{"index":502,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2009.11462","reference_contexts":"First, before instruction-tuning, language models are more prone to generating low-quality text, which may be inconsistent in tone, overly verbose or concise, or simply unhelpful or even toxic [6, 7, 13, 20, 35] ; ChatGPT-level [2] performance is achieved with effective instruction-tuning [20]","reference_id":"NzMzNzI4MTkzMTMwNzUyMDA4"} +{"index":503,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2106.09685","reference_contexts":"A number of techniques have been developed to personalize or adapt text-to-image models to a new subject or style [12, 15, 28]","reference_id":"NDEyMzA3NjY2NzYwNTkwMzI1NA=="} +{"index":504,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2302.03917","reference_contexts":"Recent advances in generative models have enabled them to generate various high-quality content, such as text [2, 33] , image [21, 25] , music [16] , video [32] , and even 3D scenes [19, 22, 34] , which has fundamentally revolutionized generative artificial intelligence (AI)","reference_id":"MTgyMzk1NzYxNjAxMTk2NTQ2MTk="} +{"index":505,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2305.01569","reference_contexts":"As argued in many recent papers (e.g., [18, 21] ), FID scores do not correlate well with human assessment of the performance of generative models.","reference_id":"NjM2MjQ1NjI0NjcxNzM0ODA3"} +{"index":506,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"The denoising process can occur either in pixel space or latent space, resulting in pixel diffusion models [5, 25, 30] or latent diffusion models [21, 27] , which feature higher efficiency by reducing the size of spatial features","reference_id":"OTUwOTU1NjA1NDYyMjQ3MjAzMA=="} +{"index":507,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"On the commonly used PartiPrompts and our Open User Input Prompts, we carefully evaluate our model against a publicly available state-of-the-art text-to-image model (SDXLv1.0 [21] ) as well as our pre-trained LDM","reference_id":"MTAzODA5Njk5NjA5NzQ2MDEzNzQ="} +{"index":508,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"As argued in many recent papers (e.g., [18, 21] ), FID scores do not correlate well with human assessment of the performance of generative models.","reference_id":"NjM2MjQ1NjI0NjcxNzM0ODA3"} +{"index":509,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"In fact, it significantly outperform a state-of-the-art publicly available model SDXLv1.0 [21] on visual appeal","reference_id":"MTgyMjg0Njc2MzMyMzEwOTE0Mg=="} +{"index":510,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"The model is trained with progressively increasing resolutions, similar to [21]","reference_id":"NTk1NTA0ODYwMDEyMDA1ODUxMA=="} +{"index":511,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"To place the visual appeal of our generated images in context with a current state-of-the-art model, we compare our model to SDXLv1.0 [21]","reference_id":"MTQxMTA0MjIwNzExODQyMjA5"} +{"index":512,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"Recent advances in generative models have enabled them to generate various high-quality content, such as text [2, 33] , image [21, 25] , music [16] , video [32] , and even 3D scenes [19, 22, 34] , which has fundamentally revolutionized generative artificial intelligence (AI)","reference_id":"MTgyMzk1NzYxNjAxMTk2NTQ2MTk="} +{"index":513,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2307.01952","reference_contexts":"To place the visual appeal of our generated images in the context of current state-of-the-art, we compare Emu with SDXLv1.0 [21]","reference_id":"NjcyMTg1NTc2MzQ5NzgxMTIz"} +{"index":514,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2209.14988","reference_contexts":"Recent advances in generative models have enabled them to generate various high-quality content, such as text [2, 33] , image [21, 25] , music [16] , video [32] , and even 3D scenes [19, 22, 34] , which has fundamentally revolutionized generative artificial intelligence (AI)","reference_id":"MTgyMzk1NzYxNjAxMTk2NTQ2MTk="} +{"index":515,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2204.06125","reference_contexts":"Recent advances in generative models have enabled them to generate various high-quality content, such as text [2, 33] , image [21, 25] , music [16] , video [32] , and even 3D scenes [19, 22, 34] , which has fundamentally revolutionized generative artificial intelligence (AI)","reference_id":"MTgyMzk1NzYxNjAxMTk2NTQ2MTk="} +{"index":516,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2204.06125","reference_contexts":"The denoising process can occur either in pixel space or latent space, resulting in pixel diffusion models [5, 25, 30] or latent diffusion models [21, 27] , which feature higher efficiency by reducing the size of spatial features","reference_id":"OTUwOTU1NjA1NDYyMjQ3MjAzMA=="} +{"index":517,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2301.09515","reference_contexts":"Generative adversarial network [17, 31] also show remarkable capability of generating realistic images","reference_id":"MTUwMjIzMzgyODcwMTIzMzg4MA=="} +{"index":518,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2209.14792","reference_contexts":"Recent advances in generative models have enabled them to generate various high-quality content, such as text [2, 33] , image [21, 25] , music [16] , video [32] , and even 3D scenes [19, 22, 34] , which has fundamentally revolutionized generative artificial intelligence (AI)","reference_id":"MTgyMzk1NzYxNjAxMTk2NTQ2MTk="} +{"index":519,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2302.13971","reference_contexts":"Recent advances in generative models have enabled them to generate various high-quality content, such as text [2, 33] , image [21, 25] , music [16] , video [32] , and even 3D scenes [19, 22, 34] , which has fundamentally revolutionized generative artificial intelligence (AI)","reference_id":"MTgyMzk1NzYxNjAxMTk2NTQ2MTk="} +{"index":520,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2302.13971","reference_contexts":"Fine-tuning has become a critical step in building high-quality LLMs [1, 20, 33] .","reference_id":"MTQyNzA0NTkyNjY1NTE4Njk0ODE="} +{"index":521,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2302.13971","reference_contexts":"Effective fine-tuning of LLMs can be achieved with a relatively small but highquality fine-tuning dataset, e.g., using 27K prompts in [33] .","reference_id":"NzIwNzQ0MzQ4NjA0NTg0NjQ0Ng=="} +{"index":522,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2302.13971","reference_contexts":"Llama2 [33] has been fine-tuned on 27K high-quality prompts, which can be considered a very small quantity compared to the billions or trillions of pre-training tokens","reference_id":"Mjg5OTYyMzY1NzQyNTYxMDEzOA=="} +{"index":523,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2305.16213","reference_contexts":"Recent advances in generative models have enabled them to generate various high-quality content, such as text [2, 33] , image [21, 25] , music [16] , video [32] , and even 3D scenes [19, 22, 34] , which has fundamentally revolutionized generative artificial intelligence (AI)","reference_id":"MTgyMzk1NzYxNjAxMTk2NTQ2MTk="} +{"index":524,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2112.04359","reference_contexts":"First, before instruction-tuning, language models are more prone to generating low-quality text, which may be inconsistent in tone, overly verbose or concise, or simply unhelpful or even toxic [6, 7, 13, 20, 35] ; ChatGPT-level [2] performance is achieved with effective instruction-tuning [20]","reference_id":"NzMzNzI4MTkzMTMwNzUyMDA4"} +{"index":525,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"1905.00546","reference_contexts":"Lastly, to balance images from various domains and categories, we leverage visual concept classification [36] to source images from specific domains (e.g., portrait, food, animal, landscape, car, etc)","reference_id":"MTYwMjA2MjgwODcwOTc0ODQ2MDg="} +{"index":526,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2206.10789","reference_contexts":"We evaluate on two large sets of prompts: 1600 PartiPrompts [37] which is commonly used for text-toimage generation benchmarking, and our 2100 Open User Input (OUI) Prompts","reference_id":"MTc1NjA5MjkxMDk4NjI1NTkyOTc="} +{"index":527,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2206.10789","reference_contexts":"Generative transformer methods usually train autoregressive [4, 10, 11, 26, 37, 38] or non-autoregressive (masked) [9] transformers in discrete token space to model the generation process","reference_id":"ODg1NTE4Mjg4MzUyNzQ5OTg3NQ=="} +{"index":528,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2309.02591","reference_contexts":"Generative transformer methods usually train autoregressive [4, 10, 11, 26, 37, 38] or non-autoregressive (masked) [9] transformers in discrete token space to model the generation process","reference_id":"ODg1NTE4Mjg4MzUyNzQ5OTg3NQ=="} +{"index":529,"citer_arxiv_id":"2309.15807","cited_arxiv_id":"2302.05543","reference_contexts":"Control-Net [39] provides additional control to the generation process by additionally conditioning on pose, sketch, edges, depth, etc","reference_id":"MTIzNjYxODI0OTI0NDE2NTI0MzI="} +{"index":530,"citer_arxiv_id":"1306.1091","cited_arxiv_id":"1305.2982","reference_contexts":"An interesting source of inspiration regarding this question is a recent paper on estimating or propagating gradients through stochastic neurons (Bengio, 2013)","reference_id":"MTA1ODk5NDg0MzI5MTM2NzU5NTI="} +{"index":531,"citer_arxiv_id":"1306.1091","cited_arxiv_id":"1207.0580","reference_contexts":"For example, one could use a convolutional architecture with max-pooling for parametric parsimony and computational efficiency, or dropout (Hinton et al., 2012) to prevent co-adaptation of hidden representations.","reference_id":"MjE4NzEzNTg2Mjk3NzA4MDc4Nw=="} +{"index":532,"citer_arxiv_id":"1306.1091","cited_arxiv_id":"1207.0580","reference_contexts":"The latest breakthrough in object recognition (Krizhevsky et al., 2012) was achieved with fairly deep convolutional networks with a form of noise injection in the input and hidden layers during training, called dropout (Hinton et al., 2012)","reference_id":"MTg0MjI0NjQ2ODg4ODg2NDcwNTE="} +{"index":533,"citer_arxiv_id":"1306.1091","cited_arxiv_id":"1306.0733","reference_contexts":"The advantage of that formulation is that one can directly back-propagated the reconstruction loglikelihood log P (X 1 = x 0 |H 1 = f (X 0 , Z 0 , H 0 )) into all the parameters of f and g (a similar idea was independently proposed in (Kingma, 2013) and also exploited in (Rezende et al., 2014) ).","reference_id":"NjQ2ODMxNjc3Mzc4Mzc5NzY="} +{"index":534,"citer_arxiv_id":"1306.1091","cited_arxiv_id":"1401.4082","reference_contexts":"The advantage of that formulation is that one can directly back-propagated the reconstruction loglikelihood log P (X 1 = x 0 |H 1 = f (X 0 , Z 0 , H 0 )) into all the parameters of f and g (a similar idea was independently proposed in (Kingma, 2013) and also exploited in (Rezende et al., 2014) ).","reference_id":"NjQ2ODMxNjc3Mzc4Mzc5NzY="} +{"index":535,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1511.01844","reference_contexts":"Despite a number of subtleties in generative model assessment [1] , in a quest to improve generative models of images, researchers, and particularly those who have focused on Generative Adversarial Networks [2] , have identified desirable properties such as \"sample quality\" and \"diversity\" and proposed automatic metrics to measure these desiderata","reference_id":"MTUxNDc4NTc0MDk3OTEyNDc4NTk="} +{"index":536,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1511.01844","reference_contexts":"The likelihood, however, is a fairly poor measure of sample quality [1] and often scores out-of-domain data more highly than in-domain data [23] .","reference_id":"NzA0OTM3MTEwNTQ0ODM0MzYwMQ=="} +{"index":537,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1809.11096","reference_contexts":"While ultimately the utility of these models is their performance in downstream tasks, the focus on these metrics has led to models whose samplers now generate nearly photorealistic images [3] [4] [5]","reference_id":"Mzk4MTIzNTczNzI0OTI1MjA0Mg=="} +{"index":538,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1809.11096","reference_contexts":"As noted in [3] , lower truncation values seem to improve sample quality at the expense of diversity","reference_id":"MTUwMjMxOTgyOTE3MDA4NjM1Mzc="} +{"index":539,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1710.10196","reference_contexts":"While ultimately the utility of these models is their performance in downstream tasks, the focus on these metrics has led to models whose samplers now generate nearly photorealistic images [3] [4] [5]","reference_id":"Mzk4MTIzNTczNzI0OTI1MjA0Mg=="} +{"index":540,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1812.04948","reference_contexts":"While ultimately the utility of these models is their performance in downstream tasks, the focus on these metrics has led to models whose samplers now generate nearly photorealistic images [3] [4] [5]","reference_id":"Mzk4MTIzNTczNzI0OTI1MjA0Mg=="} +{"index":541,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1812.01608","reference_contexts":"Examples include: Subscale Pixel Networks [8] , unconditional autoregressive models of 128×128 ImageNet that achieve state-of-the-art test set log-likelihoods; Hierarchical Autoregressive Models (HAMs) [9] , class-conditional autoregressive models of 128×128 and 256×256 ImageNet; and the recently introduced Vector-Quantized Variational Autoencoder-2 (VQ-VAE-2) [10] , a variational autoencoder that uses vector quantization and an autoregressive prior to produce high-quality samples","reference_id":"OTkyMzc3OTA2NzAwMzU5Mjc3OQ=="} +{"index":542,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1903.04933","reference_contexts":"Examples include: Subscale Pixel Networks [8] , unconditional autoregressive models of 128×128 ImageNet that achieve state-of-the-art test set log-likelihoods; Hierarchical Autoregressive Models (HAMs) [9] , class-conditional autoregressive models of 128×128 and 256×256 ImageNet; and the recently introduced Vector-Quantized Variational Autoencoder-2 (VQ-VAE-2) [10] , a variational autoencoder that uses vector quantization and an autoregressive prior to produce high-quality samples","reference_id":"OTkyMzc3OTA2NzAwMzU5Mjc3OQ=="} +{"index":543,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1906.00446","reference_contexts":"Examples include: Subscale Pixel Networks [8] , unconditional autoregressive models of 128×128 ImageNet that achieve state-of-the-art test set log-likelihoods; Hierarchical Autoregressive Models (HAMs) [9] , class-conditional autoregressive models of 128×128 and 256×256 ImageNet; and the recently introduced Vector-Quantized Variational Autoencoder-2 (VQ-VAE-2) [10] , a variational autoencoder that uses vector quantization and an autoregressive prior to produce high-quality samples","reference_id":"OTkyMzc3OTA2NzAwMzU5Mjc3OQ=="} +{"index":544,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1703.01560","reference_contexts":"[12] first introduced the metric (denoted \"adversarial accuracy\") to measure their proposed Layer-Recursive GAN and connected image classification to approximate inference","reference_id":"MTU2Mzk5NTI4OTAxNDcxOTE3MTk="} +{"index":545,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1703.01560","reference_contexts":"this approach has been used for evaluating smaller scale GANs [12] [13] [14] [15] [16] )","reference_id":"MTU3MzMzNDUwOTkwNzc0NDkzMzc="} +{"index":546,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1703.01560","reference_contexts":"To that end, we use a general framework (first posed in [12] and further studied in [13] ) in which we use conditional generative models to perform approximate inference and measure the quality of that inference","reference_id":"MTQ3ODYxNDUzMDgzODQ1NDM1ODg="} +{"index":547,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1711.00970","reference_contexts":"[13] more systematically studied this idea of approximate inference to measure the boundary distortion induced by GANs","reference_id":"MzA0NzEzMTkxNTQwMDM4MjM5Mw=="} +{"index":548,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1711.00970","reference_contexts":"this approach has been used for evaluating smaller scale GANs [12] [13] [14] [15] [16] )","reference_id":"MTU3MzMzNDUwOTkwNzc0NDkzMzc="} +{"index":549,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1711.00970","reference_contexts":"To that end, we use a general framework (first posed in [12] and further studied in [13] ) in which we use conditional generative models to perform approximate inference and measure the quality of that inference","reference_id":"MTQ3ODYxNDUzMDgzODQ1NDM1ODg="} +{"index":550,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1706.02633","reference_contexts":"this approach has been used for evaluating smaller scale GANs [12] [13] [14] [15] [16] )","reference_id":"MTU3MzMzNDUwOTkwNzc0NDkzMzc="} +{"index":551,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1706.02633","reference_contexts":"Predating [13] , [14] used \"Train on Synthetic, Test on Real\" to measure a recurrent conditional GAN for medical data","reference_id":"MTM1MDM2NDY3Njk0NTE1MjUzODY="} +{"index":552,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1312.6114","reference_contexts":"• (Inexact) Likelihood models using variational inference (e.g., VAE [17, 18] )","reference_id":"Mjg1MDgxNjU4Mzc1NDY5OTQ2Mg=="} +{"index":553,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1401.4082","reference_contexts":"• (Inexact) Likelihood models using variational inference (e.g., VAE [17, 18] )","reference_id":"Mjg1MDgxNjU4Mzc1NDY5OTQ2Mg=="} +{"index":554,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1605.08803","reference_contexts":"• Likelihood models based on bijections (e.g., GLOW [20] , rNVP [21] )","reference_id":"MTY2NTI4NDk4OTU3Nzg5Mzk3NjQ="} +{"index":555,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1810.09136","reference_contexts":"The likelihood, however, is a fairly poor measure of sample quality [1] and often scores out-of-domain data more highly than in-domain data [23] .","reference_id":"NzA0OTM3MTEwNTQ0ODM0MzYwMQ=="} +{"index":556,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1801.01401","reference_contexts":"[24] proposed KID, an unbiased alternative to FID, but this metric suffers from variance too large to be reliable when using the number of per-class samples in the ImageNet training set (roughly 1,000 per class), much less when using the","reference_id":"OTUzNDE4MDgyNzU3NTYyNTgwMQ=="} +{"index":557,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1801.01401","reference_contexts":"More principled versions of this procedure are Frechet Inception Distance (FID) [7] and Kernel Inception Distance (KID) [24] , which both use variants of two-sample tests in a learned \"perceptual\" feature space, the Inception pool3 space, to assess distribution matching","reference_id":"MTY4NTc5NzQyMDQwNzk1MTIwODU="} +{"index":558,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1801.01401","reference_contexts":"FID, on the other hand, suffers from a high degree of bias [24]","reference_id":"MjA5MjU1MDAzNTg4NjExNzc2Mg=="} +{"index":559,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1801.01401","reference_contexts":"While in theory one could calculate a per-class FID, FID is known to suffer from high bias [24] for a low number of samples, suggesting that per-class estimates would be unreliable","reference_id":"OTk2Nzg0MjA4MDY2MDQwMzcyNA=="} +{"index":560,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1904.06991","reference_contexts":"Even more recent work [26, 27] calculate 1 and 2 independently by calculating precision and recall.","reference_id":"NTcyNTQ1NjY5OTkwODg0NDE1NA=="} +{"index":561,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1801.01973","reference_contexts":"IS does not penalize a lack of intra-class diversity, and certain out-of-distribution samples produce Inception Scores three times higher than that of the data [28]","reference_id":"OTc2MzQ4MjAwNjkyMDgyOTgwOQ=="} +{"index":562,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1801.01973","reference_contexts":"One can easily diagnose the issue with IS: as noted in [28] , IS does not account for intra-class diversity, and a training set with little intra-class diversity may make the classifier fail to generalize to a more diverse test set","reference_id":"MTc5NDc4OTA2MDExMjMxNDQ2Mw=="} +{"index":563,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1904.01121","reference_contexts":"Moreover, the pool3 feature layer may not even correlate well with human judgment of sample quality [29]","reference_id":"Mjg0MjUzMjg2NzEzOTMxNzc4Ng=="} +{"index":564,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1802.05637","reference_contexts":"Unlike previous work, they tested on larger datasets such as 128×128 ImageNet, but with smaller scale models such as SNGAN [30] .","reference_id":"MTcxMDQ4OTc3OTQwNTAwNzY4OTQ="} +{"index":565,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1802.05637","reference_contexts":"We compare four models: BigGAN, cGAN with Projection Discriminator [30] , PixelCNN [19] , and PixelIQN [40]","reference_id":"MzE5NDk5NjU2MzYyODQyODQ3NQ=="} +{"index":566,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1802.02664","reference_contexts":"[31] constructs approximate manifolds from data and samples, and applies the method to GAN samples to determine whether mode collapse occurred","reference_id":"MTI3MDA5Njg3NDc4NTUyNzE2Mzc="} +{"index":567,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1706.08224","reference_contexts":"[32] attempts to determine the support size of GANs by using a Birthday Paradox test, though the procedure requires a human to identify two nearly-identical samples","reference_id":"MTQyNTE4NTY5MDE5ODk2OTQ1NDU="} +{"index":568,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1806.04936","reference_contexts":"Procedurally similar to our method, [34] proposes a \"reverse LM score\", which trains a language model on GAN data and tests on a real held-out set","reference_id":"MTAzNDk1MjY5MjU0MDQ5NTUzNzE="} +{"index":569,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1706.02677","reference_contexts":"It mirrors the 8,192 batch setup of [39] with gradual warmup","reference_id":"NzgxMzkxNDc4MTU3MDI4OTY4Nw=="} +{"index":570,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1706.02677","reference_contexts":"We believe, however, that now is the right time for such a metric due to a few key advances in training classifiers: 1) the training of ImageNet classifiers has been reduced to minutes [39] , 2) with cloud services, the variance due to implementation details of such a metric is largely mitigated, and 3) the price and time cost of training classifiers on cloud services is reasonable and will only improve over time","reference_id":"MzU5Mzg5NDg5NTc0MjAyMjUxMg=="} +{"index":571,"citer_arxiv_id":"1905.10887","cited_arxiv_id":"1806.05575","reference_contexts":"We compare four models: BigGAN, cGAN with Projection Discriminator [30] , PixelCNN [19] , and PixelIQN [40]","reference_id":"MzE5NDk5NjU2MzYyODQyODQ3NQ=="} +{"index":572,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1703.10135","reference_contexts":"Most of deep learning-based speech synthesis systems consist of two parts: (1) a text-to-speech (TTS) model that converts text into an acoustic feature such as mel-spectrogram [1] [2] [3] [4] and (2) a vocoder that generates time-domain speech waveform using this acoustic feature [5] [6] [7] [8]","reference_id":"MTM5NDg0Mzg0NTMxODg1Njc5ODE="} +{"index":573,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1703.10135","reference_contexts":"Diff-TTS was trained with 400 time-step and synthesized speech samples with the decimation factor γ ∈ [1, 7, 21, 57]","reference_id":"MTkzMDU2MDE5OTMyMjczNDE3NA=="} +{"index":574,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1710.07654","reference_contexts":"Most of deep learning-based speech synthesis systems consist of two parts: (1) a text-to-speech (TTS) model that converts text into an acoustic feature such as mel-spectrogram [1] [2] [3] [4] and (2) a vocoder that generates time-domain speech waveform using this acoustic feature [5] [6] [7] [8]","reference_id":"MTM5NDg0Mzg0NTMxODg1Njc5ODE="} +{"index":575,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2010.05646","reference_contexts":"Most of deep learning-based speech synthesis systems consist of two parts: (1) a text-to-speech (TTS) model that converts text into an acoustic feature such as mel-spectrogram [1] [2] [3] [4] and (2) a vocoder that generates time-domain speech waveform using this acoustic feature [5] [6] [7] [8]","reference_id":"MTM5NDg0Mzg0NTMxODg1Njc5ODE="} +{"index":576,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1609.03499","reference_contexts":"Most of deep learning-based speech synthesis systems consist of two parts: (1) a text-to-speech (TTS) model that converts text into an acoustic feature such as mel-spectrogram [1] [2] [3] [4] and (2) a vocoder that generates time-domain speech waveform using this acoustic feature [5] [6] [7] [8]","reference_id":"MTM5NDg0Mzg0NTMxODg1Njc5ODE="} +{"index":577,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1609.03499","reference_contexts":"Diff-TTS was trained with 400 time-step and synthesized speech samples with the decimation factor γ ∈ [1, 7, 21, 57]","reference_id":"MTkzMDU2MDE5OTMyMjczNDE3NA=="} +{"index":578,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1807.07281","reference_contexts":"Most of deep learning-based speech synthesis systems consist of two parts: (1) a text-to-speech (TTS) model that converts text into an acoustic feature such as mel-spectrogram [1] [2] [3] [4] and (2) a vocoder that generates time-domain speech waveform using this acoustic feature [5] [6] [7] [8]","reference_id":"MTM5NDg0Mzg0NTMxODg1Njc5ODE="} +{"index":579,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2005.05957","reference_contexts":"Recently, various non-AR TTS models have been proposed to overcome the shortcomings of the AR models [10] [11] [12] [13] [14]","reference_id":"NzIwNzUzMTY5ODA3MjQzNzMyOQ=="} +{"index":580,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2006.06873","reference_contexts":"Recently, various non-AR TTS models have been proposed to overcome the shortcomings of the AR models [10] [11] [12] [13] [14]","reference_id":"NzIwNzUzMTY5ODA3MjQzNzMyOQ=="} +{"index":581,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2006.03575","reference_contexts":"Recently, various non-AR TTS models have been proposed to overcome the shortcomings of the AR models [10] [11] [12] [13] [14]","reference_id":"NzIwNzUzMTY5ODA3MjQzNzMyOQ=="} +{"index":582,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1905.09263","reference_contexts":"For instance, since feed-forward models such as FastSpeech1,2 [15, 16] and Speedyspeech [17] cannot produce diverse synthetic speech since they are optimized by a simple regression objective function without any probabilistic modeling","reference_id":"MTM3NzA1MDU1MjU5NDg5ODc5MTI="} +{"index":583,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2006.04558","reference_contexts":"For instance, since feed-forward models such as FastSpeech1,2 [15, 16] and Speedyspeech [17] cannot produce diverse synthetic speech since they are optimized by a simple regression objective function without any probabilistic modeling","reference_id":"MTM3NzA1MDU1MjU5NDg5ODc5MTI="} +{"index":584,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2006.04558","reference_contexts":"Duration Predictor and Length Regulator: The Diff-TTS employs the length regulator in [16] to match the length of phoneme and mel-spectrogram sequence","reference_id":"MTQ5MzYxMTk1ODg5NTU1ODEwMTg="} +{"index":585,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2005.11129","reference_contexts":"Additionally, the flowbased generative models such as Flow-TTS [18] and Glow-TTS [19] are parameter-inefficient due to architectural constraints imposed on normalizing flow-based models.","reference_id":"MTAwMDA3NDM2MDQ2OTI1Mzc0Mzk="} +{"index":586,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2006.11239","reference_contexts":"Meanwhile, another generative framework called denoising diffusion has shown state-of-the-art performance on image generation, and raw audio synthesis [20] [21] [22] [23] [24]","reference_id":"NTA3ODU3NzcxNTEyMzgxNzE1Mw=="} +{"index":587,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2006.11239","reference_contexts":"is intractable, we utilize the parameterization trick which is demonstrated in [20] to calculate the variational lower bound of the log-likelihood in a closed form","reference_id":"NDY3NjQ5NzAwMTAyMjU1NjQxMw=="} +{"index":588,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2010.00654","reference_contexts":"Meanwhile, another generative framework called denoising diffusion has shown state-of-the-art performance on image generation, and raw audio synthesis [20] [21] [22] [23] [24]","reference_id":"NTA3ODU3NzcxNTEyMzgxNzE1Mw=="} +{"index":589,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2010.00654","reference_contexts":"Diff-TTS was trained with 400 time-step and synthesized speech samples with the decimation factor γ ∈ [1, 7, 21, 57]","reference_id":"MTkzMDU2MDE5OTMyMjczNDE3NA=="} +{"index":590,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2102.09672","reference_contexts":"Meanwhile, another generative framework called denoising diffusion has shown state-of-the-art performance on image generation, and raw audio synthesis [20] [21] [22] [23] [24]","reference_id":"NTA3ODU3NzcxNTEyMzgxNzE1Mw=="} +{"index":591,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2009.00713","reference_contexts":"Meanwhile, another generative framework called denoising diffusion has shown state-of-the-art performance on image generation, and raw audio synthesis [20] [21] [22] [23] [24]","reference_id":"NTA3ODU3NzcxNTEyMzgxNzE1Mw=="} +{"index":592,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2009.09761","reference_contexts":"Meanwhile, another generative framework called denoising diffusion has shown state-of-the-art performance on image generation, and raw audio synthesis [20] [21] [22] [23] [24]","reference_id":"NTA3ODU3NzcxNTEyMzgxNzE1Mw=="} +{"index":593,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2009.09761","reference_contexts":"Decoder and Step Encoder: In this part, we adapt the similar architecture as in [24]","reference_id":"MTI3NDA0MTkwNTQyOTg4MDY0MDk="} +{"index":594,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2010.02502","reference_contexts":"In addition, since Diff-TTS has Markov-chain constraint, which accompanies slow inference speed, we also introduce the accelerated sampling method demonstrated in [25]","reference_id":"Mjk4MjI2OTk3NzM5MzA2MjQ4Nw=="} +{"index":595,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"2010.02502","reference_contexts":"To mitigate this problem, the denoising diffusion implicit model (DDIM) [25] introduces a new sampling method called accelerated sampling for diffusion models that boosts up the inference speed without model retraining","reference_id":"NTMzNTU0NTIwNzU1ODE3ODYw"} +{"index":596,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1607.06450","reference_contexts":"The dilations of the convolution are [1, 2, 4, 1, 2, 4, 1, 2, 4, 1] from the bottom to the top, and the kernel size is 4 with ReLU activation function followed by layer normalization [26]","reference_id":"MTM2MzM3OTU0Nzg3MTQ1NDMyNTU="} +{"index":597,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1710.05941","reference_contexts":"The step encoder is constructed using two FC layers and Swish activations [28]","reference_id":"MTYyNzY2NTQ3MzA3MTA3NTM2Nw=="} +{"index":598,"citer_arxiv_id":"2104.01409","cited_arxiv_id":"1412.6980","reference_contexts":"The Diff-TTS was trained for 700K iterations using Adam optimizer [31] on a single NVIDIA 2080Ti GPU","reference_id":"MTI1NjM2NjI3ODYzOTMzNTA4ODI="} +{"index":599,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"1810.04805","reference_contexts":"Given an (Image, Text) pair, we extract multi-scale image features with an image backbone like Swin Transformer [32] , and text features with a text backbone like BERT [8]","reference_id":"NjI5NzIzMjU5NTE3NjQ3NzAxNA=="} +{"index":600,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"1810.04805","reference_contexts":"We leveraged BERT-base [8] from Hugging Face [49] as text backbones","reference_id":"MTE5MTc4MDk5MTE1MTc2NDM2NjI="} +{"index":601,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2110.04544","reference_contexts":"Word level representation [12, 18] enables encoding multiple category names with one forward but introduces unnecessary dependencies among categories, especially when the input text is a concatenation of multiple category names in an arbitrary order.","reference_id":"MTE1ODM0MjMwOTQwOTI5MDYxMw=="} +{"index":602,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2110.04544","reference_contexts":"GLIP [12] formulates object detection as a grounding problem and leverages additional grounding data to help learn aligned semantics","reference_id":"MzU1MzUwNjYwMDc3MDI5Njg0"} +{"index":603,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2101.07448","reference_contexts":"DETR was first proposed in [2] and then has been improved from many directions [4, 5, 13, 17, 33, 48, 64] in the past few years","reference_id":"Nzg5NjQyODU3NzIxMjU2NzcyMA=="} +{"index":604,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"We follow GLIP [25] and use contrastive loss between predicted objects and language tokens for classification","reference_id":"MTAzMjQ3OTUwMTg1OTA3OTkxMDE="} +{"index":605,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"We thank the authors of GLIP [25] : Liunian Harold Li , Pengchuan Zhang , and Haotian Zhang for their helpful discussions and instructions","reference_id":"MTE1NjUwNzQwNDkyODE3NTc4OTg="} +{"index":606,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"Following previous works [25, 52] , we pre-train the Grounding DINO T on O365v1 for a fair comparison","reference_id":"ODkxNjkxNzEwNjYxMTQzMzg3Mg=="} +{"index":607,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"The key to open-set detection is introducing language for unseen object generalization [1, 7, 25]","reference_id":"NTczMTA2NDU3NTcyMjkxODg3"} +{"index":608,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"We leverage GLIP [25] as our baseline","reference_id":"ODgyMDE5NTk0NDgyMjg1ODczMg=="} +{"index":609,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"Swin-T --49.7 / -DyHead-L [5] Swin-L --58.4 / 58.7 DyHead-L [5] Swin-L O365,ImageNet21K -60.3 / 60.6 SoftTeacher [50] Swin-L O365,SS-COCO -60.7 / 61.3 DINO(Swin-L) [57] Swin-L O365 -62.5 / -DyHead-T † [5] Swin-T O365 43.6 53.3 / -GLIP-T (B) [25] Swin-T O365 44.9 53.8 / -GLIP-T (C) [25] Swin-T O365,GoldG 46.7 55.1 / -GLIP-L [25] Swin-L FourODs,GoldG,Cap24M","reference_id":"MTY1NjM4NzcwMTkyNDY0MDE2MjY="} +{"index":610,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"Following GLIP [25] , we reformulate the object detection task to a phrase grounding task by concatenating the category names into text prompts","reference_id":"OTgwNTk5NTY4NTc0ODkzNjc0MQ=="} +{"index":611,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"Inspired by GLIP [25] , we add an image-to-text and a textto-image cross-attention modules for feature fusion","reference_id":"MTQ1MTQyNTQzNDU4NjQ3MTUzMzc="} +{"index":612,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"For example, GLIP [25] performs early fusion in the neck module (phase A), and OV-DETR [55] uses language-aware queries as head inputs (phase B)","reference_id":"NzQ0OTY5NjU4NTQ3MzgxNzE4Nw=="} +{"index":613,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"Following GLIP [25] , we concatenate all category names as input texts for object detection tasks","reference_id":"MTUxNTAzOTQyMDE5MTg5MzU2NTg="} +{"index":614,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2112.03857","reference_contexts":"In contrast, GLIP [25] presents a different way by reformulating object detection as a phrase grounding task and introducing contrastive training between object regions and language phrases on large-scale data","reference_id":"NTE5MDQ5NDEyNzA3NTA2NzU5NQ=="} +{"index":615,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2103.14030","reference_contexts":"Given an (Image, Text) pair, we extract multi-scale image features with an image backbone like Swin Transformer [32] , and text features with a text backbone like BERT [8]","reference_id":"NjI5NzIzMjU5NTE3NjQ3NzAxNA=="} +{"index":616,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2103.14030","reference_contexts":"We trained two model variants, Grounding DINO T with Swin-T [32] , and Grounding DINO L with Swin-L [32] as an image backbone, respectively","reference_id":"MzM1NTk3Nzk0MTk4Njg5ODE2Mg=="} +{"index":617,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2108.06152","reference_contexts":"Following previous DETR-like works [2, 24, 31, 33, 57, 64] , we use the L1 loss and the GIOU [40] loss for bounding box regressions","reference_id":"MTU3OTE4OTY3NTQ5NDY4ODU1ODQ="} +{"index":618,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2108.06152","reference_contexts":"Each decoder query contains two parts: content part and positional part [33] , respectively","reference_id":"MTcxMjI4OTM5NTY5MTgxMTQ1MzI="} +{"index":619,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2108.06152","reference_contexts":"DETR was first proposed in [2] and then has been improved from many directions [4, 5, 13, 17, 33, 48, 64] in the past few years","reference_id":"Nzg5NjQyODU3NzIxMjU2NzcyMA=="} +{"index":620,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"1910.03771","reference_contexts":"We leveraged BERT-base [8] from Hugging Face [49] as text backbones","reference_id":"MTE5MTc4MDk5MTE1MTc2NDM2NjI="} +{"index":621,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"2106.09018","reference_contexts":"Swin-T --49.7 / -DyHead-L [5] Swin-L --58.4 / 58.7 DyHead-L [5] Swin-L O365,ImageNet21K -60.3 / 60.6 SoftTeacher [50] Swin-L O365,SS-COCO -60.7 / 61.3 DINO(Swin-L) [57] Swin-L O365 -62.5 / -DyHead-T † [5] Swin-T O365 43.6 53.3 / -GLIP-T (B) [25] Swin-T O365 44.9 53.8 / -GLIP-T (C) [25] Swin-T O365,GoldG 46.7 55.1 / -GLIP-L [25] Swin-L FourODs,GoldG,Cap24M","reference_id":"MTY1NjM4NzcwMTkyNDY0MDE2MjY="} +{"index":622,"citer_arxiv_id":"2303.05499","cited_arxiv_id":"1904.07850","reference_contexts":"We use the notations \"CC\", \"SBU\", \"VG\", \"OI\", \"O365\", and \"YFCC\" for Conceptual Captions [44] , SBU Captions [36] , Visual Genome [20] , OpenImage [22] , Objects365 [63] , YFCC100M [47] respectively","reference_id":"NTg2MTg5Mjc3MjA0ODYyNjM3MA=="} +{"index":623,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2103.10951","reference_contexts":"Instead of directly training a text-to-image generator, a recent surge of methods leverage a pre-trained generator, and use a pre-trained CLIP [35] to guide the generation process by text [3, 12, 25, 33]","reference_id":"MTAyNDkyMTg1NzU5NTIxNzg3NDg="} +{"index":624,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2103.10951","reference_contexts":"The richness of the enormous visual and textual space spanned by CLIP has been demonstrated by various recent image editing methods (e.g., [2, 3, 11, 12, 33] )","reference_id":"Mzc1OTUwNTg1NDQyMDM3MTE4Nw=="} +{"index":625,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2106.14843","reference_contexts":"Our task is ill-posed -numerous possible edits can satisfy the target text according to CLIP, some of which include noisy or undesired solutions [11, 25]","reference_id":"MTYxODc4Mzg4NzE2NTM2NzQ2MTg="} +{"index":626,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2106.14843","reference_contexts":"Several works [11, 14, 21, 28] take a test-time optimization approach and leverage CLIP without using a pre-trained generator","reference_id":"MTI1MjEwODE4MTEwNjgzMDQ5MjE="} +{"index":627,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2106.14843","reference_contexts":"For example, CLIPDraw [11] renders a drawing that matches a target text by directly optimizing a set of vector strokes","reference_id":"OTA4MjA1ODI4NDIxODE3OTM4MQ=="} +{"index":628,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2106.14843","reference_contexts":"The richness of the enormous visual and textual space spanned by CLIP has been demonstrated by various recent image editing methods (e.g., [2, 3, 11, 12, 33] )","reference_id":"Mzc1OTUwNTg1NDQyMDM3MTE4Nw=="} +{"index":629,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2108.00946","reference_contexts":"The richness of the enormous visual and textual space spanned by CLIP has been demonstrated by various recent image editing methods (e.g., [2, 3, 11, 12, 33] )","reference_id":"Mzc1OTUwNTg1NDQyMDM3MTE4Nw=="} +{"index":630,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2108.00946","reference_contexts":"The second term controls the direction of edit in CLIP space [12, 33] and is given by: L dir = Dcos(Eim(Io)-Eim(Is), Etxt(T ) -Etxt(T ROI ))","reference_id":"NTQ4NzQwOTE2NTAyMzEwODcwMw=="} +{"index":631,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2108.00946","reference_contexts":"StyleCLIP [33] and StyleGAN-NADA [12] use a pre-trained StyleGAN2 [17] for image manipulation, by either controlling the GAN's latent code [33] , or by fine-tuning the StyleGAN's output domain [12]","reference_id":"MTM1MTAzNzA3MzMwMTQ5MjU5OTc="} +{"index":632,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2108.00946","reference_contexts":"Instead of directly training a text-to-image generator, a recent surge of methods leverage a pre-trained generator, and use a pre-trained CLIP [35] to guide the generation process by text [3, 12, 25, 33]","reference_id":"MTAyNDkyMTg1NzU5NTIxNzg3NDg="} +{"index":633,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2108.00946","reference_contexts":"In the SM, we also include additional qualitative comparison to the StyleGAN text-guided editing methods [33, 12]","reference_id":"MjQ1MTc2NzYwMzgwMDM1Mjk4Nw=="} +{"index":634,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2110.02711","reference_contexts":"[19] takes this approach a step forward by optimizing the diffusion process itself","reference_id":"NTIxODY4NTM3MjkwMDAzNDM3Nw=="} +{"index":635,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.00374","reference_contexts":"CLIPStyler [21] takes a similar approach for global stylization","reference_id":"Njk0ODc0MjUxNTg2MjY2NTc5Mg=="} +{"index":636,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.00374","reference_contexts":"Several works [11, 14, 21, 28] take a test-time optimization approach and leverage CLIP without using a pre-trained generator","reference_id":"MTI1MjEwODE4MTEwNjgzMDQ5MjE="} +{"index":637,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"1701.03754","reference_contexts":"Existing approaches for consistent video editing can be roughly divided into: (i) propagation-based methods, which use keyframes [15, 46] or optical flow [41] to propagate edits through the video, and (ii) video layering-based methods, in which a layered representation of the video is estimated and then edited [18, 23, 26, 27, 37]","reference_id":"MTE4NzI3NDM0MDYxMzU0MTg2MTA="} +{"index":638,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.01573","reference_contexts":"Our task is ill-posed -numerous possible edits can satisfy the target text according to CLIP, some of which include noisy or undesired solutions [11, 25]","reference_id":"MTYxODc4Mzg4NzE2NTM2NzQ2MTg="} +{"index":639,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.01573","reference_contexts":"Instead of directly training a text-to-image generator, a recent surge of methods leverage a pre-trained generator, and use a pre-trained CLIP [35] to guide the generation process by text [3, 12, 25, 33]","reference_id":"MTAyNDkyMTg1NzU5NTIxNzg3NDg="} +{"index":640,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.03221","reference_contexts":"Several works [11, 14, 21, 28] take a test-time optimization approach and leverage CLIP without using a pre-trained generator","reference_id":"MTI1MjEwODE4MTEwNjgzMDQ5MjE="} +{"index":641,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.03221","reference_contexts":"Finally, as acknowledged by prior works (e.g., [28] ), we also noticed that slightly different text prompts describing similar concepts may lead to slightly different flavors of edits.","reference_id":"NDIyODM0NDIxMTE2MzgwMDMzMA=="} +{"index":642,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.03221","reference_contexts":"Other works use CLIP to synthesize [14] or edit [28] a single 3D representation (NeRF or mesh)","reference_id":"MTgwNjExMDM2MjUzMjUwMzAyMDU="} +{"index":643,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.10741","reference_contexts":"GLIDE [30] takes this approach further, supporting both text-to-image generation and inpainting","reference_id":"MTU0OTI2MzkwMTEwOTA2NjcxNTc="} +{"index":644,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2112.10741","reference_contexts":"Inpainting methods: Blended-Diffusion [2] and GLIDE [30] , both require user-provided editing mask","reference_id":"MzI3MTY3MzE5NDA3NjkwMjU5Ng=="} +{"index":645,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"1704.00675","reference_contexts":"The video set contains seven videos from DAVIS dataset [34]","reference_id":"NDY3NjI1MjQ4OTk0NjM5NzA2OA=="} +{"index":646,"citer_arxiv_id":"2204.02491","cited_arxiv_id":"2101.05278","reference_contexts":"With GANs, the domain of images is restricted and requires to invert the input image to the GAN's latent space -a challenging task by itself [49]","reference_id":"NDIxMzExNjUyODk5NjkwODkwNg=="} +{"index":647,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2204.02849","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":648,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2211.14305","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":649,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2204.02491","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":650,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2103.10951","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":651,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2005.07727","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":652,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"1809.11096","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":653,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2204.08583","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":654,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2111.14822","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":655,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2208.01626","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":656,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":657,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"1312.6114","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":658,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2112.05744","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":659,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"1511.02793","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":660,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2112.10741","reference_contexts":"2020; Nichol and Dhariwal 2021; Sohl-Dickstein et al","reference_id":"MTU4OTU3NTE3MDMyMDUxODk2MzU="} +{"index":661,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2112.10741","reference_contexts":"2020; Nichol and Dhariwal 2021] .","reference_id":"NDQzMzYwNTM4MzE2NTY0MjI5Nw=="} +{"index":662,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2204.04908","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":663,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2103.00020","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":664,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2204.06125","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":665,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2102.12092","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":666,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2106.05744","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":667,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2205.11487","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":668,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2111.02114","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":669,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2203.00386","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":670,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2101.05278","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":671,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2206.10789","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":672,"citer_arxiv_id":"2206.02779","cited_arxiv_id":"2111.13792","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":673,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2108.02818","reference_contexts":"Our framework uses a pre-trained CLIP [44] model, which has been shown to contain bias [1]","reference_id":"NjI0MjkzMDI5NzI1NzkxNTAwMg=="} +{"index":674,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2108.02818","reference_contexts":"Our framework utilizes a pre-trained CLIP embedding space, which has been shown to contain bias [1]","reference_id":"MTE0NDc1NTQzOTE4NDkzMjY0MDQ="} +{"index":675,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"1512.03012","reference_contexts":"We consider a variety of sources including: COSEG [54] , Thingi10K [71] , Shapenet [5] , Turbo Squid [57] , and ModelNet [59]","reference_id":"MTM0MjkwMDUyMDk2ODA2Mzk1NzE="} +{"index":676,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"1512.03012","reference_contexts":"Another concurrent work uses the ShapeNet dataset [5] and CLIP to perform unconditional 3D voxel generation [48]","reference_id":"MTUwMDQzNzMxODU2Nzk0NzIxNA=="} +{"index":677,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2110.12427","reference_contexts":"Concurrent work uses CLIP to fine-tune a pre-trained StyleGAN [11] , and for image stylization [6]","reference_id":"NzE4NjkwMDc4MjgwMTU2MDkwNA=="} +{"index":678,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2106.14843","reference_contexts":"In [10] random 2D augmentations are necessary to generate meaningful CLIP-guided drawings","reference_id":"MTI0MjUwNzIzMzEwMDU2MzAyODM="} +{"index":679,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2106.14843","reference_contexts":"The first to leverage CLIP for synthesis without the need for a pre-trained network or dataset is CLIPDraw [10]","reference_id":"MTA1NjI5MDQ5MzgxMzcyNzMyMTI="} +{"index":680,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2108.00946","reference_contexts":"Concurrent work uses CLIP to fine-tune a pre-trained StyleGAN [11] , and for image stylization [6]","reference_id":"NzE4NjkwMDc4MjgwMTU2MDkwNA=="} +{"index":681,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2109.12922","reference_contexts":"Concurrent work [25] uses CLIP to optimize over parameters of the SMPL human body model to create digital creatures","reference_id":"MTYxNzY0OTQ5MzM0NjY5MTM5NTk="} +{"index":682,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2011.12948","reference_contexts":"NeRF [40] and follow ups [42, 65, 69] have demonstrated success on 3D scene modeling","reference_id":"NDU2NzEzNzM2OTMwNTYyMzQ3MQ=="} +{"index":683,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2103.00020","reference_contexts":"Our framework uses a pre-trained CLIP [44] model, which has been shown to contain bias [1]","reference_id":"NjI0MjkzMDI5NzI1NzkxNTAwMg=="} +{"index":684,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2103.00020","reference_contexts":"Our neural optimization is guided by the multi-modal embedding space provided by a pre-trained CLIP [44] model","reference_id":"MTQ3NzY2NDU2NTA5MDk4MDA1NTA="} +{"index":685,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2103.00020","reference_contexts":"This is facilitated by recent developments in joint embeddings of text and images, such as CLIP [44]","reference_id":"NzU2MjQ4MDQwOTc2NDg5NzUxNA=="} +{"index":686,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2103.00020","reference_contexts":"Our work is similar in spirit to image manipulation techniques controlled through textual descriptions embedded by CLIP [44]","reference_id":"MTM0MDg4ODIwMjk1MTE4NjEzMQ=="} +{"index":687,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2110.02624","reference_contexts":"Another concurrent work uses the ShapeNet dataset [5] and CLIP to perform unconditional 3D voxel generation [48]","reference_id":"MTUwMDQzNzMxODU2Nzk0NzIxNA=="} +{"index":688,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2006.10739","reference_contexts":"Specifically, our stylization network acts as a neural prior, which leverages positional encoding [56] to synthesize fine-grained stylization details.","reference_id":"MjA0MjQ1MjcyMzM1NjM2ODc0MA=="} +{"index":689,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2006.10739","reference_contexts":"In order to produce accurate styles which Input Steve Jobs Yeti Astronaut Buzz Lightyear Ninja Lawyer Messi Batman Hulk contain high-frequency content with high fidelity, we use a frequency-based positional encoding [56]","reference_id":"ODYwMjc0ODYzNDI2NDI3NTE0Mw=="} +{"index":690,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2006.10739","reference_contexts":"To synthesize high-frequency details, we apply a positional encoding using Fourier feature mappings, which enables MLPs to overcome the spectral bias and learn to interpolate highfrequency functions [56]","reference_id":"MTc4MzQ0MDg5Nzc0NTQ4MDIzMDY="} +{"index":691,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"2010.07492","reference_contexts":"NeRF [40] and follow ups [42, 65, 69] have demonstrated success on 3D scene modeling","reference_id":"NDU2NzEzNzM2OTMwNTYyMzQ3MQ=="} +{"index":692,"citer_arxiv_id":"2112.03221","cited_arxiv_id":"1605.04797","reference_contexts":"We consider a variety of sources including: COSEG [54] , Thingi10K [71] , Shapenet [5] , Turbo Squid [57] , and ModelNet [59]","reference_id":"MTM0MjkwMDUyMDk2ODA2Mzk1NzE="} +{"index":693,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2204.14198","reference_contexts":"Some methods freeze the language model to use the knowledge from LLMs for vision-to-language generation tasks (Tsimpoukelli et al., 2021; Alayrac et al., 2022; Chen et al., 2022a; Mañas et al., 2023; Tiong et al., 2022; Guo et al., 2022)","reference_id":"NjE4MjUxNjIyMTUyMjg2MTIyMg=="} +{"index":694,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2204.14198","reference_contexts":"Frozen (Tsimpoukelli et al., 2021) , Flamingo (Alayrac et al., 2022) ) resort to an image-to-text generation loss, which we show is insufficient to bridge the modality gap.","reference_id":"NDAyMTg0MjgxNDY5NjYyOTUzNw=="} +{"index":695,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2204.14198","reference_contexts":"Flamingo (Alayrac et al., 2022) inserts new cross-attention layers into the LLM to inject visual features, and pre-trains the new layers on billions of image-text pairs","reference_id":"MTU1NjI0NDY5NzY5MzAxNDc4Mzc="} +{"index":696,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2204.14198","reference_contexts":"CIDEr SPICE TR@1 IR@1 BLIP (Li et al., 2022) 583M ✓ -113.2 14.8 96.7 86.7 SimVLM (Wang et al., 2021b) 1.4B ✗ -112.2 ---BEIT-3 (Wang et al., 2022b) 1.9B ✗ ---94.9 81.5 Flamingo (Alayrac et al., 2022) Table 2","reference_id":"Njc2NzkyODM2NDQ0NjUzMzk0NQ=="} +{"index":697,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2204.14198","reference_contexts":"For example, BLIP-2 outperforms Flamingo (Alayrac et al., 2022 ) by 8.7% on zero-shot VQAv2, while using 54× fewer trainable parameters","reference_id":"MzUwMjYyMzcxNjEyODM0NDQ1MA=="} +{"index":698,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2209.06794","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":699,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2102.02779","reference_contexts":"Depending on the downstream task, different model architectures have been proposed, including the dual-encoder architecture (Radford et al., 2021; Jia et al., 2021) , the fusion-encoder architecture (Tan & Bansal, 2019; Li et al., 2021) , the encoder-decoder architecture (Cho et al., 2021; Wang et al., 2021b; Chen et al., 2022b) , and more recently, the unified transformer architecture (Li et al., 2022; Wang et al., 2022b)","reference_id":"MzgyMzc1NTc3MDAzMTExMzU1Mw=="} +{"index":700,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2210.11416","reference_contexts":"Moreover, it is inflexible for end-to-end pre-trained models to leverage readily-available unimodal pre-trained models, such as LLMs (Brown et al., 2020; Zhang et al., 2022; Chung et al., 2022) .","reference_id":"OTM5NTk4NjQ4NTc1NDk1MzQ2Ng=="} +{"index":701,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2210.11416","reference_contexts":"OPT (Zhang et al., 2022) , FlanT5 (Chung et al., 2022) ), BLIP-2 can be prompted to perform zero-shot image-to-text generation that follows natural language instructions, which enables emerging capabilities such as visual knowledge reasoning, visual conversation, etc","reference_id":"MTc0NzA3NDA2MDE0NzI1MDA3MTk="} +{"index":702,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2210.11416","reference_contexts":"For the frozen language model, we explore the unsupervised-trained OPT model family (Zhang et al., 2022) for decoder-based LLMs, and the instruction-trained FlanT5 model family (Chung et al., 2022) for encoder-decoder-based LLMs.","reference_id":"MTAxMDI4NjM2ODI5MDkyMDE2MDk="} +{"index":703,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2211.07636","reference_contexts":"For the frozen image encoder, we explore two state-of-the-art pre-trained vision transformer models: (1) ViT-L/14 from CLIP (Radford et al., 2021) and (2) ViT-g/14 from EVA-CLIP (Fang et al., 2022)","reference_id":"NTQ3Mjc1NTk3NDExNjU5NjIzNg=="} +{"index":704,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2203.15556","reference_contexts":"We hypothesis that this is because OK-VQA focuses more on open-world knowledge than visual understanding, and the 70B Chinchilla (Hoffmann et al., 2022) language model from Flamingo80B possesses more knowledge than the 11B FlanT5 XXL .","reference_id":"NDMxNjI3MzY1Mjg3NDkxMjczOQ=="} +{"index":705,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2102.05918","reference_contexts":"Depending on the downstream task, different model architectures have been proposed, including the dual-encoder architecture (Radford et al., 2021; Jia et al., 2021) , the fusion-encoder architecture (Tan & Bansal, 2019; Li et al., 2021) , the encoder-decoder architecture (Cho et al., 2021; Wang et al., 2021b; Chen et al., 2022b) , and more recently, the unified transformer architecture (Li et al., 2022; Wang et al., 2022b)","reference_id":"MzgyMzc1NTc3MDAzMTExMzU1Mw=="} +{"index":706,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"1711.05101","reference_contexts":"We use the AdamW (Loshchilov & Hutter, 2017) optimizer with β 1 = 0.9, β 1 = 0.98, and a weight decay of 0.05","reference_id":"NDMwNDAyMDU0Njk5OTUzNjA2NA=="} +{"index":707,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2103.00020","reference_contexts":"Depending on the downstream task, different model architectures have been proposed, including the dual-encoder architecture (Radford et al., 2021; Jia et al., 2021) , the fusion-encoder architecture (Tan & Bansal, 2019; Li et al., 2021) , the encoder-decoder architecture (Cho et al., 2021; Wang et al., 2021b; Chen et al., 2022b) , and more recently, the unified transformer architecture (Li et al., 2022; Wang et al., 2022b)","reference_id":"MzgyMzc1NTc3MDAzMTExMzU1Mw=="} +{"index":708,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2103.00020","reference_contexts":"Text → Image R@1 R@5 R@10 R@1 R@5 R@10 R@1 R@5 R@10 R@1 R@5 R@10 Dual-encoder models CLIP (Radford et al., 2021)","reference_id":"ODA0NDg2NzU2OTg0NTQ5MTk3Mw=="} +{"index":709,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2103.00020","reference_contexts":"Vision-language pre-training (VLP) research has witnessed a rapid advancement in the past few years, where pre-trained models with increasingly larger scale have been developed to continuously push the state-of-the-art on various downstream tasks (Radford et al., 2021; Li et al., 2021; 2022; Figure 1","reference_id":"NTc2NDE1ODExMTE3NDY0MTcxNg=="} +{"index":710,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2103.00020","reference_contexts":"Some methods freeze the image encoder, including the early work which adopts a frozen object detector to extract visual features (Chen et al., 2020; Li et al., 2020; Zhang et al., 2021) , and the recent LiT (Zhai et al., 2022) which uses a frozen pre-trained image encoder for CLIP (Radford et al., 2021) pre-training","reference_id":"NjI2MjkyMzAyNTkwMjI1MTAzNQ=="} +{"index":711,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2111.02114","reference_contexts":"We use the same pre-training dataset as BLIP with 129M images in total, including COCO (Lin et al., 2014) , Visual Genome (Krishna et al., 2017) , CC3M (Sharma et al., 2018) , CC12M (Changpinyo et al., 2021) , SBU (Ordonez et al., 2011) , and 115M images from the LAION400M dataset (Schuhmann et al., 2021)","reference_id":"NzgzNzk4Nzc0MDkyNjg4NTQ1OA=="} +{"index":712,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2111.02358","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":713,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2208.10442","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":714,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2108.10904","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":715,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2205.01917","reference_contexts":"Various pre-training objectives have also been proposed over the years, and have progressively converged to a few time-tested ones: image-text contrastive learning (Radford et al., 2021; Yao et al., 2022; Li et al., 2021; 2022) , image-text matching (Li et al., 2021; 2022; Wang et al., 2021a) , and (masked) language modeling (Li et al., 2021; 2022; Yu et al., 2022; Wang et al., 2022b) .","reference_id":"MjM4ODE0OTk2NTM5OTAwNDI3"} +{"index":716,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2111.11432","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":717,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2101.00529","reference_contexts":"Some methods freeze the image encoder, including the early work which adopts a frozen object detector to extract visual features (Chen et al., 2020; Li et al., 2020; Zhang et al., 2021) , and the recent LiT (Zhai et al., 2022) which uses a frozen pre-trained image encoder for CLIP (Radford et al., 2021) pre-training","reference_id":"NjI2MjkyMzAyNTkwMjI1MTAzNQ=="} +{"index":718,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2205.01068","reference_contexts":"Moreover, it is inflexible for end-to-end pre-trained models to leverage readily-available unimodal pre-trained models, such as LLMs (Brown et al., 2020; Zhang et al., 2022; Chung et al., 2022) .","reference_id":"OTM5NTk4NjQ4NTc1NDk1MzQ2Ng=="} +{"index":719,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2205.01068","reference_contexts":"OPT (Zhang et al., 2022) , FlanT5 (Chung et al., 2022) ), BLIP-2 can be prompted to perform zero-shot image-to-text generation that follows natural language instructions, which enables emerging capabilities such as visual knowledge reasoning, visual conversation, etc","reference_id":"MTc0NzA3NDA2MDE0NzI1MDA3MTk="} +{"index":720,"citer_arxiv_id":"2301.12597","cited_arxiv_id":"2205.01068","reference_contexts":"For the frozen language model, we explore the unsupervised-trained OPT model family (Zhang et al., 2022) for decoder-based LLMs, and the instruction-trained FlanT5 model family (Chung et al., 2022) for encoder-decoder-based LLMs.","reference_id":"MTAxMDI4NjM2ODI5MDkyMDE2MDk="} +{"index":721,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2112.10684","reference_contexts":"For example, through the use of conditional computation large MoE models like the 1.7 trillion parameter Switch transformer (Fedus et al., 2021) , the 1.2 Trillion parameter GLaM model (Du et al., 2021) , and others (Artetxe et al., 2021; Zoph et al., 2022) are able to provide a large effective model size despite using relatively fewer training and inference FLOPs","reference_id":"NDI3OTI5ODUxMDg5MTExODc2OQ=="} +{"index":722,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2202.01169","reference_contexts":"However, for very large models the computational benefits of routed models seems to diminish (Clark et al., 2022)","reference_id":"MTEzNzc1ODc5NDE2NDY3NDk0NDY="} +{"index":723,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2112.06905","reference_contexts":"For example, through the use of conditional computation large MoE models like the 1.7 trillion parameter Switch transformer (Fedus et al., 2021) , the 1.2 Trillion parameter GLaM model (Du et al., 2021) , and others (Artetxe et al., 2021; Zoph et al., 2022) are able to provide a large effective model size despite using relatively fewer training and inference FLOPs","reference_id":"NDI3OTI5ODUxMDg5MTExODc2OQ=="} +{"index":724,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2101.03961","reference_contexts":"For example, through the use of conditional computation large MoE models like the 1.7 trillion parameter Switch transformer (Fedus et al., 2021) , the 1.2 Trillion parameter GLaM model (Du et al., 2021) , and others (Artetxe et al., 2021; Zoph et al., 2022) are able to provide a large effective model size despite using relatively fewer training and inference FLOPs","reference_id":"NDI3OTI5ODUxMDg5MTExODc2OQ=="} +{"index":725,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2101.00027","reference_contexts":"modelling pubmed_abstracts nih_exporter uspto_backgrounds pubmed_central pile_cc bookcorpus2 stackexchange opensubtitles openwebtext2 hackernews dm_mathematics arxiv freelaw books3 philpapers github ubuntu_irc europarl gutenberg_pg_19 0.00 0.02 0.04 0.06 0.08 0.10 Decrease in bpb compared to Gopher Chinchilla significantly outperforms Gopher on all evaluation subsets of The Pile (Gao et al., 2020) , as shown in Figure 5","reference_id":"NDEzNjYzNjc5OTAzMTg1ODU5NA=="} +{"index":726,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2101.00027","reference_contexts":"In Table A5 we show the bits-per-byte (bpb) on The Pile (Gao et al., 2020) of Chinchilla, Gopher, and Jurassic-1","reference_id":"MTY2NjAxNzM2ODQ4NDgwNzkwNTU="} +{"index":727,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2101.00027","reference_contexts":"• Language modelling on LAMBADA, Wikitext103 (Merity et al., 2017) , C4 (Raffel et al., 2020a) , PG-19 (Rae et al., 2020) and the Pile (Gao et al., 2020)","reference_id":"NjUyOTc1MzU1MTc2MDAyMDI3Mg=="} +{"index":728,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2009.03300","reference_contexts":"We thus place more emphasis on other tasks for which leakage is less of a concern, such as MMLU (Hendrycks et al., 2020) and BIG-bench (BIG-bench collaboration, 2021) along with various closed-book question answering and common sense analyses.","reference_id":"MTc1NzAyMjYwMTMxOTMyNzMwMjM="} +{"index":729,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2009.03300","reference_contexts":"The Massive Multitask Language Understanding (MMLU) benchmark (Hendrycks et al., 2020) consists of a range of exam-like questions on academic subjects","reference_id":"MTQ5ODQyNzM3NzI4NDQ3NTE0NDM="} +{"index":730,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2009.03300","reference_contexts":"• Language understanding, real world knowledge, mathematical and logical reasoning on the Massive Multitask Language Understanding (MMLU) benchmark (Hendrycks et al., 2020) and on the \"Beyond the Imitation Game Benchmark\" (BIG-bench) (BIG-bench collaboration, 2021)","reference_id":"MTk2NTY5NDI1ODY2OTQyNDAzMQ=="} +{"index":731,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"1705.03551","reference_contexts":"On TriviaQA (Joshi et al., 2017) we show results for both the filtered (previously used in retrieval and open-book work) and unfiltered set (previously used in large language model evaluations)","reference_id":"MjI3NDM2MjIwODQ0MzI4NDc5NA=="} +{"index":732,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"1705.03551","reference_contexts":"• Question answering (closed book) on Natural Questions (Kwiatkowski et al., 2019) and TriviaQA (Joshi et al., 2017 )","reference_id":"MTM1MjY0MzAwMjQ1NDI0OTQ4MTE="} +{"index":733,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2001.08361","reference_contexts":"We show a comparison between our calculation and that using the common approximation 𝐶 = 6𝐷𝑁 (Kaplan et al., 2020) where 𝐶 is FLOPs, 𝐷 is the number of training tokens, and 𝑁 is the number of parameters in Table A4","reference_id":"NDM2NjkyMjY4NTczNjc3MTAyMA=="} +{"index":734,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2001.08361","reference_contexts":"(2020) and the training setup of GPT-3 (Brown et al., 2020) , many of the recently trained large models have been trained for approximately 300 billion tokens (Table 1 ), in line with the approach of predominantly increasing model size when increasing compute.","reference_id":"NDgxNjg0NDM1NjE2NTE1NzczNw=="} +{"index":735,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2001.08361","reference_contexts":"We can approximate the functions 𝑁 𝑜𝑝𝑡 and 𝐷 𝑜𝑝𝑡 by minimizing the parametric loss L under the constraint FLOPs(𝑁, 𝐷) ≈ 6𝑁 𝐷 (Kaplan et al., 2020)","reference_id":"MTgzNjU0NDU3NTc4NTE1MzgxNzE="} +{"index":736,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2001.08361","reference_contexts":"Recently a series of Large Language Models (LLMs) have been introduced (Brown et al., 2020; Lieber et al., 2021; Rae et al., 2021; Smith et al., 2022; Thoppilan et al., 2022) , with the largest dense language models now having over 500 billion parameters","reference_id":"MzgyNTUyOTMzNTA5MjIzOTgw"} +{"index":737,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"1412.6980","reference_contexts":"Specifically, Gopher was trained with Adam (Kingma and Ba, 2014) whereas Chinchilla was trained with AdamW (Loshchilov and Hutter, 2019)","reference_id":"MTM4NDk5NDAwNzE1OTYxMzk3NTI="} +{"index":738,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"1412.6980","reference_contexts":"• We use AdamW (Loshchilov and Hutter, 2019) for Chinchilla rather than Adam (Kingma and Ba, 2014) as this improves the language modelling loss and the downstream task performance after finetuning","reference_id":"MzQxODMzNjkyNzMxMDY5NjEzMQ=="} +{"index":739,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"1808.06226","reference_contexts":"foot_7 • We train Chinchilla with a slightly modified SentencePiece (Kudo and Richardson, 2018) tokenizer that does not apply NFKC normalisation","reference_id":"MzY0NjY4ODQ0NzMwNDM3MTg0NQ=="} +{"index":740,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2006.12467","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":741,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2109.07958","reference_contexts":"On TruthfulQA (Lin et al., 2021) , Chinchilla reaches 43.6%, 58.5%, and 66.7% accuracy with 0-shot, 5-shot, and 10-shot respectively","reference_id":"NjI5ODM1OTA4MjIxOTc4NjQ1NA=="} +{"index":742,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"1811.03600","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":743,"citer_arxiv_id":"2203.15556","cited_arxiv_id":"2201.11990","reference_contexts":"Recently a series of Large Language Models (LLMs) have been introduced (Brown et al., 2020; Lieber et al., 2021; Rae et al., 2021; Smith et al., 2022; Thoppilan et al., 2022) , with the largest dense language models now having over 500 billion parameters","reference_id":"MzgyNTUyOTMzNTA5MjIzOTgw"} +{"index":744,"citer_arxiv_id":"2009.03300","cited_arxiv_id":"1803.05457","reference_contexts":"Most either cover easy topics like grade school subjects for which models can already achieve strong performance (Clark et al., 2018; Khot et al., 2019; Mihaylov et al., 2018; Clark et al., 2019) , or are focused on linguistic understanding in the form of reading comprehension (Lai et al., 2017; Richardson et al., 2013)","reference_id":"MTUzMjU1MDY3NDEyMDA0Njk3MTg="} +{"index":745,"citer_arxiv_id":"2009.03300","cited_arxiv_id":"1909.01958","reference_contexts":"Most either cover easy topics like grade school subjects for which models can already achieve strong performance (Clark et al., 2018; Khot et al., 2019; Mihaylov et al., 2018; Clark et al., 2019) , or are focused on linguistic understanding in the form of reading comprehension (Lai et al., 2017; Richardson et al., 2013)","reference_id":"MTUzMjU1MDY3NDEyMDA0Njk3MTg="} +{"index":746,"citer_arxiv_id":"2009.03300","cited_arxiv_id":"1810.04805","reference_contexts":"Until recently, researchers primarily used fine-tuned models on downstream tasks (Devlin et al., 2019)","reference_id":"MTA3MTkyMzY4NjA0MDk3ODg4NTE="} +{"index":747,"citer_arxiv_id":"2009.03300","cited_arxiv_id":"1907.07174","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":748,"citer_arxiv_id":"2009.03300","cited_arxiv_id":"1909.11942","reference_contexts":"et al., 2019), ALBERT-xxlarge (223 million parameters) (Lan et al., 2020) , and GPT-2 (1,558 million parameters) (Radford et al., 2019)","reference_id":"MTMwOTQ4MzU2NDQ3ODUzOTM5MTQ="} +{"index":749,"citer_arxiv_id":"2009.03300","cited_arxiv_id":"1907.11692","reference_contexts":"After fine-tuning a RoBERTa-base model (Liu et al., 2019) using this custom training set, our model attained 32.8% test accuracy","reference_id":"MTIxNzEzNDc0MjEyMjIyOTE5OTE="} +{"index":750,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1606.06565","reference_contexts":"These vulnerabilities Cummings, 2004; Scharre, 2018 Biggio et al., 2012 Szegedy et al., 2013 Amodei, Olah, et al., 2016 p.1 8","reference_id":"NDk3NDk2MzA4NzUwNzM4NDY5OA=="} +{"index":751,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1406.2661","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":752,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1606.08813","reference_contexts":"As an example of policymaking in this domain that has surfaced several difficulties, the European Union's General Data Protection Regulation is a commonly-discussed example of a policy that is hard to interpret and apply in the context of current machine learning algorithms (Goodman and Flaxman, 2016) .","reference_id":"MTYzNTcwODk4OTc1NjU0MDA0Ng=="} +{"index":753,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1410.8516","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":754,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1705.08417","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":755,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1707.08945","reference_contexts":"Szegedy et al., 2013; Papernot et al., 2016; Evtimov et al., 2017; Carlini et al., 2016 Rubinstein et al., 2009; Šrndic and Laskov, 2014 e.g","reference_id":"MTU4Nzc2ODk3MzA0MTEwNTIyNQ=="} +{"index":756,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1611.05397","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":757,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1702.01135","reference_contexts":"See also Katz et al., 2017; Selsam, Liang, and Dill, 2017; and Carlini et al., 2017","reference_id":"MjE4NTE1MzI1MzA4MTkzNDQxMw=="} +{"index":758,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1708.06733","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":759,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1705.08807","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":760,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1704.03732","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":761,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1703.04730","reference_contexts":"Koh and Liang (2017) for related ideas)","reference_id":"Mjg2NjQ5Nzg4Nzc0NzE4ODY3OA=="} +{"index":762,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1602.02697","reference_contexts":"Bass, 2017 Papernot et al., 2016b Knight, 2017 though see e.g","reference_id":"ODMwOTk2NDc4MDU4ODA1NjM5NA=="} +{"index":763,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1511.06434","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":764,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1709.02349","reference_contexts":"Seymour and Tully, 2016 Everett et al., 2016 Shao et al., 2017 Chung, Jamaludin, and Zisserman, 2017 Browne, 2017 Adams, 2017; Serban et al., 2017","reference_id":"Nzk3MDUwODA3NTA1NDIxMTM1NA=="} +{"index":765,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1312.6199","reference_contexts":"These vulnerabilities Cummings, 2004; Scharre, 2018 Biggio et al., 2012 Szegedy et al., 2013 Amodei, Olah, et al., 2016 p.1 8","reference_id":"NDk3NDk2MzA4NzUwNzM4NDY5OA=="} +{"index":766,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1312.6199","reference_contexts":"Szegedy et al., 2013; Papernot et al., 2016; Evtimov et al., 2017; Carlini et al., 2016 Rubinstein et al., 2009; Šrndic and Laskov, 2014 e.g","reference_id":"MTU4Nzc2ODk3MzA0MTEwNTIyNQ=="} +{"index":767,"citer_arxiv_id":"1802.07228","cited_arxiv_id":"1703.01161","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":768,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2012.05672","reference_contexts":"There is also a related line of work on instruction following for navigation, where models are trained to follow natural language instructions to navigate in a simulated environment (Bahdanau et al., 2018; Abramson et al., 2020; Zhao et al., 2021) .","reference_id":"MTU0OTI2MzkxMzI5NDk5ODMwMDQ="} +{"index":769,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1705.08439","reference_contexts":"For example, one could explore expert iteration (Anthony et al., 2017; Silver et al., 2017) , or simpler behavior cloning methods that use a subset of the comparison data","reference_id":"MjU3NzAxMDQyMzg0MzI1NDg1Mg=="} +{"index":770,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2111.10952","reference_contexts":"There has been a range of work in this domain (Yi et al., 2019; Mishra et al., 2021; Wei et al., 2021; Khashabi et al., 2020; Sanh et al., 2021; Aribandi et al., 2021) , which differ in training and evaluation data, formatting of instructions, size of pretrained models, and other experimental details","reference_id":"NjcwNjQwMzAxNTAyODgwNDc2NA=="} +{"index":771,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2112.00861","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":772,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1607.07086","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":773,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1806.01946","reference_contexts":"There is also a related line of work on instruction following for navigation, where models are trained to follow natural language instructions to navigate in a simulated environment (Bahdanau et al., 2018; Abramson et al., 2020; Zhao et al., 2021) .","reference_id":"MTU0OTI2MzkxMzI5NDk5ODMwMDQ="} +{"index":774,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2005.14050","reference_contexts":"Making significant progress on these problems is hard since well-intentioned interventions on LM behavior can have side-effects (Welbl et al., 2021; Blodgett et al., 2020) ; for instance, efforts to reduce the toxicity of LMs can reduce their ability to model text from under-represented groups, due to prejudicial correlations in the training data (Xu et al., 2021) .","reference_id":"MTgyNzk4MTMyNDkxMzc1NDEwMDQ="} +{"index":775,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.01214","reference_contexts":"Originally developed for training simple robots in simulated environments and Atari games (Christiano et al., 2017; Ibarz et al., 2018) , it has recently been applied to fine-tuning language models to summarize text (Ziegler et al., 2019; Stiennon et al., 2020; Böhm et al., 2019; Wu et al., 2021)","reference_id":"NTUyODE1NjA0NTM1MTM4Nzk2Ng=="} +{"index":776,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2108.07258","reference_contexts":"However, these models often express unintended behaviors such as making up facts, generating biased or toxic text, or simply not following user instructions (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021; Gehman et al., 2020)","reference_id":"OTMxMzE0NjE0NzU3NTAzNTE1"} +{"index":777,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2108.07258","reference_contexts":"These risks have been extensively documented (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021)","reference_id":"NTg5Nzc1NDA1MTk4NTE1MDk4OQ=="} +{"index":778,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2005.14165","reference_contexts":"We start with a pretrained language model (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022) , a distribution of prompts on which we want our model to produce aligned outputs, and a team of trained human labelers (see Sections 3.4 for details)","reference_id":"Mjg0NTE1NzAyNDEzMDA3MjQ3OQ=="} +{"index":779,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2005.14165","reference_contexts":"The cost of collecting our data and the compute for training runs, including experimental runs is a fraction of what was spent to train GPT-3: training our 175B SFT model requires 4.9 petaflops/s-days and training our 175B PPO-ptx model requires 60 petaflops/s-days, compared to 3,640 petaflops/s-days for GPT-3 (Brown et al., 2020)","reference_id":"OTYwOTA5MDkyNjc5NTY3NzA1Ng=="} +{"index":780,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2005.14165","reference_contexts":"All model architectures use the GPT-3 architecture (Brown et al., 2020)","reference_id":"ODUzMDE5MDk0ODkyNTE1MjcwNw=="} +{"index":781,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2005.14165","reference_contexts":"used for many recent large LMs-predicting the next token on a webpage from the internet-is different from the objective \"follow the user's instructions helpfully and safely\" (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022)","reference_id":"MTI1NTkzMzI4NTA1NTk4NTIyMDE="} +{"index":782,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2107.03374","reference_contexts":"The definition of alignment has historically been a vague and confusing topic, with various competing proposals (Chen et al., 2021; Leike et al., 2018; Gabriel, 2020)","reference_id":"MTgzNDc4NDU5MDM5NDg1NDE1MTY="} +{"index":783,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1811.00511","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":784,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1810.08575","reference_contexts":"Alignment research has historically been rather abstract, focusing on either theoretical results (Soares et al., 2015) , small synthetic domains (Christiano et al., 2018; Leike et al., 2017) , or training ML models on public NLP datasets (Ziegler et al., 2019; Stiennon et al., 2020)","reference_id":"MTA3OTgwNjM4ODM3Mjg0MzQ0NTc="} +{"index":785,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1810.08575","reference_contexts":"Moreover, the alignment technique we use here, RLHF, is an important building block in several proposals to align superhuman systems (Leike et al., 2018; Irving et al., 2018; Christiano et al., 2018)","reference_id":"MTI1NzE0OTA2ODA2NTY2NDg4NTc="} +{"index":786,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1912.02164","reference_contexts":"Getting models to do what we want is directly related to the steerability and controllability literature (Dathathri et al., 2019; Krause et al., 2020)","reference_id":"NjQzNTg4NjA0MzM5NjgyNzcyNw=="} +{"index":787,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1912.02164","reference_contexts":"A promising future path is combining RLHF with other methods of steerability, for example using control codes (Keskar et al., 2019) , or modifying the sampling procedure at inference time using a smaller model (Dathathri et al., 2019) .","reference_id":"MTgyNzA0OTUyMjA1NTUwOTAyMTY="} +{"index":788,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1912.02164","reference_contexts":"There is also work on steering the generation of language models using a second (usually smaller) language model (Dathathri et al., 2019; Krause et al., 2020) , and variants of this idea have been applied to reducing language model toxicity (Schick et al., 2021)","reference_id":"MTc0NTU1MDM4MDk5ODg3NTY3Nw=="} +{"index":789,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1911.03842","reference_contexts":"(2020) use a variety of approaches to improve the safety of chatbots, including data filtering, blocking certain words or n-grams during generation, safety-specific control tokens (Keskar et al., 2019; Dinan et al., 2019a) , and human-in-theloop data collection (Dinan et al., 2019b)","reference_id":"NzIxNDc1ODM5OTUwMDU5MTQ2MQ=="} +{"index":790,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1911.03842","reference_contexts":"Other approaches for mitigating the generated bias by LMs use word embedding regularization (Liu et al., 2019; Huang et al., 2019) , data augmentation (Liu et al., 2019; Dinan et al., 2019a; Sheng et al., 2019) , null space projection to make the distribution over sensitive tokens more uniform (Liang et al., 2021) , different objective functions (Qian et al., 2019) , or causal mediation analysis (Vig et al., 2020)","reference_id":"MTcyNTk3ODY4MTE4OTYxNTMxMjI="} +{"index":791,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1908.06083","reference_contexts":"(2020) use a variety of approaches to improve the safety of chatbots, including data filtering, blocking certain words or n-grams during generation, safety-specific control tokens (Keskar et al., 2019; Dinan et al., 2019a) , and human-in-theloop data collection (Dinan et al., 2019b)","reference_id":"NzIxNDc1ODM5OTUwMDU5MTQ2MQ=="} +{"index":792,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1908.06083","reference_contexts":"We believe both these behaviors could be dramatically reduced with adversarial data collection (Dinan et al., 2019b) .","reference_id":"ODkwODYwMjg5ODM3NTM1OTg2OQ=="} +{"index":793,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1908.06083","reference_contexts":"For example, one could use an adversarial set-up where labelers find the worst-case behaviors of the model, which are then labeled and added to the dataset (Dinan et al., 2019b)","reference_id":"MzcwNjUzOTEzNzU3NjA5MjI3Mw=="} +{"index":794,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1908.06083","reference_contexts":"Deploying language models in specific domains gives rise to new risks and challenges, for example in dialog systems (Henderson et al., 2018; Xu et al., 2020; Dinan et al., 2019b)","reference_id":"MTI0Mjg5NzA4MjU2MjI3MTM0MjA="} +{"index":795,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1903.00161","reference_contexts":"We perform automatic evaluations on the following benchmark datasets: Winogender (Rudinger et al., 2018) , CrowS-Pairs (Nangia et al., 2020) , RealToxicityPrompts (Gehman et al., 2020) , TruthfulQA (Lin et al., 2021) , DROP (Dua et al., 2019) , QuAC (Choi et al., 2018) , SquadV2 (Rajpurkar et al., 2018) , Hellaswag (Zellers et al., 2019) , SST (Socher et al., 2013) , RTE and WSC (both part of Super-GLUE (Wang et al., 2019) ), WMT 15 Fr → En (Bojar et al., 2015) , CNN/Daily Mail Summarization (Nallapati et al., 2016) , and Reddit TLDR Summarization (Völske et al., 2017) .","reference_id":"MjIyNDIxNzI2OTgzMDAxODA0Nw=="} +{"index":796,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1903.00161","reference_contexts":"During RLHF fine-tuning, we observe performance regressions compared to GPT-3 on certain public NLP datasets, notably SQuAD (Rajpurkar et al., 2018) , DROP (Dua et al., 2019) , HellaSwag (Zellers et al., 2019) , and WMT 2015 French to English translation (Bojar et al., 2015)","reference_id":"MTI5MzM0ODY4NjU4MTE2ODkxNjE="} +{"index":797,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2101.03961","reference_contexts":"We start with a pretrained language model (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022) , a distribution of prompts on which we want our model to produce aligned outputs, and a team of trained human labelers (see Sections 3.4 for details)","reference_id":"Mjg0NTE1NzAyNDEzMDA3MjQ3OQ=="} +{"index":798,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2101.03961","reference_contexts":"used for many recent large LMs-predicting the next token on a webpage from the internet-is different from the objective \"follow the user's instructions helpfully and safely\" (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022)","reference_id":"MTI1NTkzMzI4NTA1NTk4NTIyMDE="} +{"index":799,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.11462","reference_contexts":"However, these models often express unintended behaviors such as making up facts, generating biased or toxic text, or simply not following user instructions (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021; Gehman et al., 2020)","reference_id":"OTMxMzE0NjE0NzU3NTAzNTE1"} +{"index":800,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.11462","reference_contexts":"We perform automatic evaluations on the following benchmark datasets: Winogender (Rudinger et al., 2018) , CrowS-Pairs (Nangia et al., 2020) , RealToxicityPrompts (Gehman et al., 2020) , TruthfulQA (Lin et al., 2021) , DROP (Dua et al., 2019) , QuAC (Choi et al., 2018) , SquadV2 (Rajpurkar et al., 2018) , Hellaswag (Zellers et al., 2019) , SST (Socher et al., 2013) , RTE and WSC (both part of Super-GLUE (Wang et al., 2019) ), WMT 15 Fr → En (Bojar et al., 2015) , CNN/Daily Mail Summarization (Nallapati et al., 2016) , and Reddit TLDR Summarization (Völske et al., 2017) .","reference_id":"MjIyNDIxNzI2OTgzMDAxODA0Nw=="} +{"index":801,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.11462","reference_contexts":"To measure toxicity, we use the RealToxicityPrompts dataset (Gehman et al., 2020) and conduct both automatic and human evaluations","reference_id":"MTE0MzQ2OTg2OTUxNTEzNjI2ODY="} +{"index":802,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.11462","reference_contexts":"We first evaluate our models on the RealToxicityPrompts dataset (Gehman et al., 2020)","reference_id":"MTU1MDY3OTQzMDAwMDUzNzA3MTE="} +{"index":803,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.11462","reference_contexts":"There is a nascent but growing field that aims to build benchmarks to concretely evaluate these harms, particularly around toxicity (Gehman et al., 2020) , stereotypes (Nadeem et al., 2020) , and social bias (Dhamala et al., 2021; Nangia et al., 2020; Rudinger et al., 2018)","reference_id":"MTQ2ODk1MjQ1MTMzMTEwMzMwNjI="} +{"index":804,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.11462","reference_contexts":"We also conduct human evaluations of toxicity on the RealToxicityPrompts dataset (Gehman et al., 2020)","reference_id":"NzI4NDI3NjkzNTU4ODI3NzQ5Mw=="} +{"index":805,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.11462","reference_contexts":"We also benchmark our model on datasets intended to measure bias and toxicity, such as RealToxicityPrompts (Gehman et al., 2020) and CrowS-Pairs (Nangia et al., 2020) .","reference_id":"NDY0MTk5OTA3ODE1MjIyNDMxMA=="} +{"index":806,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1901.05415","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":807,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1911.03064","reference_contexts":"Other approaches for mitigating the generated bias by LMs use word embedding regularization (Liu et al., 2019; Huang et al., 2019) , data augmentation (Liu et al., 2019; Dinan et al., 2019a; Sheng et al., 2019) , null space projection to make the distribution over sensitive tokens more uniform (Liang et al., 2021) , different objective functions (Qian et al., 2019) , or causal mediation analysis (Vig et al., 2020)","reference_id":"MTcyNTk3ODY4MTE4OTYxNTMxMjI="} +{"index":808,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1805.00899","reference_contexts":"Moreover, the alignment technique we use here, RLHF, is an important building block in several proposals to align superhuman systems (Leike et al., 2018; Irving et al., 2018; Christiano et al., 2018)","reference_id":"MTI1NzE0OTA2ODA2NTY2NDg4NTc="} +{"index":809,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1907.00456","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":810,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2103.14659","reference_contexts":"However, these models often express unintended behaviors such as making up facts, generating biased or toxic text, or simply not following user instructions (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021; Gehman et al., 2020)","reference_id":"OTMxMzE0NjE0NzU3NTAzNTE1"} +{"index":811,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2103.14659","reference_contexts":"These risks have been extensively documented (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021)","reference_id":"NTg5Nzc1NDA1MTk4NTE1MDk4OQ=="} +{"index":812,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.05858","reference_contexts":"(2020) use a variety of approaches to improve the safety of chatbots, including data filtering, blocking certain words or n-grams during generation, safety-specific control tokens (Keskar et al., 2019; Dinan et al., 2019a) , and human-in-theloop data collection (Dinan et al., 2019b)","reference_id":"NzIxNDc1ODM5OTUwMDU5MTQ2MQ=="} +{"index":813,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.05858","reference_contexts":"A promising future path is combining RLHF with other methods of steerability, for example using control codes (Keskar et al., 2019) , or modifying the sampling procedure at inference time using a smaller model (Dathathri et al., 2019) .","reference_id":"MTgyNzA0OTUyMjA1NTUwOTAyMTY="} +{"index":814,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2005.00700","reference_contexts":"There has been a range of work in this domain (Yi et al., 2019; Mishra et al., 2021; Wei et al., 2021; Khashabi et al., 2020; Sanh et al., 2021; Aribandi et al., 2021) , which differ in training and evaluation data, formatting of instructions, size of pretrained models, and other experimental details","reference_id":"NjcwNjQwMzAxNTAyODgwNDc2NA=="} +{"index":815,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.06367","reference_contexts":"Getting models to do what we want is directly related to the steerability and controllability literature (Dathathri et al., 2019; Krause et al., 2020)","reference_id":"NjQzNTg4NjA0MzM5NjgyNzcyNw=="} +{"index":816,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.06367","reference_contexts":"There is also work on steering the generation of language models using a second (usually smaller) language model (Dathathri et al., 2019; Krause et al., 2020) , and variants of this idea have been applied to reducing language model toxicity (Schick et al., 2021)","reference_id":"MTc0NTU1MDM4MDk5ODg3NTY3Nw=="} +{"index":817,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1804.05958","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":818,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1805.01252","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":819,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1811.07871","reference_contexts":"The definition of alignment has historically been a vague and confusing topic, with various competing proposals (Chen et al., 2021; Leike et al., 2018; Gabriel, 2020)","reference_id":"MTgzNDc4NDU5MDM5NDg1NDE1MTY="} +{"index":820,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1811.07871","reference_contexts":"Moreover, the alignment technique we use here, RLHF, is an important building block in several proposals to align superhuman systems (Leike et al., 2018; Irving et al., 2018; Christiano et al., 2018)","reference_id":"MTI1NzE0OTA2ODA2NTY2NDg4NTc="} +{"index":821,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1811.07871","reference_contexts":"Even though this work focuses on our current language model systems, we seek general and scalable methods that work for future AI systems (Leike et al., 2018)","reference_id":"MTUxNzQ1ODY1ODk5OTAwNTE2Mzc="} +{"index":822,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1811.07871","reference_contexts":"We make progress on aligning language models by training them to act in accordance with the user's intention (Leike et al., 2018)","reference_id":"NTk3NzUxODk0MDg2NTA2Mjk5OA=="} +{"index":823,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1711.09883","reference_contexts":"Alignment research has historically been rather abstract, focusing on either theoretical results (Soares et al., 2015) , small synthetic domains (Christiano et al., 2018; Leike et al., 2017) , or training ML models on public NLP datasets (Ziegler et al., 2019; Stiennon et al., 2020)","reference_id":"MTA3OTgwNjM4ODM3Mjg0MzQ0NTc="} +{"index":824,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.07958","reference_contexts":"We perform automatic evaluations on the following benchmark datasets: Winogender (Rudinger et al., 2018) , CrowS-Pairs (Nangia et al., 2020) , RealToxicityPrompts (Gehman et al., 2020) , TruthfulQA (Lin et al., 2021) , DROP (Dua et al., 2019) , QuAC (Choi et al., 2018) , SquadV2 (Rajpurkar et al., 2018) , Hellaswag (Zellers et al., 2019) , SST (Socher et al., 2013) , RTE and WSC (both part of Super-GLUE (Wang et al., 2019) ), WMT 15 Fr → En (Bojar et al., 2015) , CNN/Daily Mail Summarization (Nallapati et al., 2016) , and Reddit TLDR Summarization (Völske et al., 2017) .","reference_id":"MjIyNDIxNzI2OTgzMDAxODA0Nw=="} +{"index":825,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.07958","reference_contexts":"Instead, we measure truthfulness-whether the model's statements about the world are true-using two metrics: (1) evaluating our model's tendency to make up information on closed domain tasks (\"hallucinations\"), and (2) using the TruthfulQA dataset (Lin et al., 2021)","reference_id":"MTEyNTY4NDg0NjE4NzczNjY3MDY="} +{"index":826,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1910.10486","reference_contexts":"Other approaches for mitigating the generated bias by LMs use word embedding regularization (Liu et al., 2019; Huang et al., 2019) , data augmentation (Liu et al., 2019; Dinan et al., 2019a; Sheng et al., 2019) , null space projection to make the distribution over sensitive tokens more uniform (Liang et al., 2021) , different objective functions (Qian et al., 2019) , or causal mediation analysis (Vig et al., 2020)","reference_id":"MTcyNTk3ODY4MTE4OTYxNTMxMjI="} +{"index":827,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2104.08773","reference_contexts":"There has been a range of work in this domain (Yi et al., 2019; Mishra et al., 2021; Wei et al., 2021; Khashabi et al., 2020; Sanh et al., 2021; Aribandi et al., 2021) , which differ in training and evaluation data, formatting of instructions, size of pretrained models, and other experimental details","reference_id":"NjcwNjQwMzAxNTAyODgwNDc2NA=="} +{"index":828,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2004.09456","reference_contexts":"There is a nascent but growing field that aims to build benchmarks to concretely evaluate these harms, particularly around toxicity (Gehman et al., 2020) , stereotypes (Nadeem et al., 2020) , and social bias (Dhamala et al., 2021; Nangia et al., 2020; Rudinger et al., 2018)","reference_id":"MTQ2ODk1MjQ1MTMzMTEwMzMwNjI="} +{"index":829,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2112.09332","reference_contexts":"Similarly, one could combine our approach with methods that improve models' truthfulness, such as WebGPT (Nakano et al., 2021) .","reference_id":"NDM2MzM3NzM1NzE1MjU1MDU3OA=="} +{"index":830,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1602.06023","reference_contexts":"We perform automatic evaluations on the following benchmark datasets: Winogender (Rudinger et al., 2018) , CrowS-Pairs (Nangia et al., 2020) , RealToxicityPrompts (Gehman et al., 2020) , TruthfulQA (Lin et al., 2021) , DROP (Dua et al., 2019) , QuAC (Choi et al., 2018) , SquadV2 (Rajpurkar et al., 2018) , Hellaswag (Zellers et al., 2019) , SST (Socher et al., 2013) , RTE and WSC (both part of Super-GLUE (Wang et al., 2019) ), WMT 15 Fr → En (Bojar et al., 2015) , CNN/Daily Mail Summarization (Nallapati et al., 2016) , and Reddit TLDR Summarization (Völske et al., 2017) .","reference_id":"MjIyNDIxNzI2OTgzMDAxODA0Nw=="} +{"index":831,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.05863","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":832,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1905.12801","reference_contexts":"Other approaches for mitigating the generated bias by LMs use word embedding regularization (Liu et al., 2019; Huang et al., 2019) , data augmentation (Liu et al., 2019; Dinan et al., 2019a; Sheng et al., 2019) , null space projection to make the distribution over sensitive tokens more uniform (Liang et al., 2021) , different objective functions (Qian et al., 2019) , or causal mediation analysis (Vig et al., 2020)","reference_id":"MTcyNTk3ODY4MTE4OTYxNTMxMjI="} +{"index":833,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2112.11446","reference_contexts":"We start with a pretrained language model (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022) , a distribution of prompts on which we want our model to produce aligned outputs, and a team of trained human labelers (see Sections 3.4 for details)","reference_id":"Mjg0NTE1NzAyNDEzMDA3MjQ3OQ=="} +{"index":834,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2112.11446","reference_contexts":"used for many recent large LMs-predicting the next token on a webpage from the internet-is different from the objective \"follow the user's instructions helpfully and safely\" (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022)","reference_id":"MTI1NTkzMzI4NTA1NTk4NTIyMDE="} +{"index":835,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1806.03822","reference_contexts":"We perform automatic evaluations on the following benchmark datasets: Winogender (Rudinger et al., 2018) , CrowS-Pairs (Nangia et al., 2020) , RealToxicityPrompts (Gehman et al., 2020) , TruthfulQA (Lin et al., 2021) , DROP (Dua et al., 2019) , QuAC (Choi et al., 2018) , SquadV2 (Rajpurkar et al., 2018) , Hellaswag (Zellers et al., 2019) , SST (Socher et al., 2013) , RTE and WSC (both part of Super-GLUE (Wang et al., 2019) ), WMT 15 Fr → En (Bojar et al., 2015) , CNN/Daily Mail Summarization (Nallapati et al., 2016) , and Reddit TLDR Summarization (Völske et al., 2017) .","reference_id":"MjIyNDIxNzI2OTgzMDAxODA0Nw=="} +{"index":836,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1806.03822","reference_contexts":"During RLHF fine-tuning, we observe performance regressions compared to GPT-3 on certain public NLP datasets, notably SQuAD (Rajpurkar et al., 2018) , DROP (Dua et al., 2019) , HellaSwag (Zellers et al., 2019) , and WMT 2015 French to English translation (Bojar et al., 2015)","reference_id":"MTI5MzM0ODY4NjU4MTE2ODkxNjE="} +{"index":837,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2110.08207","reference_contexts":"InstructGPT) to GPT-3 fine-tuned on two different compilations of public NLP tasks: the FLAN (Wei et al., 2021) and T0 (Sanh et al., 2021) (in particular, the T0++ variant)","reference_id":"NzQ4ODgxNTEyMzMzMzA0MjE1"} +{"index":838,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2110.08207","reference_contexts":"6 We additionally compare InstructGPT to fine-tuning 175B GPT-3 on the FLAN (Wei et al., 2021) and T0 (Sanh et al., 2021) datasets, which both consist of a variety of NLP tasks, combined with natural language instructions for each task (the datasets differ in the NLP datasets included, and the style of instructions used)","reference_id":"OTY1MDEyOTYzMTczMzA3Mjk0Nw=="} +{"index":839,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2110.08207","reference_contexts":"In Figure 5 , we also compare InstructGPT to our 175B GPT-3 baselines fine-tuned on the FLAN (Wei et al., 2021) and T0 (Sanh et al., 2021) datasets (see Appendix C for details)","reference_id":"NzMwMDI1Mjk2MDgzMDg3NDY2OA=="} +{"index":840,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2110.08207","reference_contexts":"There has been a range of work in this domain (Yi et al., 2019; Mishra et al., 2021; Wei et al., 2021; Khashabi et al., 2020; Sanh et al., 2021; Aribandi et al., 2021) , which differ in training and evaluation data, formatting of instructions, size of pretrained models, and other experimental details","reference_id":"NjcwNjQwMzAxNTAyODgwNDc2NA=="} +{"index":841,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2103.00453","reference_contexts":"There is also work on steering the generation of language models using a second (usually smaller) language model (Dathathri et al., 2019; Krause et al., 2020) , and variants of this idea have been applied to reducing language model toxicity (Schick et al., 2021)","reference_id":"MTc0NTU1MDM4MDk5ODg3NTY3Nw=="} +{"index":842,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1707.06347","reference_contexts":"In other words, each batch is randomly split into 8 minibatches and is trained on for only a single inner epoch (Schulman et al., 2017)","reference_id":"ODUwNDkxOTUxNjQ3MDUzNzAyOA=="} +{"index":843,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1707.06347","reference_contexts":"Finally, we use this RM as a reward function and fine-tune our supervised learning baseline to maximize this reward using the PPO algorithm (Schulman et al., 2017)","reference_id":"MTY4ODQ3NDc5MzE3ODU4MDk1Njc="} +{"index":844,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1707.06347","reference_contexts":"(2020) , we fine-tuned the SFT model on our environment using PPO (Schulman et al., 2017)","reference_id":"NDg1ODgxMTU4NjI4MzEwMzIwMw=="} +{"index":845,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1707.06347","reference_contexts":"We fine-tune the supervised policy to optimize this reward using the PPO algorithm (Schulman et al., 2017) .","reference_id":"Njk2OTQ2MDY0OTQ4Mzc4MzM4Nw=="} +{"index":846,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.01326","reference_contexts":"Other approaches for mitigating the generated bias by LMs use word embedding regularization (Liu et al., 2019; Huang et al., 2019) , data augmentation (Liu et al., 2019; Dinan et al., 2019a; Sheng et al., 2019) , null space projection to make the distribution over sensitive tokens more uniform (Liang et al., 2021) , different objective functions (Qian et al., 2019) , or causal mediation analysis (Vig et al., 2020)","reference_id":"MTcyNTk3ODY4MTE4OTYxNTMxMjI="} +{"index":847,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1712.01815","reference_contexts":"For example, one could explore expert iteration (Anthony et al., 2017; Silver et al., 2017) , or simpler behavior cloning methods that use a subset of the comparison data","reference_id":"MjU3NzAxMDQyMzg0MzI1NDg1Mg=="} +{"index":848,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1908.09203","reference_contexts":"Language models can produce biased outputs (Dhamala et al., 2021; Liang et al., 2021; Manela et al., 2021; Caliskan et al., 2017; Kirk et al., 2021) , leak private data (Carlini et al., 2021) , generate misinformation (Solaiman et al., 2019; Buchanan et al., 2021) , and be used maliciously; for a thorough review we direct the reader to Weidinger et al","reference_id":"MTM1ODc3MjUxNDExNDEwNTIzMTE="} +{"index":849,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2106.10328","reference_contexts":"Solaiman and Dennison (2021) fine-tune LMs on a small, value-targeted dataset, which improves the models' ability to adhere to these values on a question answering task","reference_id":"NTU1ODY5NTA4MDkwMjA4Mjg0Mw=="} +{"index":850,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.01325","reference_contexts":"Compared to earlier work that collects human preference data on the task of summarization (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021) , our inputs span a much broader range of tasks, and can occasionally include controversial and sensitive topics","reference_id":"MTY5NzkyOTkyNzAxNjYwNDAzNTQ="} +{"index":851,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.01325","reference_contexts":"Alignment research has historically been rather abstract, focusing on either theoretical results (Soares et al., 2015) , small synthetic domains (Christiano et al., 2018; Leike et al., 2017) , or training ML models on public NLP datasets (Ziegler et al., 2019; Stiennon et al., 2020)","reference_id":"MTA3OTgwNjM4ODM3Mjg0MzQ0NTc="} +{"index":852,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.01325","reference_contexts":"Specifically, we use reinforcement learning from human feedback (RLHF; Christiano et al., 2017; Stiennon et al., 2020) to fine-tune GPT-3 to follow a broad class of written instructions (see Figure 2 )","reference_id":"NDY5MjA1ODYxMDkyNDQwNDA2OA=="} +{"index":853,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.01325","reference_contexts":"Originally developed for training simple robots in simulated environments and Atari games (Christiano et al., 2017; Ibarz et al., 2018) , it has recently been applied to fine-tuning language models to summarize text (Ziegler et al., 2019; Stiennon et al., 2020; Böhm et al., 2019; Wu et al., 2021)","reference_id":"NTUyODE1NjA0NTM1MTM4Nzk2Ng=="} +{"index":854,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2009.01325","reference_contexts":"This research is part of our broader research program to align AI systems with human intentions (Christiano et al., 2017; Ziegler et al., 2019; Stiennon et al., 2020)","reference_id":"MzA5MzgwNTY0ODgyMTY4MzI0NQ=="} +{"index":855,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2102.02503","reference_contexts":"However, these models often express unintended behaviors such as making up facts, generating biased or toxic text, or simply not following user instructions (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021; Gehman et al., 2020)","reference_id":"OTMxMzE0NjE0NzU3NTAzNTE1"} +{"index":856,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2102.02503","reference_contexts":"These risks have been extensively documented (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021)","reference_id":"NTg5Nzc1NDA1MTk4NTE1MDk4OQ=="} +{"index":857,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2201.08239","reference_contexts":"We start with a pretrained language model (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022) , a distribution of prompts on which we want our model to produce aligned outputs, and a team of trained human labelers (see Sections 3.4 for details)","reference_id":"Mjg0NTE1NzAyNDEzMDA3MjQ3OQ=="} +{"index":858,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2201.08239","reference_contexts":"used for many recent large LMs-predicting the next token on a webpage from the internet-is different from the objective \"follow the user's instructions helpfully and safely\" (Radford et al., 2019; Brown et al., 2020; Fedus et al., 2021; Rae et al., 2021; Thoppilan et al., 2022)","reference_id":"MTI1NTkzMzI4NTA1NTk4NTIyMDE="} +{"index":859,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1905.00537","reference_contexts":"We perform automatic evaluations on the following benchmark datasets: Winogender (Rudinger et al., 2018) , CrowS-Pairs (Nangia et al., 2020) , RealToxicityPrompts (Gehman et al., 2020) , TruthfulQA (Lin et al., 2021) , DROP (Dua et al., 2019) , QuAC (Choi et al., 2018) , SquadV2 (Rajpurkar et al., 2018) , Hellaswag (Zellers et al., 2019) , SST (Socher et al., 2013) , RTE and WSC (both part of Super-GLUE (Wang et al., 2019) ), WMT 15 Fr → En (Bojar et al., 2015) , CNN/Daily Mail Summarization (Nallapati et al., 2016) , and Reddit TLDR Summarization (Völske et al., 2017) .","reference_id":"MjIyNDIxNzI2OTgzMDAxODA0Nw=="} +{"index":860,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.01652","reference_contexts":"InstructGPT) to GPT-3 fine-tuned on two different compilations of public NLP tasks: the FLAN (Wei et al., 2021) and T0 (Sanh et al., 2021) (in particular, the T0++ variant)","reference_id":"NzQ4ODgxNTEyMzMzMzA0MjE1"} +{"index":861,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.01652","reference_contexts":"6 We additionally compare InstructGPT to fine-tuning 175B GPT-3 on the FLAN (Wei et al., 2021) and T0 (Sanh et al., 2021) datasets, which both consist of a variety of NLP tasks, combined with natural language instructions for each task (the datasets differ in the NLP datasets included, and the style of instructions used)","reference_id":"OTY1MDEyOTYzMTczMzA3Mjk0Nw=="} +{"index":862,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.01652","reference_contexts":"In Figure 5 , we also compare InstructGPT to our 175B GPT-3 baselines fine-tuned on the FLAN (Wei et al., 2021) and T0 (Sanh et al., 2021) datasets (see Appendix C for details)","reference_id":"NzMwMDI1Mjk2MDgzMDg3NDY2OA=="} +{"index":863,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.01652","reference_contexts":"There has been a range of work in this domain (Yi et al., 2019; Mishra et al., 2021; Wei et al., 2021; Khashabi et al., 2020; Sanh et al., 2021; Aribandi et al., 2021) , which differ in training and evaluation data, formatting of instructions, size of pretrained models, and other experimental details","reference_id":"NjcwNjQwMzAxNTAyODgwNDc2NA=="} +{"index":864,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2112.04359","reference_contexts":"However, these models often express unintended behaviors such as making up facts, generating biased or toxic text, or simply not following user instructions (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021; Gehman et al., 2020)","reference_id":"OTMxMzE0NjE0NzU3NTAzNTE1"} +{"index":865,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2112.04359","reference_contexts":"These risks have been extensively documented (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021)","reference_id":"NTg5Nzc1NDA1MTk4NTE1MDk4OQ=="} +{"index":866,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.07445","reference_contexts":"Making significant progress on these problems is hard since well-intentioned interventions on LM behavior can have side-effects (Welbl et al., 2021; Blodgett et al., 2020) ; for instance, efforts to reduce the toxicity of LMs can reduce their ability to model text from under-represented groups, due to prejudicial correlations in the training data (Xu et al., 2021) .","reference_id":"MTgyNzk4MTMyNDkxMzc1NDEwMDQ="} +{"index":867,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.10862","reference_contexts":"For example, RLHF was a central method in recent work on summarizing books, a task that exhibits some of the difficulties of aligning superhuman AI systems as it is difficult for humans to evaluate directly (Wu et al., 2021) .","reference_id":"OTgzNDQ2NTU0MTYwMzI1NjUxNw=="} +{"index":868,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.10862","reference_contexts":"Compared to earlier work that collects human preference data on the task of summarization (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021) , our inputs span a much broader range of tasks, and can occasionally include controversial and sensitive topics","reference_id":"MTY5NzkyOTkyNzAxNjYwNDAzNTQ="} +{"index":869,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2109.10862","reference_contexts":"Originally developed for training simple robots in simulated environments and Atari games (Christiano et al., 2017; Ibarz et al., 2018) , it has recently been applied to fine-tuning language models to summarize text (Ziegler et al., 2019; Stiennon et al., 2020; Böhm et al., 2019; Wu et al., 2021)","reference_id":"NTUyODE1NjA0NTM1MTM4Nzk2Ng=="} +{"index":870,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2104.06390","reference_contexts":"Making significant progress on these problems is hard since well-intentioned interventions on LM behavior can have side-effects (Welbl et al., 2021; Blodgett et al., 2020) ; for instance, efforts to reduce the toxicity of LMs can reduce their ability to model text from under-represented groups, due to prejudicial correlations in the training data (Xu et al., 2021) .","reference_id":"MTgyNzk4MTMyNDkxMzc1NDEwMDQ="} +{"index":871,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2010.07079","reference_contexts":"Deploying language models in specific domains gives rise to new risks and challenges, for example in dialog systems (Henderson et al., 2018; Xu et al., 2020; Dinan et al., 2019b)","reference_id":"MTI0Mjg5NzA4MjU2MjI3MTM0MjA="} +{"index":872,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1904.13015","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":873,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1904.13015","reference_contexts":"There has been a range of work in this domain (Yi et al., 2019; Mishra et al., 2021; Wei et al., 2021; Khashabi et al., 2020; Sanh et al., 2021; Aribandi et al., 2021) , which differ in training and evaluation data, formatting of instructions, size of pretrained models, and other experimental details","reference_id":"NjcwNjQwMzAxNTAyODgwNDc2NA=="} +{"index":874,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"2002.05058","reference_contexts":"This work is in turn influenced by similar work using human feedback as a reward in domains such as dialogue (Jaques et al., 2019; Yi et al., 2019; Hancock et al., 2019) , translation (Kreutzer et al., 2018; Bahdanau et al., 2016) , semantic parsing (Lawrence and Riezler, 2018) , story generation (Zhou and Xu, 2020) , review generation (Cho et al., 2018) , and evidence extraction (Perez et al., 2019)","reference_id":"MTYyOTg2ODAwODczNTUzNjkwMDQ="} +{"index":875,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.08593","reference_contexts":"Compared to earlier work that collects human preference data on the task of summarization (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021) , our inputs span a much broader range of tasks, and can occasionally include controversial and sensitive topics","reference_id":"MTY5NzkyOTkyNzAxNjYwNDAzNTQ="} +{"index":876,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.08593","reference_contexts":"Alignment research has historically been rather abstract, focusing on either theoretical results (Soares et al., 2015) , small synthetic domains (Christiano et al., 2018; Leike et al., 2017) , or training ML models on public NLP datasets (Ziegler et al., 2019; Stiennon et al., 2020)","reference_id":"MTA3OTgwNjM4ODM3Mjg0MzQ0NTc="} +{"index":877,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.08593","reference_contexts":"Originally developed for training simple robots in simulated environments and Atari games (Christiano et al., 2017; Ibarz et al., 2018) , it has recently been applied to fine-tuning language models to summarize text (Ziegler et al., 2019; Stiennon et al., 2020; Böhm et al., 2019; Wu et al., 2021)","reference_id":"NTUyODE1NjA0NTM1MTM4Nzk2Ng=="} +{"index":878,"citer_arxiv_id":"2203.02155","cited_arxiv_id":"1909.08593","reference_contexts":"This research is part of our broader research program to align AI systems with human intentions (Christiano et al., 2017; Ziegler et al., 2019; Stiennon et al., 2020)","reference_id":"MzA5MzgwNTY0ODgyMTY4MzI0NQ=="} +{"index":879,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2005.14165","reference_contexts":"These dimensions are comparable to GPT-3 [8]","reference_id":"MjE3OTUwNTIxNTY3MjU5NTg3NA=="} +{"index":880,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2005.14165","reference_contexts":"Recent research has shown the benefit of scaling up the Transformer model [1, 8, 13, 18, 33] by increasing the number of layers and the sizes of dimensions M, H and N, which could reach hundreds of billions of parameters","reference_id":"NDE5NDE0MDQxNjg5ODM3MTE0Mg=="} +{"index":881,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2005.14165","reference_contexts":"This constraint is practical because a common way to scale up models is to stack layers of the same pattern [8, 18] .","reference_id":"OTc5NDYyNDk3MDExMzcyOTE5NA=="} +{"index":882,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"Pipeline parallelism partitions the training graph into multiple stages that run on different devices [18] to help build large models","reference_id":"MTM4OTQyMjIzNTE1MzE2ODMyOA=="} +{"index":883,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"Parallelism may be introduced in several ways: data parallelism [22] partitions training data, pipeline parallelism [18, 25, 27] partitions the computation graph, and within-layer model parallelism [36] partitions the weight and computation of each model layer.","reference_id":"MTQyODQyMzE4MDQ1NjU3ODE5MTk="} +{"index":884,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"Recent research has shown the benefit of scaling up the Transformer model [1, 8, 13, 18, 33] by increasing the number of layers and the sizes of dimensions M, H and N, which could reach hundreds of billions of parameters","reference_id":"NDE5NDE0MDQxNjg5ODM3MTE0Mg=="} +{"index":885,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"Pipelining requires data to be organized into microbatches; each stage loops over them one at a time [18]","reference_id":"NjY5ODA0NjM3OTIxNTE2NjUwNQ=="} +{"index":886,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"Pipelining algorithms [15, 18, 25, 27, 37] focus on one type of parallelism, while GSPMD can be used either to express similar ideas with the help of vectorization (Section 3.3), or to work in combination of these implementations by additionally partitioning each pipeline stage","reference_id":"MTQyNDExNTAwMzc0MjA0NzM2OTQ="} +{"index":887,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"For heterogeneous stages that cannot be supported, we recommend integrating GSPMD with other pipeline implementations [18, 25, 27] and sharding each stage","reference_id":"MTM0NzU2MzkyOTIyOTU0MzQ2MDU="} +{"index":888,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"The extra iterations are equivalent to the bubbles in earlier work [18] that describe the idle time due to data dependency, although the waiting devices compute on padded data instead of being idle.","reference_id":"NzIzNTcwODI2ODAxNTgxMjU1NQ=="} +{"index":889,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"We follow GPipe's approach [18] , where a key difference from non-pipelined configurations is that we recompute forward pass intermediate results during the backward pass, a technique known as rematerialization [10] to lower peak memory usage","reference_id":"MTY4NDkxMDUyOTM4MDc5MjAyNTY="} +{"index":890,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1811.06965","reference_contexts":"This constraint is practical because a common way to scale up models is to stack layers of the same pattern [8, 18] .","reference_id":"OTc5NDYyNDk3MDExMzcyOTE5NA=="} +{"index":891,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2006.16668","reference_contexts":"Partially Tiled: • Replicated in subgroups • Each subgroup has a different subset of data • Device mesh tensor , inner-most dim is replication • • Partially tiled (an extension to GShard [23] )","reference_id":"MTgyMTM1NDMwNTMzMDE2MjA3MTE="} +{"index":892,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2006.16668","reference_contexts":"It was originally used to work around cases where GSPMD was inefficient (see Section 3.2 in [23] ), but as the software matures and advanced optimizations are added, most of the use cases are no longer needed.","reference_id":"MTkxNTM5MjY0OTYzODg5MDA2NQ=="} +{"index":893,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2006.16668","reference_contexts":"GSPMD is generalized from the backend of GShard [23] based on our experiences of model scaling beyond the mixtureof-expert (MoE) use case, and it has helped Google to scale many deep learning models across several domains, including language (e.g., LamBDA [1] , GShard-M4 [24] ), image (e.g., MetNet-2 [14] ), and speech (e.g., BigSSL [41] )","reference_id":"MTIzNTc4Mjk3NzI4Nzk4OTIxMzE="} +{"index":894,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2006.16668","reference_contexts":"GSPMD is generalized from the back-end system used in GShard [23]","reference_id":"MTAxNTQyNzAzMDIwMDMzMjAwMQ=="} +{"index":895,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1910.02054","reference_contexts":"We present GSPMD, a system that uses simple tensor sharding annotations to achieve different parallelism paradigms in a unified way, including data parallelism, in-layer model parallelism, and novel strategies like image spatial partitioning [11] and weight-update/optimizer-state sharding [30, 40]","reference_id":"MTc1NzM3MTU1MjY5MTA1NTM1MDM="} +{"index":896,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1910.02054","reference_contexts":"Zero [30] presents a set of optimizations to reduce memory redundancy in parallel training devices, by partitioning weights, activations, and optimizer state separately","reference_id":"NDkzMzE0NjMzMjg5NDc2NDcxNg=="} +{"index":897,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1910.02054","reference_contexts":"Weight-update sharding or optimizer-state sharding is an enhancement to data parallelism where the computation to apply combined gradients onto weights is sharded across replica devices [30, 40]","reference_id":"ODU5MjkxNzkzMDc2MDUzMDgxNg=="} +{"index":898,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1910.02054","reference_contexts":"This behavior along X is conceptually the same as the weight-update/optimizer-state sharding technique [30, 40]","reference_id":"MTU3MzQ3MDE3OTUxMDM1Mjk0NzM="} +{"index":899,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1910.02054","reference_contexts":"If the user further partitions the weights along the other mesh dimension, i.e., df = mesh_split ( df , mesh , [0 , 1]) it will additionally trigger the weight-update sharding optimization [30, 40] where the weight will be unsharded on demand on the D dimension right before this layer in the forward pass to reduce peak memory usage, and gradients will be communicated with ReduceScatter instead of AllReduce in the backward pass and applied on a sharded optimizer.","reference_id":"MTEwODU1NTM5MTQwNzU4NjczODc="} +{"index":900,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1910.02054","reference_contexts":"This sharding configuration combines the benefit of data parallelism, weight-update sharding [30, 40] and in-layer model parallelism [36] , and only requires 7 annotations in the model","reference_id":"OTQ2ODM1NzAxMTcwMTk3ODA5MQ=="} +{"index":901,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"2102.12092","reference_contexts":"GSPMD as a shared, robust mechanism for different parallelism patterns is particularly relevant moving forward because the ML community is increasingly investing into multimodality, where text, image and audio are combined into a single model [31] .","reference_id":"OTc2ODE1OTE4Mzc0OTYzNTk5Mg=="} +{"index":902,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1701.06538","reference_contexts":"Mixture-of-experts (MoE) is a recently explored research direction in scaling up the Transformer model [13, 24, 34] , where the feed-forward layer provides many independent","reference_id":"NjI3ODc1NDM3OTc3MTYxMTgw"} +{"index":903,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1701.06538","reference_contexts":"If the layer has a sparsely activated mixture-of-expert (MoE) architecture [34] , it can additionally have an expert E dimension in the Einsum equation on both operands and the output, i.e., 𝐸𝐵𝐷, 𝐸𝐷𝐹 → 𝐸𝐵𝐹","reference_id":"NzA3MzYxODE0OTc2OTU2Mjcy"} +{"index":904,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1804.04235","reference_contexts":"We use 32-bit floating-point parameters, 16-bit floating-point activations, and Adafactor optimizer [35]","reference_id":"Mjg5MjgyOTE4OTg3MzQyODc2Ng=="} +{"index":905,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1909.08053","reference_contexts":"Parallelism may be introduced in several ways: data parallelism [22] partitions training data, pipeline parallelism [18, 25, 27] partitions the computation graph, and within-layer model parallelism [36] partitions the weight and computation of each model layer.","reference_id":"MTQyODQyMzE4MDQ1NjU3ODE5MTk="} +{"index":906,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1909.08053","reference_contexts":"Within-layer model parallelism partitions the weight tensor of a layer across multiple devices [36]","reference_id":"NTI3NTIyMzA5NTk3Mzg4NjEwNA=="} +{"index":907,"citer_arxiv_id":"2105.04663","cited_arxiv_id":"1909.08053","reference_contexts":"This sharding configuration combines the benefit of data parallelism, weight-update sharding [30, 40] and in-layer model parallelism [36] , and only requires 7 annotations in the model","reference_id":"OTQ2ODM1NzAxMTcwMTk3ODA5MQ=="} +{"index":908,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1511.06349","reference_contexts":"This idea have previously been considered in [15, Section 6.2] and more recently in [1] .","reference_id":"MTU0MTM3MTk4NTk1ODI0NTE3MDA="} +{"index":909,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1511.06349","reference_contexts":"), a phenomenon also recognized in the VAE setting [2, 1]","reference_id":"MTY2ODU1OTY4MjI0MzU3MjQ3MzU="} +{"index":910,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"), a phenomenon also recognized in the VAE setting [2, 1]","reference_id":"MTY2ODU1OTY4MjI0MzU3MjQ3MzU="} +{"index":911,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"The best log-likelihood results obtained here, -102.11, is higher than the best results from [2] at -103.38, which were obtained using more latent variables (100-50 vs 64-32-16-8-4) and further using 50 importance weighted samples for training.","reference_id":"MTE3MjQ2NDMwMDMwNzQwMjk4NTg="} +{"index":912,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"Using these VAE models as competitive baselines we show that LVAE improves the generative performance achieving as good or better performance than other (often complicated) methods for creating flexible variational distributions such as: The Variational Gaussian Processes [20] , Normalizing Flows [17] , Importance Weighted Autoencoders [2] or Auxiliary Deep Generative Models [12]","reference_id":"MTE2MjUwNDAyMDQ2NjU3Njk1MDU="} +{"index":913,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"A strictly tighter bound on the likelihood may be obtained at the expense of a K-fold increase of samples by using the importance weighted bound [2] :","reference_id":"NzkxNTM3MDIyNzcwMzM1MjcxMg=="} +{"index":914,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"Consider the L test 1 , Figure 3 b ), the VAE without batch-normalization and warm-up does not improve for additional stochastic layers beyond one whereas VAEs with batch normalization and warm-up 5 github.com/casperkaae/parmesan 6 github.com/casperkaae/LVAE 7 The OMNIGLOT data was partitioned and preprocessed as in [2] , https://github.com/yburda/iwae/tree/master/datasets/OMNIGLOT 8 The NORB dataset was downloaded in resized format from github.com/gwtaylor/convnet_matlab ≤ log p((x)) VAE 1-layer + NF [17] -85.10 IWAE, 2-layer + IW=1 [2] -85.33 IWAE, 2-layer + IW=50 [2] -82.90 VAE, 2-layer + VGP [20] -81.90 LVAE, 5-layer -82.12 LVAE, 5-layer + finetuning -81.84 LVAE, 5-layer + finetuning + IW=10 -81.74","reference_id":"OTAyNDI1NjQ5NTUwMzQ2MjQzMQ=="} +{"index":915,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"However, as previously identified [15, 2] , the latent representation is often overly sparse with few stochastic latent variables propagating useful information.","reference_id":"MzA5NDI1MTIzODc1NjQ4NTg5OA=="} +{"index":916,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"VAE inference models are parameterized as a bottom-up process similar to [2, 8]","reference_id":"MTgxNzYwNTQ1NDM5Mjc3Njc1MTE="} +{"index":917,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1509.00519","reference_contexts":"Similarly to [2] , we resample the binarized training values from the real-valued images using a Bernoulli distribution after each epoch which prevents the models from over-fitting","reference_id":"NDcxNTAyMzUyOTQyOTM0OTUy"} +{"index":918,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1502.04623","reference_contexts":"We believe this explicit parameter sharing between the inference and generative distribution can generally be beneficial in other types of recursive variational distributions such as DRAW [5] where the ideas presented here are directly applicable","reference_id":"MTQ4MTY2ODcwNjI5MTQwODc4MjQ="} +{"index":919,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1502.03167","reference_contexts":"We initially show that a warm-up period [1, 15, Section 6.2] to support stochastic units staying active in early training and batch normalization (BN) [6] can significantly improve performance of VAEs","reference_id":"NzI2ODYyMjA4NjU5ODI3ODQz"} +{"index":920,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1412.6980","reference_contexts":"The models were trained end-to-end using the Adam [7] optimizer with a mini-batch size of 256","reference_id":"MTQ4NDE4ODQ2MDEyMDU2MjY5MDA="} +{"index":921,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1312.6114","reference_contexts":"The recently introduced variational autoencoder (VAE) [9, 18] provides a framework for deep generative models","reference_id":"NzU2NDA2MDkyOTQ4OTEwOTUwMA=="} +{"index":922,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1312.6114","reference_contexts":"( 5 ) using stochastic gradient descent where we use the reparametrization trick for stochastic backpropagation through the Gaussian latent variables [9, 18]","reference_id":"MTM1NTQxNzE0MDE5MDY1MjYwODU="} +{"index":923,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1312.6114","reference_contexts":"Recently used inference are calculated purely bottom-up with no interaction between the inference and generative models [9, 17, 18]","reference_id":"MjgzOTkwMjEyODQ1NDYwNTI3Nw=="} +{"index":924,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1505.05770","reference_contexts":"Using these VAE models as competitive baselines we show that LVAE improves the generative performance achieving as good or better performance than other (often complicated) methods for creating flexible variational distributions such as: The Variational Gaussian Processes [20] , Normalizing Flows [17] , Importance Weighted Autoencoders [2] or Auxiliary Deep Generative Models [12]","reference_id":"MTE2MjUwNDAyMDQ2NjU3Njk1MDU="} +{"index":925,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1505.05770","reference_contexts":"Consider the L test 1 , Figure 3 b ), the VAE without batch-normalization and warm-up does not improve for additional stochastic layers beyond one whereas VAEs with batch normalization and warm-up 5 github.com/casperkaae/parmesan 6 github.com/casperkaae/LVAE 7 The OMNIGLOT data was partitioned and preprocessed as in [2] , https://github.com/yburda/iwae/tree/master/datasets/OMNIGLOT 8 The NORB dataset was downloaded in resized format from github.com/gwtaylor/convnet_matlab ≤ log p((x)) VAE 1-layer + NF [17] -85.10 IWAE, 2-layer + IW=1 [2] -85.33 IWAE, 2-layer + IW=50 [2] -82.90 VAE, 2-layer + VGP [20] -81.90 LVAE, 5-layer -82.12 LVAE, 5-layer + finetuning -81.84 LVAE, 5-layer + finetuning + IW=10 -81.74","reference_id":"OTAyNDI1NjQ5NTUwMzQ2MjQzMQ=="} +{"index":926,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1505.05770","reference_contexts":"Further the LVAE is orthogonal to other methods for improving the inference distribution such as Normalizing flows [17] , Variational Gaussian Process [20] or Auxiliary Deep generative models [12] and combining with these might provide further improvements.","reference_id":"NTY2OTc5ODg4NzI4MDMzMjIzMg=="} +{"index":927,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1505.05770","reference_contexts":"Recently used inference are calculated purely bottom-up with no interaction between the inference and generative models [9, 17, 18]","reference_id":"MjgzOTkwMjEyODQ1NDYwNTI3Nw=="} +{"index":928,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1401.4082","reference_contexts":"The recently introduced variational autoencoder (VAE) [9, 18] provides a framework for deep generative models","reference_id":"NzU2NDA2MDkyOTQ4OTEwOTUwMA=="} +{"index":929,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1401.4082","reference_contexts":"( 5 ) using stochastic gradient descent where we use the reparametrization trick for stochastic backpropagation through the Gaussian latent variables [9, 18]","reference_id":"MTM1NTQxNzE0MDE5MDY1MjYwODU="} +{"index":930,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1401.4082","reference_contexts":"Recently used inference are calculated purely bottom-up with no interaction between the inference and generative models [9, 17, 18]","reference_id":"MjgzOTkwMjEyODQ1NDYwNTI3Nw=="} +{"index":931,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1605.02688","reference_contexts":"The models were implemented using the Theano [19] , Lasagne [4] and Parmesan 5 frameworks","reference_id":"MTU1NTIxNjk3NDE2OTEwNjY2MjU="} +{"index":932,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1511.06499","reference_contexts":"Using these VAE models as competitive baselines we show that LVAE improves the generative performance achieving as good or better performance than other (often complicated) methods for creating flexible variational distributions such as: The Variational Gaussian Processes [20] , Normalizing Flows [17] , Importance Weighted Autoencoders [2] or Auxiliary Deep Generative Models [12]","reference_id":"MTE2MjUwNDAyMDQ2NjU3Njk1MDU="} +{"index":933,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1511.06499","reference_contexts":"Further the LVAE is orthogonal to other methods for improving the inference distribution such as Normalizing flows [17] , Variational Gaussian Process [20] or Auxiliary Deep generative models [12] and combining with these might provide further improvements.","reference_id":"NTY2OTc5ODg4NzI4MDMzMjIzMg=="} +{"index":934,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1511.06499","reference_contexts":"Consider the L test 1 , Figure 3 b ), the VAE without batch-normalization and warm-up does not improve for additional stochastic layers beyond one whereas VAEs with batch normalization and warm-up 5 github.com/casperkaae/parmesan 6 github.com/casperkaae/LVAE 7 The OMNIGLOT data was partitioned and preprocessed as in [2] , https://github.com/yburda/iwae/tree/master/datasets/OMNIGLOT 8 The NORB dataset was downloaded in resized format from github.com/gwtaylor/convnet_matlab ≤ log p((x)) VAE 1-layer + NF [17] -85.10 IWAE, 2-layer + IW=1 [2] -85.33 IWAE, 2-layer + IW=50 [2] -82.90 VAE, 2-layer + VGP [20] -81.90 LVAE, 5-layer -82.12 LVAE, 5-layer + finetuning -81.84 LVAE, 5-layer + finetuning + IW=10 -81.74","reference_id":"OTAyNDI1NjQ5NTUwMzQ2MjQzMQ=="} +{"index":935,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1411.7783","reference_contexts":"parameterization allows interactions between the bottom-up and top-down signals resembling the recently proposed Ladder Network [21, 16] , and we therefore denote it Ladder-VAE (LVAE)","reference_id":"MzI2MDI2MjA1NDczMDQ2ODcwNA=="} +{"index":936,"citer_arxiv_id":"1602.02282","cited_arxiv_id":"1601.06759","reference_contexts":"We found the Gaussian observation models to be harder to optimize compared to the Bernoulli models, a finding also recognized in [23] , which might explain the lower utilization of the topmost latent layers in these models.","reference_id":"MzI1MTg2NjI4MDE2MDUxNzAwNA=="} +{"index":937,"citer_arxiv_id":"1808.01340","cited_arxiv_id":"1708.03805","reference_contexts":"There was an improvement between the winning systems of the two challenges, with error going down from 12.4% (in 2017) to 11.0% (in 2018) [1, 4] .","reference_id":"MTU2MzE4MzY5MDU5MDcwOTM2MTE="} +{"index":938,"citer_arxiv_id":"1808.01340","cited_arxiv_id":"1705.06950","reference_contexts":"Full details can be found in the original publication [6] .","reference_id":"MzIwMjY3MzkwODc5NTc0NDY4OQ=="} +{"index":939,"citer_arxiv_id":"1808.01340","cited_arxiv_id":"1705.06950","reference_contexts":"The release of the Kinetics dataset [6] in 2017 led to marked improvements in state-of-the-art performance on a variety of action recognition datasets: UCF-101 [9] , HMDB-51 [7] , Charades [8] , AVA [3] , Thumos [5] , among others","reference_id":"MTU3NDMzODkzMDM2NTIxMjMxNjA="} +{"index":940,"citer_arxiv_id":"1808.01340","cited_arxiv_id":"1212.0402","reference_contexts":"The release of the Kinetics dataset [6] in 2017 led to marked improvements in state-of-the-art performance on a variety of action recognition datasets: UCF-101 [9] , HMDB-51 [7] , Charades [8] , AVA [3] , Thumos [5] , among others","reference_id":"MTU3NDMzODkzMDM2NTIxMjMxNjA="} +{"index":941,"citer_arxiv_id":"1502.04390","cited_arxiv_id":"1206.6464","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":942,"citer_arxiv_id":"1502.04390","cited_arxiv_id":"1212.5701","reference_contexts":"RMSProp (Tieleman & Hinton, 2012) and AdaDelta (Zeiler, 2012) were follow-up methods introduced to be practical adaptive learning methods to train large neural networks","reference_id":"MTI4NTI1NDI3Mjg1ODk0NTY0ODA="} +{"index":943,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1612.00005","reference_contexts":"So far, plenty of works have shown that GANs can play a significant role in various tasks, such as image generation [8] , image super-resolution [9] , and semi-supervised learning [10] .","reference_id":"MTM4MDQ2NjY0MjE0MzUxNjIwNjY="} +{"index":944,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1609.04802","reference_contexts":"So far, plenty of works have shown that GANs can play a significant role in various tasks, such as image generation [8] , image super-resolution [9] , and semi-supervised learning [10] .","reference_id":"MTM4MDQ2NjY0MjE0MzUxNjIwNjY="} +{"index":945,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1609.04802","reference_contexts":"Showing the powerful capability for unsupervised tasks, GANs have been applied to many specific tasks, like image generation [19] , image super-resolution [9] , text to image synthesis [20] and image to image translation [21]","reference_id":"NzA4MDcyMzA3NTY1ODM0NzMwOA=="} +{"index":946,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.07875","reference_contexts":"Many works have been proposed to address this problem by analyzing the objective functions of GANs [12, 24, 13, 25, 14]","reference_id":"NzU1NzYwNDI4MzA2MTc2NjE3"} +{"index":947,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.07875","reference_contexts":"One is to follow the comparison method in [12]","reference_id":"NDIzMjE2NjE3NDAyOTkwODQ1Nw=="} +{"index":948,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.07875","reference_contexts":"[12] have proposed a method to compare the stability of GANs learning by excluding batch normalization [15]","reference_id":"MTYzNTQxNDc0Mjk5Mjk5NDc5MDQ="} +{"index":949,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.07875","reference_contexts":"[12] analyzed the properties of four different divergences or distances over two distributions and concluded that Wasserstein distance is nicer than Jensen-Shannon divergence","reference_id":"MTM3ODg2ODU0Njc5NjU0MTg1MTA="} +{"index":950,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.07875","reference_contexts":"Recently, several papers have pointed out that the instability of GANs learning is partially caused by the objective function [12, 13, 14]","reference_id":"MTMzMDE1NDI1MjAyNzc4Nzg5NjY="} +{"index":951,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1611.02163","reference_contexts":"[13] created a unrolled objective function to enhance the generator","reference_id":"NDg2MjE2MDg0NDEyMTU1ODE2MA=="} +{"index":952,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1611.02163","reference_contexts":"Another experiment is to evaluate on a Gaussian mixture distribution dataset, which is designed in literature [13]","reference_id":"NTI2ODEwODk3NTU0NzM4NjQzNg=="} +{"index":953,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1611.02163","reference_contexts":"Recently, several papers have pointed out that the instability of GANs learning is partially caused by the objective function [12, 13, 14]","reference_id":"MTMzMDE1NDI1MjAyNzc4Nzg5NjY="} +{"index":954,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1611.02163","reference_contexts":"Many works have been proposed to address this problem by analyzing the objective functions of GANs [12, 24, 13, 25, 14]","reference_id":"NzU1NzYwNDI4MzA2MTc2NjE3"} +{"index":955,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.06264","reference_contexts":"Recently, several papers have pointed out that the instability of GANs learning is partially caused by the objective function [12, 13, 14]","reference_id":"MTMzMDE1NDI1MjAyNzc4Nzg5NjY="} +{"index":956,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.06264","reference_contexts":"Qi [14] proposed the Loss-Sensitive GAN whose loss function is based on the assumption that real samples should have smaller losses than fake samples and proved that this loss function has non-vanishing gradient almost everywhere.","reference_id":"MTM0MTY1MzU1MjM4MjY3MDY3Mjg="} +{"index":957,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1701.06264","reference_contexts":"Many works have been proposed to address this problem by analyzing the objective functions of GANs [12, 24, 13, 25, 14]","reference_id":"NzU1NzYwNDI4MzA2MTc2NjE3"} +{"index":958,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1611.07004","reference_contexts":"Showing the powerful capability for unsupervised tasks, GANs have been applied to many specific tasks, like image generation [19] , image super-resolution [9] , text to image synthesis [20] and image to image translation [21]","reference_id":"NzA4MDcyMzA3NTY1ODM0NzMwOA=="} +{"index":959,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1411.1784","reference_contexts":"One way to solve this problem is to use the conditional GANs [22] because conditioning on the label information creates the deterministic relationship between input and output","reference_id":"OTgwNzUwNDEzNjMxNTcwNjA4Ng=="} +{"index":960,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1411.1784","reference_contexts":"[20] proposed a model to synthesize images given text descriptions based on the conditional GANs [22]","reference_id":"NTc4Nzg2MDg3MjgyODA0MTc1Mw=="} +{"index":961,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1612.02136","reference_contexts":"[24] incorporated a reconstruction module and use the distance between real samples and reconstructed samples as a regularizer to get more stable gradients","reference_id":"MTE4NjkyMDcxMzI0Mzc3NDg2NTE="} +{"index":962,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1612.02136","reference_contexts":"Many works have been proposed to address this problem by analyzing the objective functions of GANs [12, 24, 13, 25, 14]","reference_id":"NzU1NzYwNDI4MzA2MTc2NjE3"} +{"index":963,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1606.00709","reference_contexts":"[25] pointed out that the objective of the original GAN [7] which is related to Jensen-Shannon divergence is a special case of divergence estimation, and generalized it to arbitrary f-divergences [27]","reference_id":"MTY2NDAzMDA5NDcyMDQwODYyMTQ="} +{"index":964,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1606.00709","reference_contexts":"Many works have been proposed to address this problem by analyzing the objective functions of GANs [12, 24, 13, 25, 14]","reference_id":"NzU1NzYwNDI4MzA2MTc2NjE3"} +{"index":965,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1609.03126","reference_contexts":"Viewing the discriminator as an energy function, [26] used an auto-encoder architecture to improve the stability of GANs learning","reference_id":"ODM0NzQ2OTA3MDE1ODg2MTA4MQ=="} +{"index":966,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1609.03126","reference_contexts":"(c) Generated by EBGANs (Reported in [26] ).","reference_id":"NzcwMzUwMzgwNjk3ODYyNTMwMg=="} +{"index":967,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1610.03483","reference_contexts":"Another generalization work is presented in literature [28]","reference_id":"MTcwMTA3MDk2NDczNDMwMzYwMjA="} +{"index":968,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1506.03365","reference_contexts":"We evaluate LSGANs on two datasets, LSUN [31] and HWDB1.0 [32]","reference_id":"NzEyNzg5MjUyMDgwNTMzNzMyNQ=="} +{"index":969,"citer_arxiv_id":"1611.04076","cited_arxiv_id":"1412.6980","reference_contexts":"Thus we evaluate the two architectures with two optimizers, Adam [34] and RMSProp [35]","reference_id":"NTA3NDgwNjI4MzU1NTg3ODI0Ng=="} +{"index":970,"citer_arxiv_id":"1402.0030","cited_arxiv_id":"1310.8499","reference_contexts":"Interestingly, single-layer fDARN (Gregor et al., 2013 ) models, which have autoregressive connections between the latent variables, perform better than any of the SBN models trained using the same algorithm","reference_id":"MTg0MDc3NzY0NzU2OTM1MjUyMzE="} +{"index":971,"citer_arxiv_id":"1402.0030","cited_arxiv_id":"1310.8499","reference_contexts":"The NVIL-trained fDARN models with 200 and 500 latent variables also outperform the fDARN (as well as the more expressive DARN) model with 400 latent variables from (Gregor et al., 2013) , which were trained using an MDL-based algorithm","reference_id":"MTE3NDcyMzMxMjk3NjY1Njg1NDg="} +{"index":972,"citer_arxiv_id":"1402.0030","cited_arxiv_id":"1312.6114","reference_contexts":"In contrast to other work on scaling up variational inference, NVIL can handle both discrete and continuous latent variables (unlike Kingma & Welling (2013) ; Rezende et al","reference_id":"NjIyOTAyMjM2NzE1MTAyNjMzNg=="} +{"index":973,"citer_arxiv_id":"1402.0030","cited_arxiv_id":"1312.6114","reference_contexts":"Recently a method for training nonlinear models with continuous latent variables, called Stochastic Gradient Variational Bayes (SGVB), has been proposed by Kingma & Welling (2013) and Rezende et al","reference_id":"MTUzMzU4MTYwNDgzNDMyNTEzOTI="} +{"index":974,"citer_arxiv_id":"1402.0030","cited_arxiv_id":"1312.6114","reference_contexts":"In fact, it is widely believed that learning variational parameters using gradient estimators of the form ( 6 ) is infeasible (Hinton & Zemel, 1994; Dayan & Hinton, 1996; Kingma & Welling, 2013)","reference_id":"MTc4MzEwMjkxMjQwNDU0NDQ4MDU="} +{"index":975,"citer_arxiv_id":"1402.0030","cited_arxiv_id":"1401.0118","reference_contexts":"Like NVIL, Black Box Variational Inference (BBVI, Ranganath et al., 2013) learns the variational parameters of the posterior by optimizing the variational bound using sampling-based gradient estimates, which makes it applicable to a large range of models","reference_id":"MTA5ODMxMDE3OTkyMjg5MTcwNjg="} +{"index":976,"citer_arxiv_id":"1402.0030","cited_arxiv_id":"1401.4082","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":977,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1909.11646","reference_contexts":"These models achieve state of the art results for speech generation, and show superior results over well-established methods, such as Wavernn [11] , Wavenet [22] , and GAN-TTS [1] .","reference_id":"MTMzMDQyOTE5Njg1MTg3OTkxNTA="} +{"index":978,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1802.04208","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":979,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2012.08125","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":980,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1406.2661","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":981,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1706.08500","reference_contexts":"The Fréchet Inception Distance (FID) [6] is used as the benchmark metric","reference_id":"MTg0Mjc0MzIwMzY3MzgyMDg5NTY="} +{"index":982,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2006.11239","reference_contexts":"We test our models with the inference procedure from DDPM [7] and DDIM [26]","reference_id":"MTcxOTYxNDc5MzkyOTgwMjU0NTA="} +{"index":983,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2006.11239","reference_contexts":"Denoising Diffusion Probabilistic Models (DDPM) [7] combine generative models based on score matching and neural Diffusion Probabilistic Models into a single model","reference_id":"NDA2NDY2MzIwMDQxNjg5OTE4Ng=="} +{"index":984,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2006.11239","reference_contexts":"Models from this last family have shown significant progress in generation capabilities during the last years, e.g., [25, 7, 2, 15] , and have achieved comparable results to the state of the art generation architecture in both images and speech.","reference_id":"NjAzMTM5ODQ4NDEyNzcwNTA1Nw=="} +{"index":985,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2006.11239","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":986,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2006.11239","reference_contexts":"The diffusion process in DDPM [7] is defined by a Markov chain that gradually adds Gaussian noise to the data according to a noise schedule","reference_id":"NDAzNjEzNjMzNjI3NTc3ODE1MA=="} +{"index":987,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2006.11239","reference_contexts":"learning rate, batch size etc) we use the same parameters that appear in DDPM [7]","reference_id":"OTY5NDU5MDI0NzI1MDgyNTQ3"} +{"index":988,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1807.03039","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":989,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1312.6114","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":990,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2009.09761","reference_contexts":"WaveGrad [2] and DiffWave [16] are conditioned on the mel-spectrogram and produce high fidelity audio samples using as few as six step of the diffusion process","reference_id":"NTU2NjM0NTgwMzY0MTk2NzcwMQ=="} +{"index":991,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2009.09761","reference_contexts":"Similarly, in [2, 16] a generative neural diffusion process based on score matching was applied to speech generation","reference_id":"ODA5NTExMzQ2NjE3MzU1MDMzMg=="} +{"index":992,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1609.03499","reference_contexts":"These models achieve state of the art results for speech generation, and show superior results over well-established methods, such as Wavernn [11] , Wavenet [22] , and GAN-TTS [1] .","reference_id":"MTMzMDQyOTE5Njg1MTg3OTkxNTA="} +{"index":993,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1609.03499","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":994,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1906.00446","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":995,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1907.05600","reference_contexts":"[27] proposed a diffusion generative model based on Langevin dynamics and the score matching method [8]","reference_id":"MTk4NTYxMzk5NTQ4NTMwODE0OA=="} +{"index":996,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2011.13456","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":997,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"2007.03898","reference_contexts":"The main architectures for generation are: (i) VAE [14] based, for example, NVAE [30] and VQ-VAE [23] , (ii) GAN [5] based, for example, StyleGAN [12] for vision application and WaveGAN [3] for speech (iii) Flow-based, for example Glow [13] (iv) Autoregessive, for example, Wavenet for speech [22] and (v) Diffusion Probabilistic Models [25] , for example, Denoising Diffusion Probabilistic Models (DDPM) [7] and its implicit version DDIM [26]","reference_id":"MTgxOTkwOTEwMzk5MjUzMTcwMjc="} +{"index":998,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1506.03365","reference_contexts":"We trained our model on two image datasets (i) CelebA 64x64 [20] and (ii) LSUN Church 256x256 [32]","reference_id":"MTI2MTA2NjQ2MDkwMjM0MDc5ODA="} +{"index":999,"citer_arxiv_id":"2106.07582","cited_arxiv_id":"1506.03365","reference_contexts":"3 we provide the FID score for the LSUN church dataset [32] (under unknown license)","reference_id":"NTg3OTA4NjE2MTk2MTkyMTQ3MQ=="} +{"index":1000,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2001.09977","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1001,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2108.07258","reference_contexts":"However, such opportunities also serve to highlight the importance of the many ethical challenges (Leidner & Plachouras, 2017; Bender et al., 2021; Bommasani et al., 2021) including representation bias (Blodgett et al., 2020) , proper selection and handling of training data (Rogers, 2021) and its documentation (Bender & Friedman, 2018) , privacy (Abadi et al., 2016b; Carlini et al., 2020) , and environmental concerns (Strubell et al., 2019; Patterson et al., 2021)","reference_id":"MTEyNDQ0MDAwNTQ0ODc5OTgxMTM="} +{"index":1002,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2012.07805","reference_contexts":"However, such opportunities also serve to highlight the importance of the many ethical challenges (Leidner & Plachouras, 2017; Bender et al., 2021; Bommasani et al., 2021) including representation bias (Blodgett et al., 2020) , proper selection and handling of training data (Rogers, 2021) and its documentation (Bender & Friedman, 2018) , privacy (Abadi et al., 2016b; Carlini et al., 2020) , and environmental concerns (Strubell et al., 2019; Patterson et al., 2021)","reference_id":"MTEyNDQ0MDAwNTQ0ODc5OTgxMTM="} +{"index":1003,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2003.10555","reference_contexts":"More recently, models that used Transformers (Vaswani et al., 2017) showed that larger models with self-supervision on unlabeled data could yield significant improvements on NLP tasks (Devlin et al., 2019; Yang et al., 2019; Liu et al., 2019; Clark et al., 2020)","reference_id":"MTU1OTE4NDI1Mjg1Mjc4Mjc2NDQ="} +{"index":1004,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1803.05457","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1005,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2101.03961","reference_contexts":"We leverage sparsely activated Mixture-of-Experts (MoE) (Shazeer et al., 2017; Fedus et al., 2021) in GLaM models","reference_id":"MTcwMDQ2ODc4NTAzNzM1NzczMjU="} +{"index":1006,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2101.03961","reference_contexts":"A sample of related models (Devlin et al., 2019; Raffel et al., 2020; Brown et al., 2020; Lieber et al., 2021; Rae et al., 2021; Shoeybi et al., 2019; Lepikhin et al., 2021; Fedus et al., 2021) pre-trained on text corpora","reference_id":"ODcyODU5MDk2MjUzMDU2NzIyNw=="} +{"index":1007,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2101.03961","reference_contexts":"The largest version of GLaM has 1.2T parameters in total with 64 experts per MoE layer (Shazeer et al., 2017; Lepikhin et al., 2021; Fedus et al., 2021) where each token in the input batch only activates a subnetwork of 96.6B (8% of 1.2T) parameters","reference_id":"MTY1MDE2MzU2NTc0MjkyOTM0NTM="} +{"index":1008,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2101.03961","reference_contexts":"As observed in previous work on sparsely-activated models (Fedus et al., 2021) , MoE models are more performant in knowledge-oriented tasks","reference_id":"MjQzNjM3MjU4Mjk2NTEyNDE0MA=="} +{"index":1009,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1606.08415","reference_contexts":"In the non-MoE Transformer feed-forward sub-layers, we replace the first linear projection and the activation function with the Gated Linear Unit (Dauphin et al., 2017; Shazeer, 2020) , which computes the component-wise product of two linear transformation of the input, followed by a Gaussian Error Linear Unit (Hendrycks & Gimpel, 2016) activation function","reference_id":"MTQzMzY4NTg3NDI4MzgwNzk1OTU="} +{"index":1010,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1712.00409","reference_contexts":"There has also been work on scaling sparsely activated MoE architectures (Hestness et al., 2017; Shazeer et al., 2018; Lepikhin et al., 2021; Kudugunta et al., 2021)","reference_id":"MzYzNzY4MzU0OTg0Mjg1MjYzNA=="} +{"index":1011,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2001.08361","reference_contexts":"The shift towards scaling with more data and larger models (Shazeer et al., 2017; Huang et al., 2019; Kaplan et al., 2020) has enabled complex natural language tasks to be performed with less labeled data","reference_id":"NTc3NzIzNTAwMTI2OTQxNzc5OA=="} +{"index":1012,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2009.06354","reference_contexts":"Coreference resolution is a capability that many applications require to perform well, including machine translation (Stanovsky et al., 2019; Webster & Pitler, 2020) and question answering (Lamm et al., 2020)","reference_id":"NzAyMzY4MTgzMzEwMDM0OTkwMg=="} +{"index":1013,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1907.11692","reference_contexts":"More recently, models that used Transformers (Vaswani et al., 2017) showed that larger models with self-supervision on unlabeled data could yield significant improvements on NLP tasks (Devlin et al., 2019; Yang et al., 2019; Liu et al., 2019; Clark et al., 2020)","reference_id":"MTU1OTE4NDI1Mjg1Mjc4Mjc2NDQ="} +{"index":1014,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1301.3781","reference_contexts":"Word embedding models and extensions such as word2vec (Mikolov et al., 2013) , GloVe (Pennington et al., 2014) and paragraph vectors (Le & Mikolov, 2014) have shown good generalization to many tasks simply by transferring the embeddings.","reference_id":"MTUyNzE3MjQ1MTQxNjI0ODgyODk="} +{"index":1015,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1301.3781","reference_contexts":"Variants of language models have been used to produce pretrained word vectors (Mikolov et al., 2013; Pennington et al., 2014) , and contextualized word vectors (Peters et al., 2018; Devlin et al., 2019) for many NLP applications","reference_id":"MTY0MTUwMzYxNzgxMzU0NjY1OTY="} +{"index":1016,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2104.10350","reference_contexts":"The datacenter PUE was 1.10 at the time of training GPT-3 (Patterson et al., 2021)","reference_id":"MTY2OTQ5NTI5NzQ5NzE5NzM3Mw=="} +{"index":1017,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2104.10350","reference_contexts":"The net tCO 2 e per MWh of the datacenter at the time was 0.088, training GLaM with 280B tokens emits a total of 18.7 net tCO 2 e, compared to 552 net tCO2e for GPT-3 (Patterson et al., 2021)","reference_id":"MzYzODcwODE5ODIxNDQzMzg3Ng=="} +{"index":1018,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2104.10350","reference_contexts":"While being effective and performant, scaling further is becoming prohibitively expensive and consumes significant amounts of energy (Patterson et al., 2021) .","reference_id":"MTIyMDM5Mjk0NjE5MTgwNzQ2OTc="} +{"index":1019,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2104.10350","reference_contexts":"However, such opportunities also serve to highlight the importance of the many ethical challenges (Leidner & Plachouras, 2017; Bender et al., 2021; Bommasani et al., 2021) including representation bias (Blodgett et al., 2020) , proper selection and handling of training data (Rogers, 2021) and its documentation (Bender & Friedman, 2018) , privacy (Abadi et al., 2016b; Carlini et al., 2020) , and environmental concerns (Strubell et al., 2019; Patterson et al., 2021)","reference_id":"MTEyNDQ0MDAwNTQ0ODc5OTgxMTM="} +{"index":1020,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2104.10350","reference_contexts":"Existing large dense language models usually require tremendous amounts of computation resources for training and serving (Patterson et al., 2021)","reference_id":"MTIwMTYwOTY4OTE4MDc1ODk1Mjc="} +{"index":1021,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2104.10350","reference_contexts":"Finally, although MoE-based sparse models are not yet common in the NLP community, our work shows that sparse decoder-only language models can be more performant than the dense architectures of similar compute FLOPs for the first time within the few-shot in-context learning setting at scale, suggesting that sparsity is one of the most promising directions to achieve high-quality NLP models while saving energy costs (Patterson et al., 2021)","reference_id":"ODA5NjA1NjczMjg4OTM0OTE3"} +{"index":1022,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1802.05365","reference_contexts":"Variants of language models have been used to produce pretrained word vectors (Mikolov et al., 2013; Pennington et al., 2014) , and contextualized word vectors (Peters et al., 2018; Devlin et al., 2019) for many NLP applications","reference_id":"MTY0MTUwMzYxNzgxMzU0NjY1OTY="} +{"index":1023,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1808.09121","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1024,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2112.11446","reference_contexts":"We also analyse toxicity degeneration similarly to Gopher (Rae et al., 2021) , and extend the analysis to consider the humanbehavioral baseline.","reference_id":"MzM1NjI0MjAxNDAwMzkxODU3MQ=="} +{"index":1025,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2112.11446","reference_contexts":"To evaluate toxicity degeneration, we adapt the methodology used in (Welbl et al., 2021; Rae et al., 2021)","reference_id":"MTUyNzQ1MjgwNDQxMDkwNTQ3MDY="} +{"index":1026,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2112.11446","reference_contexts":"A sample of related models (Devlin et al., 2019; Raffel et al., 2020; Brown et al., 2020; Lieber et al., 2021; Rae et al., 2021; Shoeybi et al., 2019; Lepikhin et al., 2021; Fedus et al., 2021) pre-trained on text corpora","reference_id":"ODcyODU5MDk2MjUzMDU2NzIyNw=="} +{"index":1027,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1804.04235","reference_contexts":"Our optimizer is Adafactor (Shazeer & Stern, 2018) with first-moment decay β 1 = 0, second-moment decay β 2 = 0.99 with a 1 -t -0.8 decay schedule, update clipping threshold of 1.0, and factored second-moment estimation","reference_id":"MTUyNjA3NTY0MDQ2NjkzNTY4NTg="} +{"index":1028,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1804.04235","reference_contexts":"There has also been work on scaling sparsely activated MoE architectures (Hestness et al., 2017; Shazeer et al., 2018; Lepikhin et al., 2021; Kudugunta et al., 2021)","reference_id":"MzYzNzY4MzU0OTg0Mjg1MjYzNA=="} +{"index":1029,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1902.08295","reference_contexts":"• We skip weight updates for a batch if there are any NaNs or Inf s in the gradients (Shen et al., 2019)","reference_id":"ODcxNjEyNDExMTY3MTU4NDE0OQ=="} +{"index":1030,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1909.08053","reference_contexts":"GPT-3 (Brown et al., 2020) and related work (Shoeybi et al., 2019; Lieber et al., 2021; Wei et al., 2021) demonstrated that scaling up language models greatly improves task-agnostic, few-shot performance","reference_id":"MTE2MzU5NjYzMTc5NTI0MDk5ODA="} +{"index":1031,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1909.08053","reference_contexts":"A sample of related models (Devlin et al., 2019; Raffel et al., 2020; Brown et al., 2020; Lieber et al., 2021; Rae et al., 2021; Shoeybi et al., 2019; Lepikhin et al., 2021; Fedus et al., 2021) pre-trained on text corpora","reference_id":"ODcyODU5MDk2MjUzMDU2NzIyNw=="} +{"index":1032,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"2105.04663","reference_contexts":"We rely on GSPMD's compiler pass (Xu et al., 2021) to automatically determine the sharding properties for the rest of the tensors.","reference_id":"NTc3NzA5OTQxNTY3MTQzNzI0Nw=="} +{"index":1033,"citer_arxiv_id":"2112.06905","cited_arxiv_id":"1810.12885","reference_contexts":"On a few tasks, such as ReCoRD (Zhang et al., 2018) and COPA (Gordon et al., 2012) , the non-normalized loss can yield better results and thus is adopted","reference_id":"NzM1MDMwNjI4MDE0OTI5MTU4MA=="} +{"index":1034,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.03143","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1035,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2302.03011","reference_contexts":"decoupling, Gen-1 [8] defines the content latents as CLIP embeddings and the structure latents as the monocular depth estimates, attaining superior decoupled controllability","reference_id":"MTAyNzM2NjQ3NzcwNzgzNDY2MTE="} +{"index":1036,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2205.11495","reference_contexts":"Recent works have utilized diffusion models [53, 37] to generate authentic videos [66, 14, 21, 59]","reference_id":"MjcyNTQzOTkyMDg1OTcyNzU3Ng=="} +{"index":1037,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2210.02303","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1038,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2210.02303","reference_contexts":"The videos generated by Make-A-Video [51] 6 and Imagen Video [16] 7 were downloaded from their official webpages","reference_id":"NzUyMDAwODE1NzUwMzc5MzMxOQ=="} +{"index":1039,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2210.02303","reference_contexts":"Imagen Video [16] synthesizes high definition videos given a text prompt by designing a video generator and a video super-resolution model","reference_id":"MTAzNDQ3MjAyODQyODIzNzc3NDk="} +{"index":1040,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2207.12598","reference_contexts":"We employ DDPM [17] with T = 1, 000 steps for training and use DDIM sampler [54] in classifierfree guidance [18] with 50 steps for inference by default","reference_id":"MTIwOTEwOTE3MTgzNTExMDM5NDY="} +{"index":1041,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2204.03458","reference_contexts":"Various approaches have been proposed, such as Video Diffusion [19] , which introduces a spatiotemporal factorized 3D Unet [6] with a novel conditional sampling mechanism","reference_id":"NDE4MzE3MTM4OTg4NzAyODEyNg=="} +{"index":1042,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2205.15868","reference_contexts":"Generating realistic videos remains challenging due to the difficulty in generating videos with high fidelity and motion continuity [52, 20, 67]","reference_id":"MTYzNjMzODQ2ODM3NzgwMzY0MDY="} +{"index":1043,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2205.15868","reference_contexts":"These temporal blocks account for 552 million parameters out of the total 1,345 million parameters in our UNet, indicating that 39% parameters of the UNet parameters are dedicated to capturing temporal Models FID-vid (↓) FVD (↓) CLIPSIM (↑) NÜWA [62] 47.68 -0.2439 CogVideo (Chinese) [20] 24.78 -0.2614 CogVideo (English) [20] 23.59 1294 0.2631 MagicVideo [71] -1290 -Video LDM [3] --0.2929 Make-A-Video [51] 13.17 -0.3049 ModelScopeT2V (ours)","reference_id":"Mjk4NDMxNTA0NTYwNjk3NzkwMQ=="} +{"index":1044,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2205.15868","reference_contexts":"While video generation [52, 20, 67] continues to pose challenges","reference_id":"MTI0MDUzNjc5MzU4NzQxNzg1NDU="} +{"index":1045,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2206.07696","reference_contexts":"Recent works have utilized diffusion models [53, 37] to generate authentic videos [66, 14, 21, 59]","reference_id":"MjcyNTQzOTkyMDg1OTcyNzU3Ng=="} +{"index":1046,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2106.09685","reference_contexts":"Potential strategies include using multi-condition approaches [25] or the LoRA technique [22]","reference_id":"MTQ5NzQ1OTQ1NDQyODEzOTQ3MQ=="} +{"index":1047,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2302.09778","reference_contexts":"Composer [24] reformulates images as various compositions, thus enabling image generation from not just texts, but also sketches, masks, depthmaps and more","reference_id":"NDk0OTk5NTUyMzQwNjI0MDI4Nw=="} +{"index":1048,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2302.09778","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1049,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2302.09778","reference_contexts":"Inspired from Stable Diffusion [46] , we augment the UNet structure with a cross-attention mechanism, which is an effective approach to condition the visual content on texts [68, 26, 24, 33]","reference_id":"NzUyNDY0MDU0NDgyMTYwMzg0Mw=="} +{"index":1050,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2210.09276","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1051,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2210.09276","reference_contexts":"Building on existing image generation framework [46, 48] , Imagic [27] achieves text-based semantic image edits by leveraging intermediate text embeddings that align with the input image and the target text.","reference_id":"MTc3Mzk3NDc3NzA4OTc0NzE0NjM="} +{"index":1052,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2210.09276","reference_contexts":"In the realm of visual content generation, research efforts have been put into image generation [39, 24, 48, 44] and editing [27] , leveraging diffusion models [53] .","reference_id":"Nzg2NDk4MTgwNjYzNDA0MjkwOQ=="} +{"index":1053,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"1312.6114","reference_contexts":"The early efforts of utilizing DPM into the image synthesis task at scale has been proven effective [28, 7] , surpassing dominant generative models, e.g., generative adversarial networks [13] and variational autoencoders [29] , in terms of diversity and fidelity","reference_id":"MTQ1MjYwNTE4NDM5ODY4OTU1MTc="} +{"index":1054,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2112.03857","reference_contexts":"Inspired from Stable Diffusion [46] , we augment the UNet structure with a cross-attention mechanism, which is an effective approach to condition the visual content on texts [68, 26, 24, 33]","reference_id":"NzUyNDY0MDU0NDgyMTYwMzg0Mw=="} +{"index":1055,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"1711.05101","reference_contexts":"We train ModelScopeT2V using the AdamW optimizer [36] with a learning rate of 5 × 10 -5","reference_id":"MTY5NTc1NzI2NjM1Njk3NjMxNTY="} +{"index":1056,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2112.10741","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1057,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2204.06125","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1058,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2204.06125","reference_contexts":"DALL-E 2 [44] generated image embeddings for a diffusion decoder with CLIP text encoder","reference_id":"ODkwNTY5NTk1MzcxMTYzNzE3Mw=="} +{"index":1059,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2204.06125","reference_contexts":"In the realm of visual content generation, research efforts have been put into image generation [39, 24, 48, 44] and editing [27] , leveraging diffusion models [53] .","reference_id":"Nzg2NDk4MTgwNjYzNDA0MjkwOQ=="} +{"index":1060,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2111.02114","reference_contexts":"LAION [50]","reference_id":"NDIwNTc5NzIzNDM0Mzc3NTMzMA=="} +{"index":1061,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2111.02114","reference_contexts":"Nonetheless, the scale of such datasets is orders of magnitude smaller compared to image-text paired datasets, such as LAION [50]","reference_id":"MTU1NTAxNTg5ODY4NTMyOTkxNjE="} +{"index":1062,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.14792","reference_contexts":"CLIPSIM are evaluated based on all captions from MSR-VTT test split following [51]","reference_id":"MjU0Mjg0ODE2Mjk2NTIyMzg5NQ=="} +{"index":1063,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.14792","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1064,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.14792","reference_contexts":"When evaluating on CLIPSIM metric, we followed previous works [51, 62] and use nearly 60k sentences from the whole test split as prompts to generate videos.","reference_id":"NjIxMjI5MjQ4NjIwMjM2NzQz"} +{"index":1065,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.14792","reference_contexts":"The CLIPSIM score of our model is only marginally lower than that of Make-A-Video [51] , while they utilize additional data from HD-VILA-100M [64] for training.","reference_id":"NDgzMDE1MzI0MTM4NTQxNzEwOA=="} +{"index":1066,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.14792","reference_contexts":"These temporal blocks account for 552 million parameters out of the total 1,345 million parameters in our UNet, indicating that 39% parameters of the UNet parameters are dedicated to capturing temporal Models FID-vid (↓) FVD (↓) CLIPSIM (↑) NÜWA [62] 47.68 -0.2439 CogVideo (Chinese) [20] 24.78 -0.2614 CogVideo (English) [20] 23.59 1294 0.2631 MagicVideo [71] -1290 -Video LDM [3] --0.2929 Make-A-Video [51] 13.17 -0.3049 ModelScopeT2V (ours)","reference_id":"Mjk4NDMxNTA0NTYwNjk3NzkwMQ=="} +{"index":1067,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.14792","reference_contexts":"Make-A-Video [51] employs off-the-shelf text-to-image generation model combined with spatio-temporal factorized diffusion models to generate high-quality videos without relying on paired video-text data","reference_id":"NzM0Njk5MjIxNjQ4MjY3MTYyMg=="} +{"index":1068,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.14792","reference_contexts":"The videos generated by Make-A-Video [51] 6 and Imagen Video [16] 7 were downloaded from their official webpages","reference_id":"NzUyMDAwODE1NzUwMzc5MzMxOQ=="} +{"index":1069,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2010.02502","reference_contexts":"To solve the first obstacle, research efforts focused on improving the sampling efficiency by learning-free sampling [54, 55, 35, 69, 37, 69] and learning based sampling [60, 70, 49]","reference_id":"MTgxNTM5MzczNTk5NzMzODgwMzc="} +{"index":1070,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2010.02502","reference_contexts":"We employ DDPM [17] with T = 1, 000 steps for training and use DDIM sampler [54] in classifierfree guidance [18] with 50 steps for inference by default","reference_id":"MTIwOTEwOTE3MTgzNTExMDM5NDY="} +{"index":1071,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2011.13456","reference_contexts":"To solve the first obstacle, research efforts focused on improving the sampling efficiency by learning-free sampling [54, 55, 35, 69, 37, 69] and learning based sampling [60, 70, 49]","reference_id":"MTgxNTM5MzczNTk5NzMzODgwMzc="} +{"index":1072,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2011.13456","reference_contexts":"To address the second obstacle, methods like LDM [46] , LSGM [55] and RDM [23] resorted to manifolds with lower intrinsic dimensionality [10, 65]","reference_id":"MTc0MzM2MjM2ODM2NTgwODI1Nw=="} +{"index":1073,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"1812.01717","reference_contexts":"We compare ModelScopeT2V with several state-of-the-art models using FID-vid [15] , FVD [57] , and CLIPSIM [61] metrics","reference_id":"MTE1NTE0NjM1NzE0ODk4NDY1MDk="} +{"index":1074,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"1812.01717","reference_contexts":"We evaluate the models with three metrics (i.e., FID-vid [15] , FVD [57] , and CLIPSIM [61] ).","reference_id":"NzIxMTMwODA1MjI1Mzc1NDI2NA=="} +{"index":1075,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2306.02018","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1076,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2306.02018","reference_contexts":"Recent works have utilized diffusion models [53, 37] to generate authentic videos [66, 14, 21, 59]","reference_id":"MjcyNTQzOTkyMDg1OTcyNzU3Ng=="} +{"index":1077,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2104.14806","reference_contexts":"We compare ModelScopeT2V with several state-of-the-art models using FID-vid [15] , FVD [57] , and CLIPSIM [61] metrics","reference_id":"MTE1NTE0NjM1NzE0ODk4NDY1MDk="} +{"index":1078,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2104.14806","reference_contexts":"We evaluate the models with three metrics (i.e., FID-vid [15] , FVD [57] , and CLIPSIM [61] ).","reference_id":"NzIxMTMwODA1MjI1Mzc1NDI2NA=="} +{"index":1079,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2209.00796","reference_contexts":"To address the second obstacle, methods like LDM [46] , LSGM [55] and RDM [23] resorted to manifolds with lower intrinsic dimensionality [10, 65]","reference_id":"MTc0MzM2MjM2ODM2NTgwODI1Nw=="} +{"index":1080,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2203.09481","reference_contexts":"Recent works have utilized diffusion models [53, 37] to generate authentic videos [66, 14, 21, 59]","reference_id":"MjcyNTQzOTkyMDg1OTcyNzU3Ng=="} +{"index":1081,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2202.10571","reference_contexts":"Generating realistic videos remains challenging due to the difficulty in generating videos with high fidelity and motion continuity [52, 20, 67]","reference_id":"MTYzNjMzODQ2ODM3NzgwMzY0MDY="} +{"index":1082,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2202.10571","reference_contexts":"While video generation [52, 20, 67] continues to pose challenges","reference_id":"MTI0MDUzNjc5MzU4NzQxNzg1NDU="} +{"index":1083,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2211.11018","reference_contexts":"This encompasses textual content [40, 5, 56] , visual content [39, 24, 48, 44, 27, 71, 16, 38, 51, 59] and auditory content [4, 32, 31]","reference_id":"MTIyMTcxMjY4NDg2NDAzMDcyNTM="} +{"index":1084,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2211.11018","reference_contexts":"Instead of modeling the video distribbution in the visual space (e.g., RGB format), MagicVideo [71] designed a generator in the latent space with a distribution adapter, which did not require temporal convolution","reference_id":"NzcyMzc3OTUxNDc0ODA4NDcz"} +{"index":1085,"citer_arxiv_id":"2308.06571","cited_arxiv_id":"2211.11018","reference_contexts":"These temporal blocks account for 552 million parameters out of the total 1,345 million parameters in our UNet, indicating that 39% parameters of the UNet parameters are dedicated to capturing temporal Models FID-vid (↓) FVD (↓) CLIPSIM (↑) NÜWA [62] 47.68 -0.2439 CogVideo (Chinese) [20] 24.78 -0.2614 CogVideo (English) [20] 23.59 1294 0.2631 MagicVideo [71] -1290 -Video LDM [3] --0.2929 Make-A-Video [51] 13.17 -0.3049 ModelScopeT2V (ours)","reference_id":"Mjk4NDMxNTA0NTYwNjk3NzkwMQ=="} +{"index":1086,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2009.11278","reference_contexts":"X-LXMERT X-LXMERT [5] is an enhanced version of LXMERT [53]","reference_id":"MTY0NzY2NDM5ODAwODk0NjQxMjg="} +{"index":1087,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2009.11278","reference_contexts":"The models we evaluate include AttnGAN [60] , X-LXMERT [5] , VQ-Diffusion [16], minDALL-E [27] , and Stable Diffusion v1.1, v1.5, and v2","reference_id":"MTQwNzY5MzYyODc3MTY5NTE4OTY="} +{"index":1088,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2202.04053","reference_contexts":"Another recent text-to-image evaluation is DALL-Eval [6] , which employs object detection to determine if the objects in the texts are in the generated images","reference_id":"MTMxMzczNjQ4Nzk1ODAwOTY2NTY="} +{"index":1089,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2202.04053","reference_contexts":"SOA [19] and DALL-Eval [6] employ object detection models to determine if objects, attributes, and relations in the text input are in the generated image","reference_id":"MTE4NzI0NzIwNTYyNTk0Mjk1MzA="} +{"index":1090,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2202.04053","reference_contexts":"[6, 19, 20] first convert the images using a captioning model, and then compare the image caption with the text using metrics like CIDEr [55] and SPICE [1]","reference_id":"OTc2NTM4OTAxMTc2MzA4NzU2Ng=="} +{"index":1091,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2202.04053","reference_contexts":"We include 161 from DrawBench used in Imagen [47] (texts that are categorized as \"misspellings\" and \"rare words\" are removed); 1420 from PartiPrompt used in Parti [61] (texts in category \"abstract\" are removed); and 500 texts from PaintSkill used in DALL-Eval [6] .","reference_id":"MTYwMTExODE1Nzk1NDkyNzg4Ng=="} +{"index":1092,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2202.04053","reference_contexts":"To promote the use of our new evaluation metric, we release TIFA v1.0, a large-scale text-to-image generation benchmark containing 4K diverse text inputs, sampled from the MSCOCO captions [34] , DrawBench [47] , Par-tiPrompts [61] , and PaintSkill [6]","reference_id":"MTEzNjI5NzE2NTI1NDA4OTMyMDc="} +{"index":1093,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2210.11416","reference_contexts":"Another model we use is BLIP-2 FlanT5-XL [32] , in which a VIT [8] is connected with a frozen FlanT5 [7] via a lightweight transformer","reference_id":"NzYzNjM0ODczNTMzNTQzMDU1MQ=="} +{"index":1094,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2212.05032","reference_contexts":"these challenges: researchers are imposing linguistic structure with diffusion guidance to produce images with multiple objects [11] ; others are designing reward models trained using human feedback to better align generations with user intention [30]","reference_id":"MTgzODEzMTY2NDIyOTMxMTY3OTc="} +{"index":1095,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2212.05032","reference_contexts":"Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]","reference_id":"MTY1Mjk5MTE4NDY1MjA4NTI1ODA="} +{"index":1096,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2212.05032","reference_contexts":"models often fail to compose multiple objects [40, 11, 35] , bind attributes to the wrong objects [11] , and struggle in generating visual text [36]","reference_id":"MTYwMDQwMTE2MDMwODg1OTA3MDE="} +{"index":1097,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2211.09699","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1098,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2302.12192","reference_contexts":"these challenges: researchers are imposing linguistic structure with diffusion guidance to produce images with multiple objects [11] ; others are designing reward models trained using human feedback to better align generations with user intention [30]","reference_id":"MTgzODEzMTY2NDIyOTMxMTY3OTc="} +{"index":1099,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2302.12192","reference_contexts":"Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]","reference_id":"MTY1Mjk5MTE4NDY1MjA4NTI1ODA="} +{"index":1100,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2301.12597","reference_contexts":"Another model we use is BLIP-2 FlanT5-XL [32] , in which a VIT [8] is connected with a frozen FlanT5 [7] via a lightweight transformer","reference_id":"NzYzNjM0ODczNTMzNTQzMDU1MQ=="} +{"index":1101,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2301.12597","reference_contexts":"This benchmark also comes with different VQA models [57, 28, 58, 32, 31, 21] that can be used to evaluate generative models and can be easily extended to use future VQA models when they become available.","reference_id":"MTAyNTQyNTA4MjcwNjkzNzY5OTM="} +{"index":1102,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2301.12597","reference_contexts":"We use the state-of-the-art BLIP-2 FlanT5-XL [32] as the captioning model","reference_id":"MTIzOTcyMjE3MTU0NTYxMTExNjI="} +{"index":1103,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2301.12597","reference_contexts":"To evaluate a generated image, we use a visual-question-answering (VQA) system (here, mPLUGlarge [31] , BLIP-2 [32] , etc.) to answer the questions given the generated image","reference_id":"MTc0Nzg3MzkzNTYwNjIyNDM1MTA="} +{"index":1104,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2206.01714","reference_contexts":"Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]","reference_id":"MTY1Mjk5MTE4NDY1MjA4NTI1ODA="} +{"index":1105,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2206.01714","reference_contexts":"models often fail to compose multiple objects [40, 11, 35] , bind attributes to the wrong objects [11] , and struggle in generating visual text [36]","reference_id":"MTYwMDQwMTE2MDMwODg1OTA3MDE="} +{"index":1106,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2211.12112","reference_contexts":"Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]","reference_id":"MTY1Mjk5MTE4NDY1MjA4NTI1ODA="} +{"index":1107,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2204.06125","reference_contexts":"While we welcome artistic freedom when we commission art from artists, images produced by deep generative models [44, 46, 43, 47, 61] should conform closely to our requests","reference_id":"MTAxNDAxODAyNzM3MjAzNTcwOTE="} +{"index":1108,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2102.12092","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1109,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"1908.10084","reference_contexts":"For each question, we first decode the free-form answer and then choose the choice that has the highest similarity with the decoded answer, measured by SBERT [45]","reference_id":"MTQzNTY5MDA1MTU0NTA0MzYzOTQ="} +{"index":1110,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2205.11487","reference_contexts":"The images are generated by the five most recent text-to-image models in our study, including VQ-Diffusion [16], minDALL-E [27] , and Stable Diffusion [47] v1.1, v1.5, and v2.1","reference_id":"NzM2ODc2MTk5MzI0OTc4MzE1Mw=="} +{"index":1111,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2205.11487","reference_contexts":"While we welcome artistic freedom when we commission art from artists, images produced by deep generative models [44, 46, 43, 47, 61] should conform closely to our requests","reference_id":"MTAxNDAxODAyNzM3MjAzNTcwOTE="} +{"index":1112,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2205.11487","reference_contexts":"We include 161 from DrawBench used in Imagen [47] (texts that are categorized as \"misspellings\" and \"rare words\" are removed); 1420 from PartiPrompt used in Parti [61] (texts in category \"abstract\" are removed); and 500 texts from PaintSkill used in DALL-Eval [6] .","reference_id":"MTYwMTExODE1Nzk1NDkyNzg4Ng=="} +{"index":1113,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2205.11487","reference_contexts":"The first text input comes from COCO [34] and the second and third text inputs come from DrawBench [47] .","reference_id":"Mjc2OTM4Njk4NzIxNzU5OTk2MA=="} +{"index":1114,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2205.11487","reference_contexts":"To promote the use of our new evaluation metric, we release TIFA v1.0, a large-scale text-to-image generation benchmark containing 4K diverse text inputs, sampled from the MSCOCO captions [34] , DrawBench [47] , Par-tiPrompts [61] , and PaintSkill [6]","reference_id":"MTEzNjI5NzE2NTI1NDA4OTMyMDc="} +{"index":1115,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"1606.03498","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1116,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2210.08402","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1117,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2307.09288","reference_contexts":"To promote deterministic benchmark generation, we fine-tune and release a LLaMA 2 (7B) [54] model that parses the captions and generates questions for arbitrary texts, using TIFA v1.0 questions as training examples","reference_id":"MTkwMDY4MjUzMjE5MjQ1NjA3MQ=="} +{"index":1118,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2205.14100","reference_contexts":"Vision-language models The general pre-trained visionlanguage model are GIT-large [57] , VILT-B/32 [28] , OFAlarge [58] , and mPLUG-large [31]","reference_id":"MjA0ODM5NDcwNzc1NzgzNzgzNA=="} +{"index":1119,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2205.14100","reference_contexts":"This benchmark also comes with different VQA models [57, 28, 58, 32, 31, 21] that can be used to evaluate generative models and can be easily extended to use future VQA models when they become available.","reference_id":"MTAyNTQyNTA4MjcwNjkzNzY5OTM="} +{"index":1120,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2202.03052","reference_contexts":"Vision-language models The general pre-trained visionlanguage model are GIT-large [57] , VILT-B/32 [28] , OFAlarge [58] , and mPLUG-large [31]","reference_id":"MjA0ODM5NDcwNzc1NzgzNzgzNA=="} +{"index":1121,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2202.03052","reference_contexts":"This benchmark also comes with different VQA models [57, 28, 58, 32, 31, 21] that can be used to evaluate generative models and can be easily extended to use future VQA models when they become available.","reference_id":"MTAyNTQyNTA4MjcwNjkzNzY5OTM="} +{"index":1122,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2206.10789","reference_contexts":"We include 161 from DrawBench used in Imagen [47] (texts that are categorized as \"misspellings\" and \"rare words\" are removed); 1420 from PartiPrompt used in Parti [61] (texts in category \"abstract\" are removed); and 500 texts from PaintSkill used in DALL-Eval [6] .","reference_id":"MTYwMTExODE1Nzk1NDkyNzg4Ng=="} +{"index":1123,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2206.10789","reference_contexts":"To promote the use of our new evaluation metric, we release TIFA v1.0, a large-scale text-to-image generation benchmark containing 4K diverse text inputs, sampled from the MSCOCO captions [34] , DrawBench [47] , Par-tiPrompts [61] , and PaintSkill [6]","reference_id":"MTEzNjI5NzE2NTI1NDA4OTMyMDc="} +{"index":1124,"citer_arxiv_id":"2303.11897","cited_arxiv_id":"2206.10789","reference_contexts":"While we welcome artistic freedom when we commission art from artists, images produced by deep generative models [44, 46, 43, 47, 61] should conform closely to our requests","reference_id":"MTAxNDAxODAyNzM3MjAzNTcwOTE="} +{"index":1125,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2112.00861","reference_contexts":"Similar issue, which is akin to the alignment tax, has been observed in language domains (Askell et al., 2021; Ouyang et al., 2022) .","reference_id":"MTE5OTA0OTI3MDY5MDY3MTE2Nzg="} +{"index":1126,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"1607.07086","reference_contexts":"Human feedback has been used to improve various AI systems, from translation (Bahdanau et al., 2016; Kreutzer et al., 2018) , to web question-answering (Nakano et al., 2021) , to story generation (Zhou & Xu, 2020) , to training RL agents without the need for hand-designed rewards (MacGlashan et al., 2017; Christiano et al., 2017; Warnell et al., 2018; Ibarz et al., 2018; Lee et al., 2021, inter alia) , and to more truthful and harmless instruction following and dialogue (Ouyang et al., 2022; Bai et al., 2022a; Ziegler et al., 2019; Wu et al., 2021; Stiennon et al., 2020; Liu et al., 2023; Scheurer et al., 2022; Bai et al., 2022b, inter alia)","reference_id":"MTg5OTcxOTk4NDQyNzY2NTYyNQ=="} +{"index":1127,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2204.05862","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1128,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2212.08073","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1129,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2005.14165","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1130,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2212.05032","reference_contexts":"For instance, current text-to-image models often fail to produce reliable visual text (Liu et al., 2022b) and struggle with compositional image generation (Feng et al., 2022) .","reference_id":"MjM5Njc1Mzk1NjYzNjMzNDc1Nw=="} +{"index":1131,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2212.05032","reference_contexts":"However, major challenges remain in domains where large-scale text-to-image models fail to generate images that are well-aligned with text prompts (Feng et al., 2022; Liu et al., 2022a; b)","reference_id":"ODI3Mzk0MzEwNzc0MTc3NzkzOQ=="} +{"index":1132,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2212.05032","reference_contexts":"However, text-to-image models frequently struggle to generate images that are well-aligned with text prompts (Feng et al., 2022; Liu et al., 2022a; b)","reference_id":"MTY5NjEwNDE2NjUzOTU5MTcwODI="} +{"index":1133,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2212.05032","reference_contexts":"Several techniques, such as character-aware text encoders (Xue et al., 2022; Liu et al., 2022b) and structured representations of language inputs (Feng et al., 2022) have been investigated to address these issues","reference_id":"NDAyOTcxMDk3OTQ2Nzg4ODYwMg=="} +{"index":1134,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2208.01618","reference_contexts":"Fine-tuning with few images (Ruiz et al., 2022; Kumari et al., 2022; Gal et al., 2022) for personalization of textto-image diffusion models is also related with our work","reference_id":"MTA0NTA0OTE3OTk4MzQ2NTk4MjY="} +{"index":1135,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2104.08718","reference_contexts":"As a baseline, we compare it with the CLIP score (Hessel et al., 2021) , which measures image-text similarity in the CLIP embedding space (Radford et al., 2021) .","reference_id":"MTg0MzIzNDQ1NzUyNTczOTIwMzU="} +{"index":1136,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2104.08718","reference_contexts":"To measure the image-text alignment, various evaluation protocols have been proposed (Madhyastha et al., 2019; Hessel et al., 2021; Saharia et al., 2022; Yu et al., 2022b)","reference_id":"MTc4NjM5NzE0MzM5NjQ0NjAwNzE="} +{"index":1137,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"1312.6114","reference_contexts":"Various deep generative models, such as variational auto-encoders (Kingma & Welling, 2013) , generative adversarial networks (Goodfellow et al., 2020) , auto-regressive models (Van Den Oord et al., 2016) , and diffusion models (Sohl-Dickstein et al., 2015; Ho et al., 2020) have been proposed for image distributions","reference_id":"MTY3NDc5OTA3MTk2OTUyNzg1MzA="} +{"index":1138,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"1804.05958","reference_contexts":"Human feedback has been used to improve various AI systems, from translation (Bahdanau et al., 2016; Kreutzer et al., 2018) , to web question-answering (Nakano et al., 2021) , to story generation (Zhou & Xu, 2020) , to training RL agents without the need for hand-designed rewards (MacGlashan et al., 2017; Christiano et al., 2017; Warnell et al., 2018; Ibarz et al., 2018; Lee et al., 2021, inter alia) , and to more truthful and harmless instruction following and dialogue (Ouyang et al., 2022; Bai et al., 2022a; Ziegler et al., 2019; Wu et al., 2021; Stiennon et al., 2020; Liu et al., 2023; Scheurer et al., 2022; Bai et al., 2022b, inter alia)","reference_id":"MTg5OTcxOTk4NDQyNzY2NTYyNQ=="} +{"index":1139,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2212.04488","reference_contexts":"Fine-tuning with few images (Ruiz et al., 2022; Kumari et al., 2022; Gal et al., 2022) for personalization of textto-image diffusion models is also related with our work","reference_id":"MTA0NTA0OTE3OTk4MzQ2NTk4MjY="} +{"index":1140,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2206.01714","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1141,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2212.10562","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1142,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"1711.05101","reference_contexts":"We update the model using AdamW (Loshchilov & Hutter, 2017) with β 1 = 0.9, β 2 = 0.999, = 1e -8 and weight decay 1e -2","reference_id":"NzA0ODY5ODQzODAwOTA1NzU2Mg=="} +{"index":1143,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2112.09332","reference_contexts":"In language modeling, learning from human feedback has emerged as a powerful solution for aligning model behavior with human intent (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021; Nakano et al., 2021; Ouyang et al., 2022; Bai et al., 2022a)","reference_id":"NzY1MzAwNjc4Nzg3MTIxNDgwNA=="} +{"index":1144,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2112.09332","reference_contexts":"Human feedback has been used to improve various AI systems, from translation (Bahdanau et al., 2016; Kreutzer et al., 2018) , to web question-answering (Nakano et al., 2021) , to story generation (Zhou & Xu, 2020) , to training RL agents without the need for hand-designed rewards (MacGlashan et al., 2017; Christiano et al., 2017; Warnell et al., 2018; Ibarz et al., 2018; Lee et al., 2021, inter alia) , and to more truthful and harmless instruction following and dialogue (Ouyang et al., 2022; Bai et al., 2022a; Ziegler et al., 2019; Wu et al., 2021; Stiennon et al., 2020; Liu et al., 2023; Scheurer et al., 2022; Bai et al., 2022b, inter alia)","reference_id":"MTg5OTcxOTk4NDQyNzY2NTYyNQ=="} +{"index":1145,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"In language modeling, learning from human feedback has emerged as a powerful solution for aligning model behavior with human intent (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021; Nakano et al., 2021; Ouyang et al., 2022; Bai et al., 2022a)","reference_id":"NzY1MzAwNjc4Nzg3MTIxNDgwNA=="} +{"index":1146,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"In prior work (Ouyang et al., 2022) , model is optimized with bigger batch for pre-training data","reference_id":"MTEzMzEzMTE4ODU4Mzk2ODUzNDU="} +{"index":1147,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"Unlike the prior work (Stiennon et al., 2020; Ouyang et al., 2022) that uses RL for optimization, we update the model using semi-supervised learning to measure model-output quality w.r.t","reference_id":"ODc0Njc5ODU4ODY3OTAyNTI3MA=="} +{"index":1148,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"However, similar to prior work in language domains (Ouyang et al., 2022) , it would be an interesting direction to use RL algorithms (Schulman et al., 2017)","reference_id":"MTA3NDY2MDk4ODA4NjExMzg3OTM="} +{"index":1149,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"Human feedback has been used to improve various AI systems, from translation (Bahdanau et al., 2016; Kreutzer et al., 2018) , to web question-answering (Nakano et al., 2021) , to story generation (Zhou & Xu, 2020) , to training RL agents without the need for hand-designed rewards (MacGlashan et al., 2017; Christiano et al., 2017; Warnell et al., 2018; Ibarz et al., 2018; Lee et al., 2021, inter alia) , and to more truthful and harmless instruction following and dialogue (Ouyang et al., 2022; Bai et al., 2022a; Ziegler et al., 2019; Wu et al., 2021; Stiennon et al., 2020; Liu et al., 2023; Scheurer et al., 2022; Bai et al., 2022b, inter alia)","reference_id":"MTg5OTcxOTk4NDQyNzY2NTYyNQ=="} +{"index":1150,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"2017 ) could be considered for updating the text-to-image model similar to RLHF finetuning (Ouyang et al., 2022)","reference_id":"MjY1Mzk4MTQ1NTkzODIyMzQ5"} +{"index":1151,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"Similar issue, which is akin to the alignment tax, has been observed in language domains (Askell et al., 2021; Ouyang et al., 2022) .","reference_id":"MTE5OTA0OTI3MDY5MDY3MTE2Nzg="} +{"index":1152,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2203.02155","reference_contexts":"More informative human feedback, such as ranking (Stiennon et al., 2020; Ouyang et al., 2022) , should prove useful when more complex or subjective text prompts are used (e.g., artistic or open-ended generation).","reference_id":"MTczNTgyMTYzMjE1MDc0MTczMTg="} +{"index":1153,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2204.06125","reference_contexts":"Most prior works (Ramesh et al., 2022; Saharia et al., 2022; Yu et al., 2022b) use the alignment score of image and text embeddings determined by pre-trained multi-modal models, such as CLIP (Radford et al., 2021) and CoCa (Yu et al., 2022a)","reference_id":"ODIxMzI4OTAyMjM5MDE5MDAyMQ=="} +{"index":1154,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2204.06125","reference_contexts":"To improve the alignment of generated images with their text prompts, we fine-tune a pre-trained text-to-image model (Ramesh et al., 2022; Saharia et al., 2022; Rombach et al., 2022) by repeating the following steps shown in Figure 1","reference_id":"MTA3Nzc2NzA4NjI4ODMyMjU4NDk="} +{"index":1155,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2204.06125","reference_contexts":"We note that the stable diffusion model struggles to generate the number of objects specified by the prompt, but reliably generates 7 Our fine-tuning method can be used readily with other textto-image models, such as Imagen (Saharia et al., 2022) , Parti (Yu et al., 2022b) and Dalle-2 (Ramesh et al., 2022)","reference_id":"NzQ0MDU1OTExNTc4NzUxNTgzNA=="} +{"index":1156,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2204.06125","reference_contexts":"Combined with the large-scale language encoders (Radford et al., 2021; Raffel et al., 2020) , these models have shown impressive results in text-to-image generation (Ramesh et al., 2022; Saharia et al., 2022; Yu et al., 2022b; Rombach et al., 2022) .","reference_id":"MTExMDgzMDAzODc4NTUwOTQwODA="} +{"index":1157,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2208.12242","reference_contexts":"DreamBooth (Ruiz et al., 2022) showed that text-to-image models can generate diverse images in a personalized way by fine-tuning with few images, and Kumari et al","reference_id":"MTE4NjAyMTA2NDUwMzQ0MTgxMjI="} +{"index":1158,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2208.12242","reference_contexts":"Fine-tuning with few images (Ruiz et al., 2022; Kumari et al., 2022; Gal et al., 2022) for personalization of textto-image diffusion models is also related with our work","reference_id":"MTA0NTA0OTE3OTk4MzQ2NTk4MjY="} +{"index":1159,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2111.02114","reference_contexts":"This success has been driven in part by the scaling of deep generative models to large-scale datasets from the web such as LAION (Schuhmann et al., 2021; 2022)","reference_id":"NTcxNjY0OTYxOTU2Njg5MDQxOQ=="} +{"index":1160,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2111.02114","reference_contexts":"For our baseline generative model, we use stable diffusion v1.5 (Rombach et al., 2022) , which has been pre-trained on large image-text datasets (Schuhmann et al., 2021; 2022)","reference_id":"NTc3NTk1NDc0NzkyOTQ0ODA0Nw=="} +{"index":1161,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2111.02114","reference_contexts":"For our baseline generative model, we use stable diffusion v1.5 (Rombach et al., 2022) , which has been pre-trained on large image-text datasets (Schuhmann et al., 2021; 2022)","reference_id":"MTY5Mzk0NjI1Njc3NTI2NTA5ODA="} +{"index":1162,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2210.08402","reference_contexts":"We also use 16K unlabeled samples for the reward-weighted loss and a 625K subset foot_8 of LAION-5B (Schuhmann et al., 2022) filtered by an aesthetic score predictor foot_9 for the pre-training loss.","reference_id":"MTAwODY5ODM5MDU2MjgxMzYyMDY="} +{"index":1163,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"1707.06347","reference_contexts":"However, similar to prior work in language domains (Ouyang et al., 2022) , it would be an interesting direction to use RL algorithms (Schulman et al., 2017)","reference_id":"MTA3NDY2MDk4ODA4NjExMzg3OTM="} +{"index":1164,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2009.01325","reference_contexts":"In language modeling, learning from human feedback has emerged as a powerful solution for aligning model behavior with human intent (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021; Nakano et al., 2021; Ouyang et al., 2022; Bai et al., 2022a)","reference_id":"NzY1MzAwNjc4Nzg3MTIxNDgwNA=="} +{"index":1165,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2009.01325","reference_contexts":"Unlike the prior work (Stiennon et al., 2020; Ouyang et al., 2022) that uses RL for optimization, we update the model using semi-supervised learning to measure model-output quality w.r.t","reference_id":"ODc0Njc5ODU4ODY3OTAyNTI3MA=="} +{"index":1166,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2009.01325","reference_contexts":"Ouyang et al., 2022; Ziegler et al., 2019; Stiennon et al., 2020) Inspired by their successes, we propose a fine-tuning method with human feedback for improving text-to-image models.","reference_id":"NjcxMjM3MzE3MzUxNjE2NTk2OQ=="} +{"index":1167,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2009.01325","reference_contexts":"Human feedback has been used to improve various AI systems, from translation (Bahdanau et al., 2016; Kreutzer et al., 2018) , to web question-answering (Nakano et al., 2021) , to story generation (Zhou & Xu, 2020) , to training RL agents without the need for hand-designed rewards (MacGlashan et al., 2017; Christiano et al., 2017; Warnell et al., 2018; Ibarz et al., 2018; Lee et al., 2021, inter alia) , and to more truthful and harmless instruction following and dialogue (Ouyang et al., 2022; Bai et al., 2022a; Ziegler et al., 2019; Wu et al., 2021; Stiennon et al., 2020; Liu et al., 2023; Scheurer et al., 2022; Bai et al., 2022b, inter alia)","reference_id":"MTg5OTcxOTk4NDQyNzY2NTYyNQ=="} +{"index":1168,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2009.01325","reference_contexts":"More informative human feedback, such as ranking (Stiennon et al., 2020; Ouyang et al., 2022) , should prove useful when more complex or subjective text prompts are used (e.g., artistic or open-ended generation).","reference_id":"MTczNTgyMTYzMjE1MDc0MTczMTg="} +{"index":1169,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2109.10862","reference_contexts":"In language modeling, learning from human feedback has emerged as a powerful solution for aligning model behavior with human intent (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021; Nakano et al., 2021; Ouyang et al., 2022; Bai et al., 2022a)","reference_id":"NzY1MzAwNjc4Nzg3MTIxNDgwNA=="} +{"index":1170,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2109.10862","reference_contexts":"Human feedback has been used to improve various AI systems, from translation (Bahdanau et al., 2016; Kreutzer et al., 2018) , to web question-answering (Nakano et al., 2021) , to story generation (Zhou & Xu, 2020) , to training RL agents without the need for hand-designed rewards (MacGlashan et al., 2017; Christiano et al., 2017; Warnell et al., 2018; Ibarz et al., 2018; Lee et al., 2021, inter alia) , and to more truthful and harmless instruction following and dialogue (Ouyang et al., 2022; Bai et al., 2022a; Ziegler et al., 2019; Wu et al., 2021; Stiennon et al., 2020; Liu et al., 2023; Scheurer et al., 2022; Bai et al., 2022b, inter alia)","reference_id":"MTg5OTcxOTk4NDQyNzY2NTYyNQ=="} +{"index":1171,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2205.01917","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1172,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"2206.10789","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1173,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"1909.08593","reference_contexts":"In language modeling, learning from human feedback has emerged as a powerful solution for aligning model behavior with human intent (Ziegler et al., 2019; Stiennon et al., 2020; Wu et al., 2021; Nakano et al., 2021; Ouyang et al., 2022; Bai et al., 2022a)","reference_id":"NzY1MzAwNjc4Nzg3MTIxNDgwNA=="} +{"index":1174,"citer_arxiv_id":"2302.12192","cited_arxiv_id":"1909.08593","reference_contexts":"Human feedback has been used to improve various AI systems, from translation (Bahdanau et al., 2016; Kreutzer et al., 2018) , to web question-answering (Nakano et al., 2021) , to story generation (Zhou & Xu, 2020) , to training RL agents without the need for hand-designed rewards (MacGlashan et al., 2017; Christiano et al., 2017; Warnell et al., 2018; Ibarz et al., 2018; Lee et al., 2021, inter alia) , and to more truthful and harmless instruction following and dialogue (Ouyang et al., 2022; Bai et al., 2022a; Ziegler et al., 2019; Wu et al., 2021; Stiennon et al., 2020; Liu et al., 2023; Scheurer et al., 2022; Bai et al., 2022b, inter alia)","reference_id":"MTg5OTcxOTk4NDQyNzY2NTYyNQ=="} +{"index":1175,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1707.02286","reference_contexts":"A popular class of on-policy algorithms is policy gradient methods (Williams, 1992; Sutton et al., 2000) , which have been shown to be effective for a diverse array of complex tasks (Heess et al., 2017; Pathak et al., 2017; Peng et al., 2018; Rajeswaran et al., 2018)","reference_id":"MzY5MTAxNDAwMDAxODcxMTA="} +{"index":1176,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1707.02286","reference_contexts":"Model-free reinforcement learning can be a general and effective methodology for training agents to acquire sophisticated behaviors with minimal assumptions on the underlying task (Mnih et al., 2015; Heess et al., 2017; Pathak et al., 2017)","reference_id":"MTUwNDc3MDk4MDA3NzE2ODE5MDE="} +{"index":1177,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1710.02298","reference_contexts":"But these methods can be notoriously unstable, and in practice, require a variety of stabilization techniques to ensure more consistent performance (Hasselt et al., 2016; Wang et al., 2016; Munos et al., 2016; Hessel et al., 2017; Fujimoto et al., 2018; Nachum et al., 2018; Fu et al., 2019)","reference_id":"MTEzNzQ5NjAzNTUzMzc0Nzg0Mjg="} +{"index":1178,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1906.00949","reference_contexts":"Q-function based methods, such as TD3 and SAC, can in principle handle off-policy data but, as discussed in prior work, tend to struggle in this setting in practice (Fujimoto et al., 2019; Kumar et al., 2019)","reference_id":"MTI3OTExMTcwMTE2ODM5OTY1NzY="} +{"index":1179,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1906.00949","reference_contexts":"Unlike Q-function based methods, AWR is less susceptible to issues from out-of-distribution actions as the policy is always trained on observed actions from the behaviour data (Kumar et al., 2019)","reference_id":"Nzk3MjM1MDM5Njc1ODI4OTc4MA=="} +{"index":1180,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1906.00949","reference_contexts":"Furthermore, it can be difficult to apply these methods to learn from fully off-policy data, where an agent is unable to collect additional environmental interactions (Fujimoto et al., 2019; Kumar et al., 2019) .","reference_id":"NDAzNjczMDcwNjcyNTkyOTU0OQ=="} +{"index":1181,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1906.00949","reference_contexts":"The comparisons also include recent off-policy methods: batch-constrained Q-learning (BCQ) (Fujimoto et al., 2019) and bootstrapping error accumulation reduction (BEAR) (Kumar et al., 2019) , which have shown strong performance on off-policy learning with static datasets","reference_id":"MTY3NjE5NTM2MTQzNjc3Nzc2MDc="} +{"index":1182,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1707.06347","reference_contexts":"We compare AWR to a number of state-of-the-art RL algorithms, including on-policy algorithms, such as TRPO (Schulman et al., 2015) and PPO (Schulman et al., 2017) , off-policy algorithms, such as DDPG (Lillicrap et al., 2016) , TD3 (Fujimoto et al., 2018) , and SAC (Haarnoja et al., 2018a) , as well as RWR (Peters & Schaal, 2007) , which we include for comparison due to its similarity to AWR","reference_id":"MTM1MTM2OTI5ODE0NjU0MDU5MDA="} +{"index":1183,"citer_arxiv_id":"1910.00177","cited_arxiv_id":"1611.01224","reference_contexts":"But these methods can be notoriously unstable, and in practice, require a variety of stabilization techniques to ensure more consistent performance (Hasselt et al., 2016; Wang et al., 2016; Munos et al., 2016; Hessel et al., 2017; Fujimoto et al., 2018; Nachum et al., 2018; Fu et al., 2019)","reference_id":"MTEzNzQ5NjAzNTUzMzc0Nzg0Mjg="} +{"index":1184,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.04160","reference_contexts":"Recently, interesting works have made remarkable progress by extending LLM to audio, e.g., KOSMOS-1 (Huang et al., 2023) , X-LLM (Chen et al., 2023) , PandaGPT (Su et al., 2023) and control systems like PaLM-E (Driess et al., 2023) and EmbodiedGPT (Mu et al., 2023)","reference_id":"MTA1MDQ2NTU1NDYwODc4MjM0MTY="} +{"index":1185,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2109.10852","reference_contexts":"An interesting genre is Pix2seq (Chen et al., 2021) , which formalizes the detection task as a sequence generation task","reference_id":"MjI3MzMwODQ5MTEzMTg2MDc3Mw=="} +{"index":1186,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2109.10852","reference_contexts":"For detect object in autoregressive model, several methods (Chen et al., 2021; Wang et al., 2022b) introduce extra vocabularies (e.g., , • • • , ) to represent coordinates for object detection in spatially discretized images, as described in Section 2.3","reference_id":"MTQ5MTg1Mjc4NDcyMTYzMDI3MzI="} +{"index":1187,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"1504.00325","reference_contexts":"For Image Captioning, we evaluate them on COCO (Chen et al., 2015) and Flickr30k (Plummer et al., 2015) in CIDEr","reference_id":"MTMwNDUyMzg5MTcyODEwMTM5NjY="} +{"index":1188,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.06500","reference_contexts":"Mini-GPT4 (Zhu et al., 2023) , mPLUG-OWL (Ye et al., 2023) , VPGTrans (Zhang et al., 2023a) , and InstructBLIP (Dai et al., 2023) retain Q-Former, replace language model to a larger one, and then tuning on meticulously collected instruction data","reference_id":"MTE5OTE3NzQwODg0Mjk0OTcxNA=="} +{"index":1189,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.06500","reference_contexts":"Our method has achieved results comparable to In-strutBLIP (Dai et al., 2023) and far surpasses recent popular MLLMs","reference_id":"MTg1MzgxNjUzNjQ0NDQ5MjMxNg=="} +{"index":1190,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.06500","reference_contexts":"In recent months, Multimodal Large Language Models (MLLMs) have witness remarkable progress (Alayrac et al., 2022; Huang et al., 2023; Liu et al., 2023a; Zhu et al., 2023; Li et al., 2023a; Gao et al., 2023; Dai et al., 2023)","reference_id":"MzYxODQzNjIwNTU3NDU3MTE1OQ=="} +{"index":1191,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2303.03378","reference_contexts":"Recently, interesting works have made remarkable progress by extending LLM to audio, e.g., KOSMOS-1 (Huang et al., 2023) , X-LLM (Chen et al., 2023) , PandaGPT (Su et al., 2023) and control systems like PaLM-E (Driess et al., 2023) and EmbodiedGPT (Mu et al., 2023)","reference_id":"MTA1MDQ2NTU1NDYwODc4MjM0MTY="} +{"index":1192,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2304.15010","reference_contexts":"In recent months, Multimodal Large Language Models (MLLMs) have witness remarkable progress (Alayrac et al., 2022; Huang et al., 2023; Liu et al., 2023a; Zhu et al., 2023; Li et al., 2023a; Gao et al., 2023; Dai et al., 2023)","reference_id":"MzYxODQzNjIwNTU3NDU3MTE1OQ=="} +{"index":1193,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.04790","reference_contexts":"Subsequently, MM-GPT (Gong et al., 2023) , and Otter (Li et al., 2023a) tune on carefully constructed instruction data for a more user-friendly interaction","reference_id":"MTMyNjA4NjIxNjQ3MDIxMTQzMDI="} +{"index":1194,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2302.14045","reference_contexts":"Recently, interesting works have made remarkable progress by extending LLM to audio, e.g., KOSMOS-1 (Huang et al., 2023) , X-LLM (Chen et al., 2023) , PandaGPT (Su et al., 2023) and control systems like PaLM-E (Driess et al., 2023) and EmbodiedGPT (Mu et al., 2023)","reference_id":"MTA1MDQ2NTU1NDYwODc4MjM0MTY="} +{"index":1195,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2302.14045","reference_contexts":"In recent months, Multimodal Large Language Models (MLLMs) have witness remarkable progress (Alayrac et al., 2022; Huang et al., 2023; Liu et al., 2023a; Zhu et al., 2023; Li et al., 2023a; Gao et al., 2023; Dai et al., 2023)","reference_id":"MzYxODQzNjIwNTU3NDU3MTE1OQ=="} +{"index":1196,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2304.02643","reference_contexts":"Some methods (Tancik et al., 2020; Kirillov et al., 2023) first encode points and boxes to positional encodings then add them to intermediate features or learned queries","reference_id":"MTc3MzQ1MDc0NDEwNTQ2OTk5OTE="} +{"index":1197,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2301.13823","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1198,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.03726","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1199,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2301.12597","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1200,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.10355","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1201,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2304.08485","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1202,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2303.05499","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1203,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2206.08916","reference_contexts":"Following Pix2seq, several methods, e.g., OFA (Wang et al., 2022b) , Unified-IO (Lu et al., 2022) , UniTab (Yang et al., 2022) , GIT (Wang et al., 2022a) , and VisionLLM (Wang et al., 2023b) introduce similar coordinate vocabulary alongside the language vocabulary for object detection and REC tasks","reference_id":"MjM1NjQyMTIxODczNjAxNTg5Ng=="} +{"index":1204,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2206.08916","reference_contexts":"GRIT refexp is the ablation split (Lu et al., 2022)","reference_id":"MzcyOTE5MjY0NzQzNTIwMjY5OQ=="} +{"index":1205,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2011.13681","reference_contexts":"Table 5: Comparing pointQA capabilities on the LookTwice-QA (Mani et al., 2020) , where the models are asked to answer question based on the input point/box","reference_id":"ODExNzQ1NDY5Mzg3NjE3NDE4Mw=="} +{"index":1206,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2011.13681","reference_contexts":"Correspondingly, to quantitatively evaluate our model's understanding of position inputs, we evaluated our model on two types PointQA datasets, LookTwice-QA of (Mani et al., 2020) and Vi-sual7W (PointQA Setting) of (Zhu et al., 2016) .","reference_id":"MTA5ODQwMTEzNTkzMzczNzYwNTE="} +{"index":1207,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2011.13681","reference_contexts":"We collection training data from public VQA, Image Captioning datset, and several datasets already containing positional annotation, such as Re-fCOCO (Kazemzadeh et al., 2014) for REC/REG, visual gemone (Krishna et al., 2017) for grounding caption, Visual-7W (Mani et al., 2020) for PointQA","reference_id":"MTUwODU1MjY1MzYwNDU5NTI2NjA="} +{"index":1208,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2011.13681","reference_contexts":"PointQA (Mani et al., 2020) requires a model answer for a visual question where the questioner queries a specific position in the picture","reference_id":"MzMxNzI3MTUxNzQyNjAyMTY1NA=="} +{"index":1209,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.15021","reference_contexts":"Recently, interesting works have made remarkable progress by extending LLM to audio, e.g., KOSMOS-1 (Huang et al., 2023) , X-LLM (Chen et al., 2023) , PandaGPT (Su et al., 2023) and control systems like PaLM-E (Driess et al., 2023) and EmbodiedGPT (Mu et al., 2023)","reference_id":"MTA1MDQ2NTU1NDYwODc4MjM0MTY="} +{"index":1210,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.16355","reference_contexts":"Recently, interesting works have made remarkable progress by extending LLM to audio, e.g., KOSMOS-1 (Huang et al., 2023) , X-LLM (Chen et al., 2023) , PandaGPT (Su et al., 2023) and control systems like PaLM-E (Driess et al., 2023) and EmbodiedGPT (Mu et al., 2023)","reference_id":"MTA1MDQ2NTU1NDYwODc4MjM0MTY="} +{"index":1211,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2205.14100","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1212,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.11172","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1213,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.11175","reference_contexts":"Following Pix2seq, several methods, e.g., OFA (Wang et al., 2022b) , Unified-IO (Lu et al., 2022) , UniTab (Yang et al., 2022) , GIT (Wang et al., 2022a) , and VisionLLM (Wang et al., 2023b) introduce similar coordinate vocabulary alongside the language vocabulary for object detection and REC tasks","reference_id":"MjM1NjQyMTIxODczNjAxNTg5Ng=="} +{"index":1214,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2306.09265","reference_contexts":"Here, we also provide VQAv2 val (83.3) and OK-VQA (53.8) results on LVLM-eHub toolbox (Xu et al., 2023) for easy comparison","reference_id":"Mzg0NDk0ODg2NTUwNzk0NDM4Ng=="} +{"index":1215,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2304.14178","reference_contexts":"Mini-GPT4 (Zhu et al., 2023) , mPLUG-OWL (Ye et al., 2023) , VPGTrans (Zhang et al., 2023a) , and InstructBLIP (Dai et al., 2023) retain Q-Former, replace language model to a larger one, and then tuning on meticulously collected instruction data","reference_id":"MTE5OTE3NzQwODg0Mjk0OTcxNA=="} +{"index":1216,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2305.01278","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1217,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2302.00923","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1218,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2304.10592","reference_contexts":"Mini-GPT4 (Zhu et al., 2023) , mPLUG-OWL (Ye et al., 2023) , VPGTrans (Zhang et al., 2023a) , and InstructBLIP (Dai et al., 2023) retain Q-Former, replace language model to a larger one, and then tuning on meticulously collected instruction data","reference_id":"MTE5OTE3NzQwODg0Mjk0OTcxNA=="} +{"index":1219,"citer_arxiv_id":"2306.15195","cited_arxiv_id":"2304.10592","reference_contexts":"In recent months, Multimodal Large Language Models (MLLMs) have witness remarkable progress (Alayrac et al., 2022; Huang et al., 2023; Liu et al., 2023a; Zhu et al., 2023; Li et al., 2023a; Gao et al., 2023; Dai et al., 2023)","reference_id":"MzYxODQzNjIwNTU3NDU3MTE1OQ=="} +{"index":1220,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2308.12966","reference_contexts":"Considering their impressive performance on complex vision-language tasks (Chen et al., 2023; Bai et al., 2023; Li et al., 2023a) , it is counter-intuitive that the hallucination problem of LVLMs is so severe","reference_id":"MjU0MDE3MjE3MzAyMDc5ODg0OA=="} +{"index":1221,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2302.04023","reference_contexts":"Especially, LLMs tend to hallucinate unintended text (Huang et al., 2021; Bang et al., 2023) , and VLPMs might generate nonexistent objects in the image (Biten et al., 2022) (termed as object hallucination)","reference_id":"NDE0MTQ3Mjg0NjIwNDc1ODk5Mw=="} +{"index":1222,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2306.15195","reference_contexts":"Considering their impressive performance on complex vision-language tasks (Chen et al., 2023; Bai et al., 2023; Li et al., 2023a) , it is counter-intuitive that the hallucination problem of LVLMs is so severe","reference_id":"MjU0MDE3MjE3MzAyMDc5ODg0OA=="} +{"index":1223,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2305.06500","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1224,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.15010","reference_contexts":"Following this line of research, a surge of studies (Zhu et al., 2023; Gao et al., 2023; Li et al., 2023a) have been proposed to enhance the visionlanguage pre-trained model (VLPM) (Gan et al., 2022) by incorporating powerful LLMs (Touvron et al., 2023; Chiang et al., 2023) , which are called large vision-language model (LVLM)","reference_id":"ODA0MjIwNzU2MjgyMTg1MzM4Ng=="} +{"index":1225,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2305.04790","reference_contexts":"They can be formulated as: We select five recently released LVLMs, i.e., mPLUG-Owl (Ye et al., 2023) , LLaVA (Liu et al., 2023) , Multimodal-GPT (Gong et al., 2023) , MiniGPT-4 (Zhu et al., 2023) and Instruct-BLIP (Dai et al., 2023a ) and prompt them with following instructions to generate captions about images in MSCOCO (Lin et al., 2014) :","reference_id":"NjYwMjk5ODk4NDEwNTM5OTUxOA=="} +{"index":1226,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2104.14839","reference_contexts":"Especially, LLMs tend to hallucinate unintended text (Huang et al., 2021; Bang et al., 2023) , and VLPMs might generate nonexistent objects in the image (Biten et al., 2022) (termed as object hallucination)","reference_id":"NDE0MTQ3Mjg0NjIwNDc1ODk5Mw=="} +{"index":1227,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2305.03726","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1228,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2301.12597","reference_contexts":"The success of LLMs drives the researchers to devise more powerful multimodal models based on the superior capacity of LLMs, to enhance the understanding of visual semantics (Alayrac et al., 2022; Li et al., 2023b)","reference_id":"NDYwMzA1NjQ0MDkyMjYzOTc3MA=="} +{"index":1229,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.08485","reference_contexts":"Since LLMs have been shown to be general task solvers in a zero-shot/few-shot manner, a number of studies are devoted to improving VLPM by integrating powerful LLMs for more accurate language understanding and generation (Zhu et al., 2023; Liu et al., 2023; Dai et al., 2023a)","reference_id":"MTc4NTcyNjM2MDE4MzU1NDI3ODM="} +{"index":1230,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.08485","reference_contexts":"They can be formulated as: We select five recently released LVLMs, i.e., mPLUG-Owl (Ye et al., 2023) , LLaVA (Liu et al., 2023) , Multimodal-GPT (Gong et al., 2023) , MiniGPT-4 (Zhu et al., 2023) and Instruct-BLIP (Dai et al., 2023a ) and prompt them with following instructions to generate captions about images in MSCOCO (Lin et al., 2014) :","reference_id":"NjYwMjk5ODk4NDEwNTM5OTUxOA=="} +{"index":1231,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.08485","reference_contexts":"In contrast, other LVLMs mostly employ the visual instructions generated by unimodal LLMs (Liu et al., 2023)","reference_id":"MTE3MjYwNDg5NjEyNzA5MDkyMg=="} +{"index":1232,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.08485","reference_contexts":"After vision-language pre-training (Alayrac et al., 2022; Li et al., 2022b) and visual instruction tuning (Liu et al., 2023) , LVLMs can fulfill complex tasks according to human instructions, demonstrating strong capacities in solving various vision-language tasks, e.g., image captioning (Ordonez et al., 2011; Hodosh et al., 2015; Sharma et al., 2018; Agrawal et al., 2019) and visual question answering (Antol et al., 2015a; Zhang et al., 2016; Goyal et al., 2017) .","reference_id":"MTc0Njk1NTk4MjYwNzMwODg1NzI="} +{"index":1233,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2303.08774","reference_contexts":"It is generally believed that the hallucination would degrade the model performance and greatly harm the user experiences in real-world applications (MacLeod et al., 2017; Ji et al., 2022)","reference_id":"MTEyMzk5MzI3MDk4MzQwODk3ODU="} +{"index":1234,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2303.08774","reference_contexts":"The hallucination phenomenon hinders the safe use of LVLMs in real-world deployment, as it may result in unexpected consequences caused by these hallucinated objects (MacLeod et al., 2017)","reference_id":"NjgzODk4MjIyNDk3NDk5NTk4Mw=="} +{"index":1235,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2302.13971","reference_contexts":"Following this line of research, a surge of studies (Zhu et al., 2023; Gao et al., 2023; Li et al., 2023a) have been proposed to enhance the visionlanguage pre-trained model (VLPM) (Gan et al., 2022) by incorporating powerful LLMs (Touvron et al., 2023; Chiang et al., 2023) , which are called large vision-language model (LVLM)","reference_id":"ODA0MjIwNzU2MjgyMTg1MzM4Ng=="} +{"index":1236,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.14178","reference_contexts":"They can be formulated as: We select five recently released LVLMs, i.e., mPLUG-Owl (Ye et al., 2023) , LLaVA (Liu et al., 2023) , Multimodal-GPT (Gong et al., 2023) , MiniGPT-4 (Zhu et al., 2023) and Instruct-BLIP (Dai et al., 2023a ) and prompt them with following instructions to generate captions about images in MSCOCO (Lin et al., 2014) :","reference_id":"NjYwMjk5ODk4NDEwNTM5OTUxOA=="} +{"index":1237,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2303.18223","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1238,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.10592","reference_contexts":"Following this line of research, a surge of studies (Zhu et al., 2023; Gao et al., 2023; Li et al., 2023a) have been proposed to enhance the visionlanguage pre-trained model (VLPM) (Gan et al., 2022) by incorporating powerful LLMs (Touvron et al., 2023; Chiang et al., 2023) , which are called large vision-language model (LVLM)","reference_id":"ODA0MjIwNzU2MjgyMTg1MzM4Ng=="} +{"index":1239,"citer_arxiv_id":"2305.10355","cited_arxiv_id":"2304.10592","reference_contexts":"Since LLMs have been shown to be general task solvers in a zero-shot/few-shot manner, a number of studies are devoted to improving VLPM by integrating powerful LLMs for more accurate language understanding and generation (Zhu et al., 2023; Liu et al., 2023; Dai et al., 2023a)","reference_id":"MTc4NTcyNjM2MDE4MzU1NDI3ODM="} +{"index":1240,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2309.16609","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1241,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2308.12966","reference_contexts":"In the realm of proprietary models, the APIs of multiple powerful VLMs have also been made publicly available to prosper downstream applications, including GPT-4v [37] , Gemini-Pro-V [44] , and Qwen-VL-Max [6]","reference_id":"Njg0NTg2NDYyNzUyMzY5NDgzNg=="} +{"index":1242,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2308.12966","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1243,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2311.12793","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1244,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"1504.00325","reference_contexts":"Problem Source Number Ratio ARAS [15] customize & selection 76 2.4% CLEVR [24] customize & selection 14 0.4% COCO [9] customize & selection 179 5.6% KonIQ-10k [21] customize & selection 32 1.0% LLaVA [33] customize 19 0.6% PISC [29] customize & selection 15 0.5% Places [54] customize & selection 59 1.8% ScienceQA [34] customize & selection 156 4.8% ShapeWorld [25] customize & selection 20 0.6% TextVQA [42] customize & selection 18 0.6% VSR [31] customize & selection 19 0.6% W3C School [1] customize 20 0.6% Internet customize 2590 80.5%","reference_id":"NTg4MTMwNjI4MTkxNDg4MDA1MA=="} +{"index":1245,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"1504.00325","reference_contexts":"The first approach involves utilizing existing public datasets [19, 9] for objective evaluation, while the second approach employs human annotators [49, 48] to perform subjective evaluations","reference_id":"ODc2Njk2ODg2ODI2OTIxMzcwOA=="} +{"index":1246,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"1504.00325","reference_contexts":"A multitude of public datasets, such as VQAv2 [19] , COCO Caption [9] , GQA [23] , and OK-VQA [35] , have long served as valuable resources for the quantitative evaluation of VLMs","reference_id":"MTUwNDk0Nzg4MTI2ODc2NjkyNTY="} +{"index":1247,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"1504.00325","reference_contexts":"COCO Caption [9] , Nocaps [3] , and Flickr30k [51] provide human-generated image captions and the corresponding task is to describe the image content in the form of text","reference_id":"MTM5NDM5NTE5NTE0MDk2MjA2NzU="} +{"index":1248,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2305.06500","reference_contexts":"Subsequently, InstructBLIP [11] extends BLIP-2 [28] with vision-language instruction tuning, achieving superior performance","reference_id":"ODMyMTQzMTA0MTA5MzcyNDYzMw=="} +{"index":1249,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2305.06500","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1250,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2401.16420","reference_contexts":"During the instruction tuning, Low-Rank Adaptation (LoRA [22] ) has been adopted by recent works [49, 12, 10] on language models to achieve better performance on multimodal understanding","reference_id":"MTAwNzY3MDY1Njc3NTY5NjA4MTE="} +{"index":1251,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2401.16420","reference_contexts":"Among open-source VLMs, InternLM-XComposer2 [12] achieves the best performance and surpass other open-source or proprietary models by a large margin, w.r.t","reference_id":"MTEwNDE1MTAyMDc5NzQ3Mjc1NTc="} +{"index":1252,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2401.16420","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1253,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2306.13394","reference_contexts":"To support this claim, we apply the quality control paradigm to other popular multimodal evaluation benchmarks (like MME [17] and SEEDBench [27] ) and try to detect the low-quality samples","reference_id":"NzE2MDE4NzM2MjEwODU1NDMwOA=="} +{"index":1254,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2306.13394","reference_contexts":"[17] introduced MME, which assesses a VLM's capabilities from various perspectives at a small scale","reference_id":"MTQyODg4NzQ2MzMxNzg4OTY1OTk="} +{"index":1255,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2106.09685","reference_contexts":"During the instruction tuning, Low-Rank Adaptation (LoRA [22] ) has been adopted by recent works [49, 12, 10] on language models to achieve better performance on multimodal understanding","reference_id":"MTAwNzY3MDY1Njc3NTY5NjA4MTE="} +{"index":1256,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"1704.04517","reference_contexts":"Problem Source Number Ratio ARAS [15] customize & selection 76 2.4% CLEVR [24] customize & selection 14 0.4% COCO [9] customize & selection 179 5.6% KonIQ-10k [21] customize & selection 32 1.0% LLaVA [33] customize 19 0.6% PISC [29] customize & selection 15 0.5% Places [54] customize & selection 59 1.8% ScienceQA [34] customize & selection 156 4.8% ShapeWorld [25] customize & selection 20 0.6% TextVQA [42] customize & selection 18 0.6% VSR [31] customize & selection 19 0.6% W3C School [1] customize 20 0.6% Internet customize 2590 80.5%","reference_id":"NTg4MTMwNjI4MTkxNDg4MDA1MA=="} +{"index":1257,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2307.16125","reference_contexts":"To support this claim, we apply the quality control paradigm to other popular multimodal evaluation benchmarks (like MME [17] and SEEDBench [27] ) and try to detect the low-quality samples","reference_id":"NzE2MDE4NzM2MjEwODU1NDMwOA=="} +{"index":1258,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2301.12597","reference_contexts":"BLIP-2 [28] introduces a Querying Transformer (Q-former) to bridge the modality gap between the frozen image encoder and the large language encoder","reference_id":"MTE0MjIwNzYyMjcyMTI2ODMzODA="} +{"index":1259,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2311.06607","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1260,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2310.03744","reference_contexts":"The learning paradigm and the multimodal instruction tuning corpus proposed by LLaVA are widely adopted by subsequent works [32, 8, 2, 10]","reference_id":"MTYzMDgxNTUzNzQzNjEwMTQ2NzM="} +{"index":1261,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2310.03744","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1262,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.08485","reference_contexts":"Notable works, such as GPT-4v [37] , Gemini-Pro-V [44] and LLaVA [33] , have demonstrated enhanced capabilities in image content recognition and reasoning within the domain of vision-language models, exhibiting superior performance compared to earlier works","reference_id":"NDA2Nzc0OTQ3NjkyNzgyODQ4Mw=="} +{"index":1263,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.08485","reference_contexts":"LLaVA [33] also utilizes GPT-4 to generate instruction-following data for vision-language tuning","reference_id":"NTU1NjcyNjY5ODEwNzEzNDEy"} +{"index":1264,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.08485","reference_contexts":"After that, models adopting the architecture of LLaVA [33] (LLaVA series and Yi-VL series) also showcase strong overall performance, which is just inferior to the state-of-the-art closedsource GPT-4v and Qwen-VL-Max","reference_id":"OTUwMzg0MzI5MzA1OTAyNDkyOA=="} +{"index":1265,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.08485","reference_contexts":"Problem Source Number Ratio ARAS [15] customize & selection 76 2.4% CLEVR [24] customize & selection 14 0.4% COCO [9] customize & selection 179 5.6% KonIQ-10k [21] customize & selection 32 1.0% LLaVA [33] customize 19 0.6% PISC [29] customize & selection 15 0.5% Places [54] customize & selection 59 1.8% ScienceQA [34] customize & selection 156 4.8% ShapeWorld [25] customize & selection 20 0.6% TextVQA [42] customize & selection 18 0.6% VSR [31] customize & selection 19 0.6% W3C School [1] customize 20 0.6% Internet customize 2590 80.5%","reference_id":"NTg4MTMwNjI4MTkxNDg4MDA1MA=="} +{"index":1266,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2303.08774","reference_contexts":"We apply three state-of-the-art LLMs, including GPT-4 [37] , Gemini-Pro [44] , and Qwen-Max [5] to infer the questions with text-only inputs under CircularEval","reference_id":"Mzc1OTE3NTQxODU4MTk2MDc1Ng=="} +{"index":1267,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2303.08774","reference_contexts":"We adopt 'majority voting' to detect text-only samples: data samples are inferred with state-of-the-art LLMs (GPT-4 [37] , Gemini-Pro [44] , etc.)","reference_id":"MTA4MjAwMDU1MDQ0NDgwNTcxOTA="} +{"index":1268,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2303.08774","reference_contexts":"For instance, the latest LLMs, such as OpenAI's ChatGPT and GPT-4 [37] , have demonstrated remarkable reasoning capabilities that are comparable to, and in some cases, even surpass human capabilities","reference_id":"ODA1NDc1NTg5MTYwNDA0NjU4Mw=="} +{"index":1269,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2303.08774","reference_contexts":"MiniGPT-4 [56] attributes the prowess of GPT-4 [37] to advanced LLMs and proposes the use of a single projection layer to align the visual representation with the language model","reference_id":"NzcxNDQ2MDU3MTMzMjc5OTI3MA=="} +{"index":1270,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2303.08774","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1271,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2305.16355","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1272,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2312.11805","reference_contexts":"We apply three state-of-the-art LLMs, including GPT-4 [37] , Gemini-Pro [44] , and Qwen-Max [5] to infer the questions with text-only inputs under CircularEval","reference_id":"Mzc1OTE3NTQxODU4MTk2MDc1Ng=="} +{"index":1273,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2312.11805","reference_contexts":"In the realm of proprietary models, the APIs of multiple powerful VLMs have also been made publicly available to prosper downstream applications, including GPT-4v [37] , Gemini-Pro-V [44] , and Qwen-VL-Max [6]","reference_id":"Njg0NTg2NDYyNzUyMzY5NDgzNg=="} +{"index":1274,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2312.11805","reference_contexts":"We adopt 'majority voting' to detect text-only samples: data samples are inferred with state-of-the-art LLMs (GPT-4 [37] , Gemini-Pro [44] , etc.)","reference_id":"MTA4MjAwMDU1MDQ0NDgwNTcxOTA="} +{"index":1275,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2312.11805","reference_contexts":"Notable works, such as GPT-4v [37] , Gemini-Pro-V [44] and LLaVA [33] , have demonstrated enhanced capabilities in image content recognition and reasoning within the domain of vision-language models, exhibiting superior performance compared to earlier works","reference_id":"NDA2Nzc0OTQ3NjkyNzgyODQ4Mw=="} +{"index":1276,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2312.11805","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1277,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2302.13971","reference_contexts":"Building upon the success of Large Language Models (LLMs) such as GPTs [41, 7, 40] , LLaMA [46] , and Vicuna [53] , recent advancements have been made in multimodal models","reference_id":"ODM1MDkwMzM2MzYwMzk5NTA3Mg=="} +{"index":1278,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2311.03079","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1279,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.14178","reference_contexts":"The first approach involves utilizing existing public datasets [19, 9] for objective evaluation, while the second approach employs human annotators [49, 48] to perform subjective evaluations","reference_id":"ODc2Njk2ODg2ODI2OTIxMzcwOA=="} +{"index":1280,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.14178","reference_contexts":"Given the aforementioned challenges, recent studies, such as OwlEval [49] and LVLM-eHub [48] propose human-involved subjective evaluation strategies, aiming to address existing methods' limitations by incorporating human judgment and perception in the evaluation process","reference_id":"MzYyMjExNjUwNTMzNTA3OTIxNA=="} +{"index":1281,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.14178","reference_contexts":"[49] constructed OwlEval, an evaluation set encompassing a variety of visual-related tasks, albeit of a limited size","reference_id":"NzgzMDUwNzkyNzY3NDkzMzMzOA=="} +{"index":1282,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.14178","reference_contexts":"During the instruction tuning, Low-Rank Adaptation (LoRA [22] ) has been adopted by recent works [49, 12, 10] on language models to achieve better performance on multimodal understanding","reference_id":"MTAwNzY3MDY1Njc3NTY5NjA4MTE="} +{"index":1283,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2311.04257","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1284,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2309.15112","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1285,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.10592","reference_contexts":"MiniGPT-4 [56] attributes the prowess of GPT-4 [37] to advanced LLMs and proposes the use of a single projection layer to align the visual representation with the language model","reference_id":"NzcxNDQ2MDU3MTMzMjc5OTI3MA=="} +{"index":1286,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.10592","reference_contexts":"Nevertheless, a large proportion of the early studies [18, 56, 33] tend to emphasize showcasing qualitative examples rather than undertaking comprehensive and quantitative experiments to thoroughly assess their model performance","reference_id":"MTA3NjQ4NjYwNzA3NjIwNTQwNjM="} +{"index":1287,"citer_arxiv_id":"2307.06281","cited_arxiv_id":"2304.10592","reference_contexts":"For the main results, we evaluate various models belonging to three major categories on MMBench: (a) Text-Only GPT-4 [37] ; (b) Open-Source VLMs including model variants of OpenFlamingo [4], MiniGPT4 [56] , InstructBLIP [11] , LLaVA [32] , IDEFICS [26] , CogVLM [47], Qwen-VL [6] , Yi-VL [2], mPLUG-Owl [50] , InternLM-XComposer [12] , and MiniCPM-V [39] ; (c) Proprietary VLMs including Qwen-VL-[Plus/Max] [6], Gemini-Pro-V [44] , and GPT-4v [37]","reference_id":"MTM1ODcxOTIzMTM0MDc5OTcyNTQ="} +{"index":1288,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2309.16609","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1289,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2308.12966","reference_contexts":"In the vision-language learning area, several initiatives [13, 16, 23, 38] have been undertaken to enhance the quality of captions within imagetext pairs","reference_id":"ODIzNTkyNDkyMDIxNDAzMTEyMA=="} +{"index":1290,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2308.12966","reference_contexts":"A recent work, VeCLIP [23] , proposes using LLMs to amalgamate information from both raw and synthetic captions","reference_id":"NzgwNDYxMTA3MjkxMjk0ODE2Mw=="} +{"index":1291,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.09478","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1292,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2306.15195","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1293,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"1504.00325","reference_contexts":"In detail, the 665k data is gathered from publicly available academic task-oriented data [1, 20, 22, 36, 37, 49, 51] and instruction-tuning data for conversational and complex reasoning tasks [31] involving natural images [29]","reference_id":"MzkzNzQzMTQ2ODA0MzY3NjkyNw=="} +{"index":1294,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2204.02311","reference_contexts":"Subsequent models like PaLM [9] extended the limits of model parameters and dataset scale, while others like InstructGPT [42] and ChatGPT [39] introduced fine-tuning and reinforcement learning techniques for improved conversational interaction","reference_id":"NTkyOTE4NTI3MzQ5NzY5MTc2NA=="} +{"index":1295,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2204.02311","reference_contexts":"Recent breakthroughs in artificial intelligence have been driven notably by the development of large language models (LLMs) [2, 4, 8, 9, 12, 53, 56]","reference_id":"MTY1NTYwMDY1OTU5MDYxNTA2NDM="} +{"index":1296,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"1810.04805","reference_contexts":"Early encoderdecoder models like BERT [11] and T5 [46] , and decodercentric models such as GPT [44], leveraged the Transformer architecture [54] to excel in various NLP tasks","reference_id":"OTQ3ODM3OTQxODQyODUyOTYzNA=="} +{"index":1297,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2103.10360","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1298,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2305.20088","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1299,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2306.13394","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1300,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.14108","reference_contexts":"In the vision-language learning area, several initiatives [13, 16, 23, 38] have been undertaken to enhance the quality of captions within imagetext pairs","reference_id":"ODIzNTkyNDkyMDIxNDAzMTEyMA=="} +{"index":1301,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.14108","reference_contexts":"Research [16, 38] explores methods to filter and blend raw and synthetic captions to enhance the CLIP model","reference_id":"NDcyMzE0MjAzODU2ODA4Njg5OA=="} +{"index":1302,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.02643","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1303,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.07699","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1304,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2307.16125","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1305,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2305.03726","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1306,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2301.12597","reference_contexts":"A series of works [26, 27] were improved upon CLIP by employing refined data strategies for more diverse data, they have been effective for basic visual tasks [28, 32, 59 ] but less so for complex tasks like visual question answering","reference_id":"NzI3MDcwNjQ4MzEyNTI0NjY5NA=="} +{"index":1307,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.03744","reference_contexts":"The replacement ratio is only 3.5% for LLaVA-1.5 [30] and Qwen-VL-Chat [3] 1 , and 14.5% for LLaVA [31] .","reference_id":"MjY5MDgzNjYwODk5MTIzMDQyMQ=="} +{"index":1308,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.03744","reference_contexts":"Following the evolution, modality unification via LLMs becomes the inevitable tendency, and visual-aligned multi-modal LLMs[3, 5, 10, 30, 31, 35, 57, [60] [61] [62] have witnessed ever-changing advances in recent days","reference_id":"MTU4MjkxMDI4NDQzNjE2MTM2NzA="} +{"index":1309,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.03744","reference_contexts":"These include 118K images from COCO [29],570K images from SAM [21], and 558K images from LLaVA-1.5 pre-training data [30] .","reference_id":"NjkxMDAxODk2NTQ2MTM3NDI1NA=="} +{"index":1310,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.03744","reference_contexts":"The ShareGPT4V-7B model follows the design of LLaVA-1.5 [30] , including three integral components: (1) A vision encoder utilizing the CLIP-Large model [45] , with a resolution of 336×336 and a patch size of 14, converting input images into 576 tokens","reference_id":"MjkxMTE5MjI2Mjk4Njg2NDExMg=="} +{"index":1311,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.03744","reference_contexts":"In previous LMM research [5, 30, 31, 62] , the vision encoder is generally not fine-tuned during pre-training, a rational approach considering the lower quality of previously used captions, where fine-tuned the vision encoder might degrade its visual knowledge extraction ability","reference_id":"MTU3Nzk1MzAyMjgyMzk2OTg3NDc="} +{"index":1312,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2310.03744","reference_contexts":"We conducted a straightforward experiment to verify the quality of the collected data: we chose a range of advanced, publicly available LMMs, including LLaVA-7B [31] , LLaVA-1.5-7B [30] , LLaVA-1.5-13B [30] , and Qwen-VL-Chat-7B [3]","reference_id":"MTY4NDY4NTk2MDUzODExMjQyNTQ="} +{"index":1313,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"The replacement ratio is only 3.5% for LLaVA-1.5 [30] and Qwen-VL-Chat [3] 1 , and 14.5% for LLaVA [31] .","reference_id":"MjY5MDgzNjYwODk5MTIzMDQyMQ=="} +{"index":1314,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"Following the evolution, modality unification via LLMs becomes the inevitable tendency, and visual-aligned multi-modal LLMs[3, 5, 10, 30, 31, 35, 57, [60] [61] [62] have witnessed ever-changing advances in recent days","reference_id":"MTU4MjkxMDI4NDQzNjE2MTM2NzA="} +{"index":1315,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"In detail, the 665k data is gathered from publicly available academic task-oriented data [1, 20, 22, 36, 37, 49, 51] and instruction-tuning data for conversational and complex reasoning tasks [31] involving natural images [29]","reference_id":"MzkzNzQzMTQ2ODA0MzY3NjkyNw=="} +{"index":1316,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"LLaVA-Instruct [31] combines human-made captions, bounding boxes, and GPT4 [39] to 'imagine' the image details, which leads to inevitable error/hallucination description (marked in red)","reference_id":"ODMxODQ0NDg5NjMxMTgyMzQxNQ=="} +{"index":1317,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"To the best of our knowledge, in the LMM area, LLaVA [31] uniquely inputs human-annotated short captions and bounding boxes into the GPT4 language model","reference_id":"NTkwNTY5NzMxODE3NzY2NDA1MQ=="} +{"index":1318,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"In previous LMM research [5, 30, 31, 62] , the vision encoder is generally not fine-tuned during pre-training, a rational approach considering the lower quality of previously used captions, where fine-tuned the vision encoder might degrade its visual knowledge extraction ability","reference_id":"MTU3Nzk1MzAyMjgyMzk2OTg3NDc="} +{"index":1319,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"The LLaVA (in the wild) benchmark [31] is composed of 60 questions, spanning three distinct tasks: conversation, complex reasoning, and detailed description","reference_id":"NjQzODU0OTI1OTIxNTQ5MTY4Mg=="} +{"index":1320,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.08485","reference_contexts":"We conducted a straightforward experiment to verify the quality of the collected data: we chose a range of advanced, publicly available LMMs, including LLaVA-7B [31] , LLaVA-1.5-7B [30] , LLaVA-1.5-13B [30] , and Qwen-VL-Chat-7B [3]","reference_id":"MTY4NDY4NTk2MDUzODExMjQyNTQ="} +{"index":1321,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2303.05499","reference_contexts":"A series of works [26, 27] were improved upon CLIP by employing refined data strategies for more diverse data, they have been effective for basic visual tasks [28, 32, 59 ] but less so for complex tasks like visual question answering","reference_id":"NzI3MDcwNjQ4MzEyNTI0NjY5NA=="} +{"index":1322,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2307.06281","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1323,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2305.15023","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1324,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2307.10350","reference_contexts":"In the vision-language learning area, several initiatives [13, 16, 23, 38] have been undertaken to enhance the quality of captions within imagetext pairs","reference_id":"ODIzNTkyNDkyMDIxNDAzMTEyMA=="} +{"index":1325,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2306.14824","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1326,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"1505.00855","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1327,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2111.02114","reference_contexts":"To maximize the diversity and comprehensiveness of our data, we compiled around 100K images from various data sources, including images for detection [29] and segmentation [21] , complex text-containing images [51] , as well as various web images [41, 48, 50] con-taining artworks, landmarks, celebrities etc","reference_id":"MTU4NjQwNDA5MzYxNjM3NzkzNzU="} +{"index":1328,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2111.02114","reference_contexts":"This includes 50K images from COCO [29], 30K images from 'LCS' (which abbreviates LAION [48] , CC-3M [50] , and SBU [41] ), 20K images from SAM [21], 500 images from TextCaps [51] , 500 images from WikiArt [47], and 1K images from webcrawled data (split evenly between images of landmarks and images of celebrities).","reference_id":"MTM3MjU5MjExNTM1MTE0MTc1MTM="} +{"index":1329,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2111.02114","reference_contexts":"'LCS' abbreviates the LAION [48] , CC [50] , and SBU [47] datasets","reference_id":"MTUyODk5NjcwNTA0NTkyNDA1ODA="} +{"index":1330,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2307.09288","reference_contexts":"(3) A LLM, based on the open-source Vicuna-v1.5 [8] , derived from LLaMA2 [53]","reference_id":"NDEwNjYxMjA4MTg4Mjc2MTkyOQ=="} +{"index":1331,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2307.09288","reference_contexts":"Recent breakthroughs in artificial intelligence have been driven notably by the development of large language models (LLMs) [2, 4, 8, 9, 12, 53, 56]","reference_id":"MTY1NTYwMDY1OTU5MDYxNTA2NDM="} +{"index":1332,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2307.09288","reference_contexts":"These developments, along with contributions from the open-source community [8, 52, 53, 53, 56] , have set new benchmarks and opened avenues for future research in NLP area.","reference_id":"MTExMTc4MTMxOTg5MTk4NjcyMjY="} +{"index":1333,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2309.14181","reference_contexts":"Q-Bench [55] assesses low-level perception, while VQA-v2 [17] and VisWiz [18] are benchmarks in the realm of traditional Visual Question Answering (VQA) tasks.","reference_id":"MTMwNDIwNDc4NDc3NjgyODk0Nzg="} +{"index":1334,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2309.14181","reference_contexts":"; MM-Vet [58] ; QBench [55] ; SQA I : ScienceQA-IMG [34] ; VQA V 2 [17]; VizWiz [18]","reference_id":"MTYwNDY5MDU0NjM5MzU0ODIzMDI="} +{"index":1335,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2309.10305","reference_contexts":"Recent breakthroughs in artificial intelligence have been driven notably by the development of large language models (LLMs) [2, 4, 8, 9, 12, 53, 56]","reference_id":"MTY1NTYwMDY1OTU5MDYxNTA2NDM="} +{"index":1336,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2309.10305","reference_contexts":"These developments, along with contributions from the open-source community [8, 52, 53, 53, 56] , have set new benchmarks and opened avenues for future research in NLP area.","reference_id":"MTExMTc4MTMxOTg5MTk4NjcyMjY="} +{"index":1337,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.14178","reference_contexts":"Following the evolution, modality unification via LLMs becomes the inevitable tendency, and visual-aligned multi-modal LLMs[3, 5, 10, 30, 31, 35, 57, [60] [61] [62] have witnessed ever-changing advances in recent days","reference_id":"MTU4MjkxMDI4NDQzNjE2MTM2NzA="} +{"index":1338,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2308.02490","reference_contexts":"; MM-Vet [58] ; QBench [55] ; SQA I : ScienceQA-IMG [34] ; VQA V 2 [17]; VizWiz [18]","reference_id":"MTYwNDY5MDU0NjM5MzU0ODIzMDI="} +{"index":1339,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2308.02490","reference_contexts":"MM-Vet [58] uses GPT4 for a six-dimensional LMM capability assessment","reference_id":"NDEzNjAzOTA5NTU1Mjc2NTM2"} +{"index":1340,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2309.15112","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1341,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2303.16199","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1342,"citer_arxiv_id":"2311.12793","cited_arxiv_id":"2304.10592","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1343,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.10403","reference_contexts":"The breakthroughs in large language models (LLMs) (Brown et al., 2020a; OpenAI, 2023c; Chowdhery et al., 2022; Anil et al., 2023; Touvron et al., 2023a; Hoffmann et al., 2022) bring generalist AI models that can solve a wide range of complicated natural language tasks, many approaching the human-expert-level performance (OpenAI, 2023c; Bubeck et al., 2023; Yang et al., 2023b)","reference_id":"Njk5NjMyMjkxMDg4NDA1OTY2"} +{"index":1344,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2308.01390","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1345,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.12712","reference_contexts":"The breakthroughs in large language models (LLMs) (Brown et al., 2020a; OpenAI, 2023c; Chowdhery et al., 2022; Anil et al., 2023; Touvron et al., 2023a; Hoffmann et al., 2022) bring generalist AI models that can solve a wide range of complicated natural language tasks, many approaching the human-expert-level performance (OpenAI, 2023c; Bubeck et al., 2023; Yang et al., 2023b)","reference_id":"Njk5NjMyMjkxMDg4NDA1OTY2"} +{"index":1346,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"1504.00325","reference_contexts":"Classic VL benchmarks focus on specific capabilities of interest, such as visual recognition (Goyal et al., 2017) , image description (Chen et al., 2015; Agrawal et al., 2019) , as well as other benchmarks for specialized capabilities such as scene text understanding (Singh et al., 2019; Sidorov et al., 2020; Yang et al., 2021) , commonsense reasoning (Zellers et al., 2019) , and outside knowledge (Marino et al., 2019)","reference_id":"MzMyMjg4OTE2MjgyODk3MDQzMw=="} +{"index":1347,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"1504.00325","reference_contexts":"As shown in Figure 1 , existing VL benchmarks (Antol et al., 2015; Chen et al., 2015; Singh et al., 2019) focus on straightforward VL tasks that test specific one or two capabilities, such as recognition, language generation, or OCR, but fall short in benchmarking more complicated tasks","reference_id":"ODYyNzkxMDMxNzM3MjkyNjMyMw=="} +{"index":1348,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"1504.00325","reference_contexts":"Vision-language models (Chen et al., 2015; Goyal et al., 2017; Lu et al., 2019; Chen et al., 2020; Li et al., 2020; Kim et al., 2021; Wang et al., 2022b; a; Yang et al., 2022a; Gan et al., 2022) approach multimodal intelligence of jointly understanding and generating vision and language signals","reference_id":"MTQ3OTA3Njc5OTk3ODg5ODY1NDg="} +{"index":1349,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.01937","reference_contexts":"Inspired by recent NLP studies (Chiang & Lee, 2023; Liu et al., 2023e; Fu et al., 2023b) that use LLMs for model evaluation, we propose an LLM-based evaluator as the evaluation metric for openended model outputs","reference_id":"MTEzNDc0OTczMjM3MDU5NjIwNzI="} +{"index":1350,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.01937","reference_contexts":"17.0 11.9 6.4 5.5 5.5 5.0 4.1 3.7 3.7 3.2 1.8 1.4 0.9 0.9 0.5 in NLP (Chiang & Lee, 2023; Liu et al., 2023e; Fu et al., 2023b)","reference_id":"MTEzNzc0NTIwMTY1ODUwNzY2OTA="} +{"index":1351,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2204.02311","reference_contexts":"The breakthroughs in large language models (LLMs) (Brown et al., 2020a; OpenAI, 2023c; Chowdhery et al., 2022; Anil et al., 2023; Touvron et al., 2023a; Hoffmann et al., 2022) bring generalist AI models that can solve a wide range of complicated natural language tasks, many approaching the human-expert-level performance (OpenAI, 2023c; Bubeck et al., 2023; Yang et al., 2023b)","reference_id":"Njk5NjMyMjkxMDg4NDA1OTY2"} +{"index":1352,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2204.02311","reference_contexts":"Inspired by the impressive quality and genericity in recent large language models (LLMs) (Brown et al., 2020b; OpenAI, 2023c; Chowdhery et al., 2022; Touvron et al., 2023a) , researchers explore large multimodal models (LMMs) that seamlessly integrate different visionlanguage capabilities to solve complicated multimodal tasks.","reference_id":"MTQzMDQ4ODEwNjQzMTA1NzQ4MjA="} +{"index":1353,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2210.11416","reference_contexts":"We mainly consider the models based on open-sourced LLMs, i.e., Flan-T5 (Chung et al., 2022) , LLaMA (Touvron et al., 2023a) , Vicuna (Zheng et al., 2023) , and LLaMA-2 (Touvron et al., 2023b) .","reference_id":"MTE4NTU4MjQ1MTAyOTQ3NDY4NjA="} +{"index":1354,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2207.04672","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1355,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.06500","reference_contexts":"Recently, thanks to the open-source of powerful LLMs like LLaMA (Touvron et al., 2023a) , more open-sourced LMMs are built, including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , and many more (Gong et al., 2023; Liu et al., 2023b; Ye et al., 2023)","reference_id":"Mjg1OTQ2MDk5MjMyOTM0MzIzOQ=="} +{"index":1356,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.06500","reference_contexts":"We consider both the end-to-end tuned models (i.e., OpenFlamingo (Alayrac et al., 2022; Awadalla et al., 2023a; b) , BLIP-2 (Li et al., 2023d) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , LLaMA-Adapter v2 (Gao et al., 2023b) , Otter (Li et al., 2023c) , and InstructBLIP (Dai et al., 2023) ), and the LLM-tool-using systems (i.e., MM-ReAct (Yang et al., 2023c) and Transformers Agent (Huggingface, 2023))","reference_id":"MTQ2MTUwNTg4MTYxMzcwNzUxNDI="} +{"index":1357,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.06500","reference_contexts":"Recent open-sourced LLMs (Zhang et al., 2022; Touvron et al., 2023a; Peng et al., 2023) also facilitate various research studies including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , and so on (Gong et al., 2023; Liu et al., 2023b; Ye et al., 2023) .","reference_id":"MTUzNDM3MTI4MzYzMzAyNDg4NTY="} +{"index":1358,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.06500","reference_contexts":"Stronger language models and the tuning data in LLava may be the reason for the good 55.5 89.2 68.6 73.9 83.3 77.5 44.5 38.9 78.2 76.5 42.9 100.0 66.7 50.0 89.0 0.0 67.7±0.3 GPT-4V-Turbo-detail:low (OpenAI, 2023a) 52.9 75.7 58.4 50.0 75.0 69.2 45.5 38.9 84.8 85.8 14.3 75.0 66.7 50.0 87.0 0.0 60.2±0.3 GPT-4V-Turbo-detail:high (OpenAI, 2023a) 50.2 77.7 82.5 85.8 75.0 80.0 54.5 38.9 81.5 78.8 42.9 100.0 66.7 95.0 90.0 0.0 67.6±0.1 InstructBLIP-14B and InstructBLIP-8B (Dai et al., 2023) achieve the best performance, which may result from the tuning data containing relevant datasets, like VQA (Goyal et al., 2017) and GQA (Hudson & Manning, 2019) .","reference_id":"MTQ4MTM1NzIwOTY0NzEzNzU5NTk="} +{"index":1359,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.06500","reference_contexts":"We evaluate two sets of multimodal systems, i.e., the end-to-end tuned LMMs including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , etc, and the LLM-tool-using systems (Yang et al., 2023c; Shen et al., 2023; Gao et al., 2023a; Huggingface, 2023) such as MM-ReAct (Yang et al., 2023c) and Transformers Agent (Huggingface, 2023)","reference_id":"OTUwNDQzODQxNzY0ODg5MDEzMQ=="} +{"index":1360,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2010.11929","reference_contexts":"This superior performance may be attributed to LLaVA's adoption of CLIP (Radford et al., 2021) ViT-L/14 (Dosovitskiy et al., 2020) as its vision model, and the inclusion of a large volume of image-OCR pairings within the training data (Liu et al., 2023f) .","reference_id":"OTQ3Nzk1NDYzODU0MzU1MTEwMQ=="} +{"index":1361,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.03378","reference_contexts":"Since more than 80% of our human being's perception, learning, cognition, and activities are mediated through vision (Politzer) , it is natural to start the exploration by equipping LLMs with \"eyes.\" One main thread of LMM works, represented by Frozen (Tsimpoukelli et al., 2021) , Flamingo (Alayrac et al., 2022) , PaLM-E (Driess et al., 2023) , GPT-4V (OpenAI, 2023c; a) , extend LLMs with the visual understanding capability via end-to-end tuning","reference_id":"MTgwMDEzMTY2ODg3NzA2MTIxMDc="} +{"index":1362,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.03378","reference_contexts":"In approaching such multimodal generalist systems, one direction is to extend LLMs with the multi-sensory ability, such as pioneer works Frozen (Tsimpoukelli et al., 2021) , Flamingo (Alayrac et al., 2022) , PaLM-E (Driess et al., 2023) and GPT-4V (OpenAI, 2023c; a)","reference_id":"MTM0NTk5Mzk4Mzk2NDE4Nzk5MDM="} +{"index":1363,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2306.13394","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1364,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2302.04166","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1365,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2210.09263","reference_contexts":"Vision-language models (Chen et al., 2015; Goyal et al., 2017; Lu et al., 2019; Chen et al., 2020; Li et al., 2020; Kim et al., 2021; Wang et al., 2022b; a; Yang et al., 2022a; Gan et al., 2022) approach multimodal intelligence of jointly understanding and generating vision and language signals","reference_id":"MTQ3OTA3Njc5OTk3ODg5ODY1NDg="} +{"index":1366,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2306.08640","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1367,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2211.10435","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1368,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.15010","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1369,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2203.15556","reference_contexts":"The breakthroughs in large language models (LLMs) (Brown et al., 2020a; OpenAI, 2023c; Chowdhery et al., 2022; Anil et al., 2023; Touvron et al., 2023a; Hoffmann et al., 2022) bring generalist AI models that can solve a wide range of complicated natural language tasks, many approaching the human-expert-level performance (OpenAI, 2023c; Bubeck et al., 2023; Yang et al., 2023b)","reference_id":"Njk5NjMyMjkxMDg4NDA1OTY2"} +{"index":1370,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"1910.13461","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1371,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2306.05425","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1372,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.03726","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1373,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2301.12597","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1374,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2306.14565","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1375,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.08485","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1376,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.08485","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1377,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2307.06281","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1378,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.16634","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1379,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.07895","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1380,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.03277","reference_contexts":"It may be beneficial from its large-scale tuning data including GPT-4-LLM (Peng et al., 2023)","reference_id":"MTkwNTc2Nzc1ODY5MjY4MDM1NA=="} +{"index":1381,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.03277","reference_contexts":"Recent open-sourced LLMs (Zhang et al., 2022; Touvron et al., 2023a; Peng et al., 2023) also facilitate various research studies including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , and so on (Gong et al., 2023; Liu et al., 2023b; Ye et al., 2023) .","reference_id":"MTUzNDM3MTI4MzYzMzAyNDg4NTY="} +{"index":1382,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2103.00020","reference_contexts":"For the vision component, two models are popular in our evaluated end-to-end LMMs, i.e., CLIP-ViT/L14 (Radford et al., 2021) (428M) and EVA-ViT-G (1.13B)","reference_id":"MTUyOTQwMzM4NTEzNTYxMjQ5MTc="} +{"index":1383,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2103.00020","reference_contexts":"This superior performance may be attributed to LLaVA's adoption of CLIP (Radford et al., 2021) ViT-L/14 (Dosovitskiy et al., 2020) as its vision model, and the inclusion of a large volume of image-OCR pairings within the training data (Liu et al., 2023f) .","reference_id":"OTQ3Nzk1NDYzODU0MzU1MTEwMQ=="} +{"index":1384,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2111.02114","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1385,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.17580","reference_contexts":"There also exists the exploration (Yang et al., 2022b; Zeng et al., 2022; Yang et al., 2023c; Shen et al., 2023; Gao et al., 2023a) on the modular combination of LLMs and image-to-text vision-language models","reference_id":"OTc3ODYzMTMxOTk2MzA1OTg0NQ=="} +{"index":1386,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.17580","reference_contexts":"We evaluate two sets of multimodal systems, i.e., the end-to-end tuned LMMs including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , etc, and the LLM-tool-using systems (Yang et al., 2023c; Shen et al., 2023; Gao et al., 2023a; Huggingface, 2023) such as MM-ReAct (Yang et al., 2023c) and Transformers Agent (Huggingface, 2023)","reference_id":"OTUwNDQzODQxNzY0ODg5MDEzMQ=="} +{"index":1387,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.17580","reference_contexts":"On the other hand, multimodal agents (Yang et al., 2023c; Shen et al., 2023; Huggingface, 2023; Gao et al., 2023a) explore chaining different vision tools with LLMs (Brown et al., 2020b; OpenAI, 2023c) to achieve integrated visionlanguage capabilities.","reference_id":"NDI0MzE4NjYzNDkwMTU5NzMxMg=="} +{"index":1388,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2302.13971","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1389,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2307.09288","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1390,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2106.13884","reference_contexts":"Since more than 80% of our human being's perception, learning, cognition, and activities are mediated through vision (Politzer) , it is natural to start the exploration by equipping LLMs with \"eyes.\" One main thread of LMM works, represented by Frozen (Tsimpoukelli et al., 2021) , Flamingo (Alayrac et al., 2022) , PaLM-E (Driess et al., 2023) , GPT-4V (OpenAI, 2023c; a) , extend LLMs with the visual understanding capability via end-to-end tuning","reference_id":"MTgwMDEzMTY2ODg3NzA2MTIxMDc="} +{"index":1391,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2106.13884","reference_contexts":"In approaching such multimodal generalist systems, one direction is to extend LLMs with the multi-sensory ability, such as pioneer works Frozen (Tsimpoukelli et al., 2021) , Flamingo (Alayrac et al., 2022) , PaLM-E (Driess et al., 2023) and GPT-4V (OpenAI, 2023c; a)","reference_id":"MTM0NTk5Mzk4Mzk2NDE4Nzk5MDM="} +{"index":1392,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2205.14100","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1393,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2305.11175","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1394,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2306.09265","reference_contexts":"Our MM-Vet is most related to the concurrent evaluation studies (Fu et al., 2023a; Liu et al., 2023d; Li et al., 2023a; Xu et al., 2023; Liu et al., 2023a) such as MME and MM-Bench, which design comprehensive evaluation samples to facilitate the LMM evaluation","reference_id":"MTQwNjg3MDY5MDk1ODIwMTcxNzU="} +{"index":1395,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2310.11441","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1396,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2309.17421","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1397,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2303.11381","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1398,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.14178","reference_contexts":"Recently, thanks to the open-source of powerful LLMs like LLaMA (Touvron et al., 2023a) , more open-sourced LMMs are built, including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , and many more (Gong et al., 2023; Liu et al., 2023b; Ye et al., 2023)","reference_id":"Mjg1OTQ2MDk5MjMyOTM0MzIzOQ=="} +{"index":1399,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.14178","reference_contexts":"Recent open-sourced LLMs (Zhang et al., 2022; Touvron et al., 2023a; Peng et al., 2023) also facilitate various research studies including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , and so on (Gong et al., 2023; Liu et al., 2023b; Ye et al., 2023) .","reference_id":"MTUzNDM3MTI4MzYzMzAyNDg4NTY="} +{"index":1400,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2204.00598","reference_contexts":"There also exists the exploration (Yang et al., 2022b; Zeng et al., 2022; Yang et al., 2023c; Shen et al., 2023; Gao et al., 2023a) on the modular combination of LLMs and image-to-text vision-language models","reference_id":"OTc3ODYzMTMxOTk2MzA1OTg0NQ=="} +{"index":1401,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2307.02469","reference_contexts":"Determining a superior model is currently not possible due to the absence of a comprehensive ablation study (Zeng et al., 2023)","reference_id":"MTc2ODA3MjMxMTAwMjQ1NTE5Mjc="} +{"index":1402,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2205.01068","reference_contexts":"Recent open-sourced LLMs (Zhang et al., 2022; Touvron et al., 2023a; Peng et al., 2023) also facilitate various research studies including OpenFlamingo (Awadalla et al., 2023a) , LLaVA (Liu et al., 2023c) , MiniGPT-4 (Zhu et al., 2023a) , Otter (Li et al., 2023c) , InstructBLIP (Dai et al., 2023) , and so on (Gong et al., 2023; Liu et al., 2023b; Ye et al., 2023) .","reference_id":"MTUzNDM3MTI4MzYzMzAyNDg4NTY="} +{"index":1403,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.10592","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1404,"citer_arxiv_id":"2308.02490","cited_arxiv_id":"2304.06939","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1405,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2005.14165","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1406,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2212.07143","reference_contexts":"We choose Stable Diffusion-v2 (Rombach et al., 2022) as our generative model, and the open-source CLIP model, OpenCLIP-ViT/H (Cherti et al., 2022) for crafting the prompt, as both share the same text encoder","reference_id":"MTcyNDg3OTQ5ODYyMzc1MTI4NjM="} +{"index":1407,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2210.11416","reference_contexts":"This \"instruction tuning\" paradigm has since become a standard way to increase the ability of large models to follow complex, taskspecific instructions (Sanh et al., 2022; Chung et al., 2022)","reference_id":"MTIyMzI2Mzc0MjQ4NzQ1OTY1NjE="} +{"index":1408,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"1602.02830","reference_contexts":"These optimizers consistently result in higher accuracy (Rastegari et al., 2016; Courbariaux et al., 2016) and avoid local minima (Li et al., 2017) .","reference_id":"MTI3OTg3ODUzMDk0NDQyMzkxODA="} +{"index":1409,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2205.12548","reference_contexts":"AutoPrompt (Shin et al., 2020) was one of the first discrete prompt optimization frameworks for transformer language models and subsequent approaches have included a gradient-free phrase editing method (Prasad et al., 2022) , an embedding optimization approach based on Langevin dynamics (Shi et al., 2022) and a reinforcement learning approach (Deng et al., 2022) .","reference_id":"Mjk1ODYyNzM2MDMxNTIxNTAxOA=="} +{"index":1410,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2208.01618","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1411,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2104.08691","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1412,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2201.12086","reference_contexts":"To generate a descriptive hard prompt, this tool first uses a pre-trained captioning model, BLIP (Li et al., 2022) to create a caption of the target image","reference_id":"MTE4NDgzOTM2NjMxNDE0MTM0NjQ="} +{"index":1413,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"1711.05101","reference_contexts":"During the prompt optimization process, we use a generic learning rate of 0.1 and run 3000 optimization steps using the AdamW optimizer (Loshchilov & Hutter, 2017)","reference_id":"NzY0MzY2MjI0MzkxNTExMDYxNg=="} +{"index":1414,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2203.07281","reference_contexts":"AutoPrompt (Shin et al., 2020) was one of the first discrete prompt optimization frameworks for transformer language models and subsequent approaches have included a gradient-free phrase editing method (Prasad et al., 2022) , an embedding optimization approach based on Langevin dynamics (Shi et al., 2022) and a reinforcement learning approach (Deng et al., 2022) .","reference_id":"Mjk1ODYyNzM2MDMxNTIxNTAxOA=="} +{"index":1415,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2210.04610","reference_contexts":"Overall, we suspect that only complete feature-based content detectors have the potential to mitigate these concerns for model owners (Rando et al., 2022) .","reference_id":"MTU5OTIwNzY1MTIxNzIxODUxNTY="} +{"index":1416,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2210.08402","reference_contexts":"We conduct experiments on four datasets with diverse distributions: LAION (Schuhmann et al., 2022) , MS COCO (Lin et al., 2014) , Celeb-A (Liu et al., 2015) , and Lexica.art (Santana, 2022)","reference_id":"MjAyMDA0NTU0OTk3MjQ5MjIwOA=="} +{"index":1417,"citer_arxiv_id":"2302.03668","cited_arxiv_id":"2205.01068","reference_contexts":"We test the transferred text on: GPT-2 XL, T5-LM-XL, OPT-2.7B, and OPT-6B (Radford et al., 2019; Lester et al., 2021b; Zhang et al., 2022) , verifying the reliability of the proposed algorithm over related techniques and testing whether the hard prompt can reliably boost performance","reference_id":"MTU3NzAxOTU1ODMwNzE1OTcyNDg="} +{"index":1418,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1710.11252","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1419,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2103.10951","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1420,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1809.11096","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1421,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1904.10509","reference_contexts":"[6, 28, 32] proposed to use local-wise sparse attention in visual synthesis tasks, which allows the models to see more contexts","reference_id":"MTIyODUzNzg3Mzc2Njg1OTg5NDE="} +{"index":1422,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1906.04341","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1423,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2105.13290","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1424,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2012.09841","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1425,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1912.12180","reference_contexts":"(15) .","reference_id":"OTAxMzIwMDM2MTYyMjc1OTY2Nw=="} +{"index":1426,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1912.12180","reference_contexts":"[15, 33, 45] proposed to use axial-wise sparse attention in visual synthesis tasks, which conducts sparse attention along the axes of visual data representations","reference_id":"MTY3Mjk4NDI3NTU5MDEzMTQzNzU="} +{"index":1427,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1705.06950","reference_contexts":"Text-to-Video (T2V) fine-tuning: We compare N ÜWA on the Kinetics [16] dataset quantitatively in Tab","reference_id":"MTY3OTQ4MDc1ODI3MjkyNDgyMg=="} +{"index":1428,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1429,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1903.01434","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1430,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1804.01523","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1431,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2003.04035","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1432,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1802.05751","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1433,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2103.00020","reference_contexts":"Following DALL-E [33], we use k blurred FID score (FID-k) and Inception Score (IS) [35] to evaluate the quality and variety respectively, and follow-ing GODIVA [45] , we use CLIPSIM metric, which incorporates a CLIP [29] model to calculate the semantic similarity between input text and the generated image","reference_id":"NjAxODgzODUwOTkyMTc1MTc4"} +{"index":1434,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1511.06434","reference_contexts":"Auto-regressive models [33, 39, 41, 45] play an important role in visual synthesis tasks, due to their explicit density modeling and stable training advantages compared with GANs [4, 30, 37, 47]","reference_id":"MzM1NDIzNTgzMjY0MjUzNjUxNg=="} +{"index":1435,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2006.10704","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1436,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1906.05909","reference_contexts":"[6, 28, 32] proposed to use local-wise sparse attention in visual synthesis tasks, which allows the models to see more contexts","reference_id":"MTIyODUzNzg3Mzc2Njg1OTg5NDE="} +{"index":1437,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2102.12092","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1438,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1409.0575","reference_contexts":"We experiment on Im-ageNet [34] and OpenImages [19]","reference_id":"NzQyMjIxMjQzNzkxNTk1ODc1NQ=="} +{"index":1439,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2008.05865","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1440,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1812.01717","reference_contexts":"Although given only one frame as condition (Cond.), N ÜWA still significantly pushes the state-of-the-art FVD [38] score from 94±2 to 86.9.","reference_id":"NDQyNzc2MzgyNTUzMDk1Mjk3NQ=="} +{"index":1441,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1606.05328","reference_contexts":"Auto-regressive models [33, 39, 41, 45] play an important role in visual synthesis tasks, due to their explicit density modeling and stable training advantages compared with GANs [4, 30, 37, 47]","reference_id":"MzM1NDIzNTgzMjY0MjUzNjUxNg=="} +{"index":1442,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1606.05328","reference_contexts":"Earlier visual auto-regressive models, such as PixelCNN [39] , PixelRNN [41] , Image Transformer [28], iGPT [5] , and Video Transformer [44], performed visual synthesis in a \"pixel-by-pixel\" manner","reference_id":"MTM4NDYyODM2NjI4NjU0NjQyMzM="} +{"index":1443,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1606.05328","reference_contexts":"Earlier visual auto-regressive models [5, 28, 39, 41, 44] performed visual synthesis in a \"pixel-by-pixel\" manner","reference_id":"MTEzMDMxNzcxNTg3MTQ5MDYzNTU="} +{"index":1444,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1711.00937","reference_contexts":"Input a raw image I ∈ R H×W ×C with height H, width W and channel C, VQ-VAE [40] trains a learnable codebook to build a bridge between raw continuous pixels and discrete tokens, as de- noted in Eq","reference_id":"MTQxMTAwNDY0NDQ0MDc2NTAzMTk="} +{"index":1445,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1711.00937","reference_contexts":"Recently, with the arise of VQ-VAE [40] as a discrete visual tokenization approach, efficient and large-scale pre-training can be applied to visual synthesis tasks for images (e.g., DALL-E [33] and CogView [9] ) and videos (e.g., GO-DIVA [45] )","reference_id":"MjE5ODcwNDkxMDU4MjA4MzMwOA=="} +{"index":1446,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"1711.00937","reference_contexts":"Recently, VQ-VAE-based [40] visual auto-regressive models were proposed for visual synthesis tasks","reference_id":"MjQyOTQ5MzQzMjA1NTQ5MTEyNQ=="} +{"index":1447,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.14806","reference_contexts":"Auto-regressive models [33, 39, 41, 45] play an important role in visual synthesis tasks, due to their explicit density modeling and stable training advantages compared with GANs [4, 30, 37, 47]","reference_id":"MzM1NDIzNTgzMjY0MjUzNjUxNg=="} +{"index":1448,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.14806","reference_contexts":"This mechanism makes training very efficient and is friendly to large-scale pre-trained models like DALL-E [33], CogView [9], and GODIVA [45]","reference_id":"NTczNjg2ODg4MjA5NDU5NDA5"} +{"index":1449,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.14806","reference_contexts":"By converting images into discrete visual tokens, such methods can conduct efficient and large-scale pre-training for textto-image generation (e.g., DALL-E [33] and CogView [9] ), text-to-video generation (e.g., GODIVA [45] ), and video prediction (e.g., LVT [31] and VideoGPT [48] ), with higher resolution of generated images or videos","reference_id":"MjU0NTc3NTA3MjYxNDIzNzY2OQ=="} +{"index":1450,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.14806","reference_contexts":"[15, 33, 45] proposed to use axial-wise sparse attention in visual synthesis tasks, which conducts sparse attention along the axes of visual data representations","reference_id":"MTY3Mjk4NDI3NTU5MDEzMTQzNzU="} +{"index":1451,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.14806","reference_contexts":"Recently, with the arise of VQ-VAE [40] as a discrete visual tokenization approach, efficient and large-scale pre-training can be applied to visual synthesis tasks for images (e.g., DALL-E [33] and CogView [9] ) and videos (e.g., GO-DIVA [45] )","reference_id":"MjE5ODcwNDkxMDU4MjA4MzMwOA=="} +{"index":1452,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.14806","reference_contexts":"Following DALL-E [33], we use k blurred FID score (FID-k) and Inception Score (IS) [35] to evaluate the quality and variety respectively, and follow-ing GODIVA [45] , we use CLIPSIM metric, which incorporates a CLIP [29] model to calculate the semantic similarity between input text and the generated image","reference_id":"NjAxODgzODUwOTkyMTc1MTc4"} +{"index":1453,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.10157","reference_contexts":"Videos can be viewed as a temporal extension of images, and recent works like VideoGPT [48] and VideoGen [51] extend convolutions in the VQ-VAE encoder from 2D to 3D and train a video-specific representation","reference_id":"MTA3ODk4MjY3OTI5NjUwNzYwNzQ="} +{"index":1454,"citer_arxiv_id":"2111.12417","cited_arxiv_id":"2104.10157","reference_contexts":"By converting images into discrete visual tokens, such methods can conduct efficient and large-scale pre-training for textto-image generation (e.g., DALL-E [33] and CogView [9] ), text-to-video generation (e.g., GODIVA [45] ), and video prediction (e.g., LVT [31] and VideoGPT [48] ), with higher resolution of generated images or videos","reference_id":"MjU0NTc3NTA3MjYxNDIzNzY2OQ=="} +{"index":1455,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1711.00464","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1456,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1705.05263","reference_contexts":"This metric has been shown to correlate with human evaluations (Jiwoong Im et al., 2018) , and if the independent critic is trained on validation data, it can also be used to measure overfitting (Danihelka et al., 2017)","reference_id":"ODQ0MjQ1NTUzNDgxMzg4NzM0NQ=="} +{"index":1457,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1605.08803","reference_contexts":"Powerful explicit posteriors given by Real Non Volume Preserving (RNVP) normalizing flows (Dinh et al., 2016) do not improve marginal distribution matching in VAEs over simple diagonal Gaussian posteriors","reference_id":"OTMxMDA0NTYyMzI5ODA1MTE0Mg=="} +{"index":1458,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1412.6515","reference_contexts":"The density ratio trick (Goodfellow et al., 2014 , Goodfellow, 2014 , Sugiyama et al., 2012) leverages the power of classifiers in order to estimate density ratios","reference_id":"OTE0OTM5OTY2NTc2ODEzNDIyMg=="} +{"index":1459,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1702.06832","reference_contexts":"Previous work has focused on finding adversarial examples as input to the VAE by finding points in data space that the VAE is unable to reconstruct (Kos et al., 2017) .","reference_id":"MTM3NzQ4NTY0MDM1OTA2MjM2MTM="} +{"index":1460,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1511.05644","reference_contexts":"Adversarial Autoencoders (AAE) (Makhzani et al., 2015) ignore the mutual information term in Equation 5 and minimize KL[q η (z)||p(z)] using adversarial training","reference_id":"Mjk0MzUyMDI0NDg4MTkzOTg2"} +{"index":1461,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1511.05644","reference_contexts":"To overcome the limitations of VAEs, it has become natural to consider borrowing the strengths of another popular type of generative algorithm, Generative Adversarial networks (GANs) (Goodfellow et al., 2014) , resulting in a fusion of VAEs and GANs (Larsen et al., 2016 , Makhzani et al., 2015 , Mescheder et al., 2017 , Pu et al., 2017 , Rosca et al., 2017 , Srivastava et al., 2017)","reference_id":"NDMyNTgxNDgzMDI2MTU0Nzk0NQ=="} +{"index":1462,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1610.03483","reference_contexts":"This is a GAN training regime, where the generator loss is given by the reverse KL generator loss (Mohamed and Lakshminarayanan, 2016) :","reference_id":"MTI5ODIzNTEyNjcwOTE5NDkwNjE="} +{"index":1463,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1610.03483","reference_contexts":"By making use of implicit latent variable models (Mohamed and Lakshminarayanan, 2016) GANs do not require observation likelihoods -they define p θ (x) by specifying a deterministic mapping G : z → x used to generate model samples","reference_id":"MTQ1MTU5NTE4Njc5NDEwMDY0MDM="} +{"index":1464,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1612.00005","reference_contexts":"Generative models currently have a wide range of applications in dimensionality reduction, denoising, reinforcement learning, few-shot learning, datasimulation and emulation, semi-supervised learning and in-painting (Higgins et al., 2017b , Kingma et al., 2014 , Mathieu et al., 2016 , Nguyen et al., 2016 , Rezende et al., 2016 , Salimans et al., 2017 , Smith, 2002) , and they continue to be deployed in new domains, from drug-discovery to high-energy physics","reference_id":"NTg1ODc5Mzc5MzQyNDQ5NTYyNA=="} +{"index":1465,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1610.09585","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1466,"citer_arxiv_id":"1802.06847","cited_arxiv_id":"1706.04987","reference_contexts":"To overcome the limitations of VAEs, it has become natural to consider borrowing the strengths of another popular type of generative algorithm, Generative Adversarial networks (GANs) (Goodfellow et al., 2014) , resulting in a fusion of VAEs and GANs (Larsen et al., 2016 , Makhzani et al., 2015 , Mescheder et al., 2017 , Pu et al., 2017 , Rosca et al., 2017 , Srivastava et al., 2017)","reference_id":"NDMyNTgxNDgzMDI2MTU0Nzk0NQ=="} +{"index":1467,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2012.05672","reference_contexts":"It would also be interesting to explore frozen pretrained models for tasks beyond predictive modeling, such as reinforcement learning (Abramson et al., 2020) .","reference_id":"MTA4NDUwMjkzODY4Mjg4NTEyNzI="} +{"index":1468,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2012.05672","reference_contexts":"Furthermore, transformers have also been used for prediction tasks over protein sequences (Jumper et al., 2021; Rao et al., 2021) , reinforcement learning (Parisotto et al., 2020) , and imitation learning (Abramson et al., 2020) .","reference_id":"MzI1OTgyNDc5NTAyNjA1OTI4MA=="} +{"index":1469,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2012.05672","reference_contexts":"The transformer architecture (Vaswani et al., 2017) has shown broad successes in deep learning, serving as the backbone of large models for tasks such as modeling natural language (Brown et al., 2020) , images (Dosovitskiy et al., 2020) , proteins (Jumper et al., 2021) , behaviors (Abramson et al., 2020) , and multimodal tasks comprising of both images and text (Lu et al., 2019; Radford et al., 2021)","reference_id":"MTA3NjM4OTE0NzAwOTA0NzcxNTM="} +{"index":1470,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1711.00482","reference_contexts":"Language might also be a natural choice of model for this blackboard, as there are hypotheses that language may serve as a good multipurpose high-level representation for cognitive behavior and conscious planning (Andreas et al., 2017; Goyal & Bengio, 2020) .","reference_id":"MTE0NTU1Mjc3Njc0ODQ2NDUwNjE="} +{"index":1471,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1910.11856","reference_contexts":"(2021) do a thorough investigation of transfer with language pretraining, notably showing transfer from English to Python, which they consider to be reasonably distanced from English; many works have also looked at transferring from one langauge to another (Artetxe et al., 2019; Ponti et al., 2019)","reference_id":"MzMzMTg5MDAxNDkwNjMwMjQ1OA=="} +{"index":1472,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1607.06450","reference_contexts":"Layer norm (Ba et al., 2016) is frequently used in recurrent and transformer architectures as a means of normalizing the activations","reference_id":"MzU2NDEwNTE5MzEwOTQyMTA1Mw=="} +{"index":1473,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2004.05150","reference_contexts":"We also investigate how other transformer architectures perform when swapped out with GPT-2: BERT (Devlin et al., 2018) , T5 (Raffel et al., 2019) , and Longformer (Beltagy et al., 2020)","reference_id":"NzkzOTExMDc4MTMyNTI5OTU3OA=="} +{"index":1474,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2005.14165","reference_contexts":"The transformer architecture (Vaswani et al., 2017) has shown broad successes in deep learning, serving as the backbone of large models for tasks such as modeling natural language (Brown et al., 2020) , images (Dosovitskiy et al., 2020) , proteins (Jumper et al., 2021) , behaviors (Abramson et al., 2020) , and multimodal tasks comprising of both images and text (Lu et al., 2019; Radford et al., 2021)","reference_id":"MTA3NjM4OTE0NzAwOTA0NzcxNTM="} +{"index":1475,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2005.14165","reference_contexts":"Transformers (Vaswani et al., 2017) were first used successfully for natural language processing (Radford et al., 2018; Devlin et al., 2018; Radford et al., 2019; Brown et al., 2020)","reference_id":"MzMxMTI0Mzk3MjcyMTM5MTk1OA=="} +{"index":1476,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1810.01365","reference_contexts":"These techniques used are similar to prior approaches such as FiLM (Perez et al., 2018) and self-modulation (Chen et al., 2018)","reference_id":"NTk1MTM3ODQ5MTc0NTM0NjI5Mw=="} +{"index":1477,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1511.01432","reference_contexts":"A common trend in deep learning models is to first train a large model on an unsupervised objective on a large dataset (Dai & Le, 2015; Radford et al., 2018) and then finetune on a small downstream dataset (e.g., by freezing the model and only finetuing the output layer)","reference_id":"MTE5NDk5NDc0NTk0MDI4NjQ4NTg="} +{"index":1478,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1810.04805","reference_contexts":"Transformers (Vaswani et al., 2017) were first used successfully for natural language processing (Radford et al., 2018; Devlin et al., 2018; Radford et al., 2019; Brown et al., 2020)","reference_id":"MzMxMTI0Mzk3MjcyMTM5MTk1OA=="} +{"index":1479,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1810.04805","reference_contexts":"We also investigate how other transformer architectures perform when swapped out with GPT-2: BERT (Devlin et al., 2018) , T5 (Raffel et al., 2019) , and Longformer (Beltagy et al., 2020)","reference_id":"NzkzOTExMDc4MTMyNTI5OTU3OA=="} +{"index":1480,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1605.09782","reference_contexts":"A common technique for evaluating the embeddings learned by an unsupervised model is to train a linear layer on top of the embeddings for a downstream task (Donahue et al., 2016; Oord et al., 2018; Chen et al., 2020b) , which is reasonable when you finetune on the same modality as the pretrained one","reference_id":"NDI3NjYxMTM1NjkwNDkxNjI4Mw=="} +{"index":1481,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2010.11929","reference_contexts":"• Image pretraining (ViT): using a pretrained Vision Transformer (Dosovitskiy et al., 2020) pretrained on ImageNet-21k (Deng et al., 2009)","reference_id":"MTI0NDgwNzY1MDcxODU5NTYyMDg="} +{"index":1482,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2010.11929","reference_contexts":"In contrast, transformers utilize selfattention layers to extract features across tokens of a sequence, such as words (Vaswani et al., 2017) or image patches (Dosovitskiy et al., 2020)","reference_id":"NTAxMTkzODA0OTAyNzE5MDYxOQ=="} +{"index":1483,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2010.11929","reference_contexts":"The transformer architecture (Vaswani et al., 2017) has shown broad successes in deep learning, serving as the backbone of large models for tasks such as modeling natural language (Brown et al., 2020) , images (Dosovitskiy et al., 2020) , proteins (Jumper et al., 2021) , behaviors (Abramson et al., 2020) , and multimodal tasks comprising of both images and text (Lu et al., 2019; Radford et al., 2021)","reference_id":"MTA3NjM4OTE0NzAwOTA0NzcxNTM="} +{"index":1484,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2010.11929","reference_contexts":"There are also many works looking at transformers specifically in the context of in-modality transfer, such as ViT for vision (Dosovitskiy et al., 2020) , T5 for language (Raffel et al., 2019) , and UDSMProt for protein sequences (Strodthoff et al., 2020)","reference_id":"NzUzNTIzMDMyMjY1MjQwNzY2"} +{"index":1485,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2003.00152","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1486,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1906.09531","reference_contexts":"Mitigating these biases is an active area of research (Grover et al., 2019; Choi et al., 2020)","reference_id":"MTYzNjk1NzkwMTkyNDI3MDcxMjM="} +{"index":1487,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1606.08415","reference_contexts":"Then, in the rest of the transformer block, a two-layer MLP is used, conventionally projecting the dimension upwards to 4 • n dim for the inner dimension and using the GELU activation function (Hendrycks & Gimpel, 2016)","reference_id":"NzA0NDI5MzAyNDEzNDA2NDM4"} +{"index":1488,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2102.01293","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1489,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1706.05137","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1490,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1909.02950","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1491,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1492,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1908.02265","reference_contexts":"The transformer architecture (Vaswani et al., 2017) has shown broad successes in deep learning, serving as the backbone of large models for tasks such as modeling natural language (Brown et al., 2020) , images (Dosovitskiy et al., 2020) , proteins (Jumper et al., 2021) , behaviors (Abramson et al., 2020) , and multimodal tasks comprising of both images and text (Lu et al., 2019; Radford et al., 2021)","reference_id":"MTA3NjM4OTE0NzAwOTA0NzcxNTM="} +{"index":1493,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1908.02265","reference_contexts":"Recent work has utilized transformers for multimodal predictive tasks, such as images and text in ViLBERT (Lu et al., 2019) and CLIP (Radford et al., 2021) ; these approaches generally use two distinct transformers to embed images and text","reference_id":"MTAzOTc4Mjk2MjQxODcwMzc3MTA="} +{"index":1494,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1807.03748","reference_contexts":"A common technique for evaluating the embeddings learned by an unsupervised model is to train a linear layer on top of the embeddings for a downstream task (Donahue et al., 2016; Oord et al., 2018; Chen et al., 2020b) , which is reasonable when you finetune on the same modality as the pretrained one","reference_id":"NDI3NjYxMTM1NjkwNDkxNjI4Mw=="} +{"index":1495,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2004.14601","reference_contexts":"Similar to our work, Papadimitriou & Jurafsky (2020) investigate transfer for LSTMs between modalities including code, different languages, and music, finding that pretraining on \"non-linguistic data with latent structure\" can transfer to language, finding grammatical structure in a modality to be important, although we generally investigate the other direction and explore more distanced modalities","reference_id":"MTgyMDY1NTI2OTgxMDA0MDQwNjg="} +{"index":1496,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1912.01703","reference_contexts":"We train all models using the Adam (Kingma & Ba, 2014) optimizer following Pytorch (Paszke et al., 2019) defaults","reference_id":"MTUyMDg0MDExNzkyMjI0MDU0MDk="} +{"index":1497,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1910.10683","reference_contexts":"There are also many works looking at transformers specifically in the context of in-modality transfer, such as ViT for vision (Dosovitskiy et al., 2020) , T5 for language (Raffel et al., 2019) , and UDSMProt for protein sequences (Strodthoff et al., 2020)","reference_id":"NzUzNTIzMDMyMjY1MjQwNzY2"} +{"index":1498,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1910.10683","reference_contexts":"We also investigate how other transformer architectures perform when swapped out with GPT-2: BERT (Devlin et al., 2018) , T5 (Raffel et al., 2019) , and Longformer (Beltagy et al., 2020)","reference_id":"NzkzOTExMDc4MTMyNTI5OTU3OA=="} +{"index":1499,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2008.02217","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1500,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1705.08045","reference_contexts":"• Layer norm parameters: as is standard practice in other finetuning works (Rebuffi et al., 2017; Houlsby et al., 2019) , we also finetune the affine layer norm parameters (scale and bias), which adapt to the statistics of the downstream task in a new domain","reference_id":"MTgyMzE0NzY0MTk0MzcwODk2NzQ="} +{"index":1501,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1705.08045","reference_contexts":"A common method used to finetune transformers are adapter networks (Rebuffi et al., 2017; Houlsby et al., 2019) , which add a fully connected residual block for each unique downstream task and also finetune the layer norm parameters","reference_id":"NTE1MDAwNDQyMzAyNTQ5NTI="} +{"index":1502,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1909.01326","reference_contexts":"For high stakes applications in the real-world, there are potential concerns with transfer of harmful biases from one modality to one another using pretrained transformer models trained on vast quantities of unlabeled, uncurated datasets (Sheng et al., 2019; Bender et al., 2021)","reference_id":"MTM1MjY1MDc4MzczNTY5NzEzMjA="} +{"index":1503,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2010.15980","reference_contexts":"A recent direction of research has explored learning prompt templates for large models (Shin et al., 2020) that simply require forward passes over the transformer","reference_id":"MTQ5Mjg1MDI5NzU3MDg4MDY5OTY="} +{"index":1504,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2011.04006","reference_contexts":"This is a modified version of the above task taken from the Long Range Arena benchmark where the images are converted to grayscale and flattened with a token length of 1 (Tay et al., 2020)","reference_id":"MTEyOTI3ODY1MTY3Njg0MTIzOTk="} +{"index":1505,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2011.04006","reference_contexts":"Frozen Pretrained Transformer Full Transformer Full LSTM Figure 1 : A frozen language-pretrained transformer (FPT) -without finetuning the self-attention and feedforward layers -can achieve strong performance compared to a transformer fully trained from scratch on a downstream modality on benchmarks from literature (Tay et al., 2020; Rao et al., 2019)","reference_id":"MTE0MjU1NTIxMjYzOTMxNjg1Ng=="} +{"index":1506,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"2012.12877","reference_contexts":"One particular modality of interest is computer vision (Chen et al., 2020a; Touvron et al., 2020) ; in particular, Dosovitskiy et al","reference_id":"MTUyMTUxNTI0NDgyMTcyMzQwNTg="} +{"index":1507,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1706.03762","reference_contexts":"In contrast, transformers utilize selfattention layers to extract features across tokens of a sequence, such as words (Vaswani et al., 2017) or image patches (Dosovitskiy et al., 2020)","reference_id":"NTAxMTkzODA0OTAyNzE5MDYxOQ=="} +{"index":1508,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1706.03762","reference_contexts":"The transformer architecture (Vaswani et al., 2017) has shown broad successes in deep learning, serving as the backbone of large models for tasks such as modeling natural language (Brown et al., 2020) , images (Dosovitskiy et al., 2020) , proteins (Jumper et al., 2021) , behaviors (Abramson et al., 2020) , and multimodal tasks comprising of both images and text (Lu et al., 2019; Radford et al., 2021)","reference_id":"MTA3NjM4OTE0NzAwOTA0NzcxNTM="} +{"index":1509,"citer_arxiv_id":"2103.05247","cited_arxiv_id":"1706.03762","reference_contexts":"Transformers (Vaswani et al., 2017) were first used successfully for natural language processing (Radford et al., 2018; Devlin et al., 2018; Radford et al., 2019; Brown et al., 2020)","reference_id":"MzMxMTI0Mzk3MjcyMTM5MTk1OA=="} +{"index":1510,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2108.02938","reference_contexts":"More recently, denoising diffusion models were used to solve inverse problems in both supervised (i.e., degradation model is known during training) [42, 41, 13, 10, 54] and unsupervised settings [22, 26, 25, 21, 46, 47, 9]","reference_id":"MzM0MDQ4MTA2NzExNjkxNTUyMg=="} +{"index":1511,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2108.02938","reference_contexts":"DDPM [19] ) that can predict x given x t+1 and t + 1 for t ∈ [0, T -1], and a noiseless measurement y = Hx, where H is a downscaling matrix, the Iterative Latent Variable Refinement (ILVR) [9] algorithm can sample from the posterior distribution p","reference_id":"MTgxNTIzMDc0NTQwMDQ0MTE2MTU="} +{"index":1512,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2108.02938","reference_contexts":"ILVR [9] suggests a diffusion-based method that handles noiseless super-resolution, and can run in 250 steps","reference_id":"OTc2NTY5MTkzMzUxMDA3MjcwOA=="} +{"index":1513,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2112.05146","reference_contexts":"More recently, denoising diffusion models were used to solve inverse problems in both supervised (i.e., degradation model is known during training) [42, 41, 13, 10, 54] and unsupervised settings [22, 26, 25, 21, 46, 47, 9]","reference_id":"MzM0MDQ4MTA2NzExNjkxNTUyMg=="} +{"index":1514,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2112.05146","reference_contexts":"We provide additional figures below showing DDRM's versatility across different datasets, inverse problems, and noise levels ( Figures 7, 8, 10, 11, and 12)","reference_id":"MTM3MjUxOTMzMDc0MjgyODU1NzE="} +{"index":1515,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"1312.6114","reference_contexts":"Here, we will use a simple property of the KL divergence between univariate Gaussians [27] :","reference_id":"MTMxNTI4NDMxNjgxMjI4OTA1Mjg="} +{"index":1516,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2103.04715","reference_contexts":"A similar use of denoisers in different iterative algorithms is proposed in [34, 16, 30]","reference_id":"MTMxMzY0MTM5MzM0MzUyMDEzMzU="} +{"index":1517,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2103.04715","reference_contexts":"An extreme example is found in [30] where a \"fast\" version of the algorithm uses 15, 000 neural function evaluations (NFEs).","reference_id":"MTMxODI5NDM3OTY2Mzk4MTM2MTQ="} +{"index":1518,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2102.09672","reference_contexts":"It is possible to learn the variances [35] or consider alternative constructions where Proposition 3.1 holds; we leave these options as future work.","reference_id":"MTMxNjUxNjc0NDIzNDQ4MDM2ODE="} +{"index":1519,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2111.05826","reference_contexts":"More recently, denoising diffusion models were used to solve inverse problems in both supervised (i.e., degradation model is known during training) [42, 41, 13, 10, 54] and unsupervised settings [22, 26, 25, 21, 46, 47, 9]","reference_id":"MzM0MDQ4MTA2NzExNjkxNTUyMg=="} +{"index":1520,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2111.05826","reference_contexts":"For a specific degradation model, image restoration can be addressed through end-to-end supervised training of neural networks, using pairs of original and degraded images [14, 58, 41]","reference_id":"MTY5OTcwODM0ODY3NDI0ODM3Njg="} +{"index":1521,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2111.05826","reference_contexts":"In terms of future work, apart from further optimizing the timestep and variance schedules, it would be interesting to investigate the following: (i) applying DDRM to non-linear inverse problems, (ii) addressing scenarios where the degradation operator is unknown, and (iii) self-supervised training techniques inspired by DDRM as well as ones used in supervised techniques [41] that further improve performance of unsupervised models for image restoration.","reference_id":"NzA3NzU5ODM0Mzc2OTY1MzIxNA=="} +{"index":1522,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2104.07636","reference_contexts":"More recently, denoising diffusion models were used to solve inverse problems in both supervised (i.e., degradation model is known during training) [42, 41, 13, 10, 54] and unsupervised settings [22, 26, 25, 21, 46, 47, 9]","reference_id":"MzM0MDQ4MTA2NzExNjkxNTUyMg=="} +{"index":1523,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"1906.09453","reference_contexts":"Under this unsupervised setting, priors based on deep neural networks have demonstrated impressive empirical results in various image restoration tasks [40, 50, 43, 38, 15]","reference_id":"NDIwODI2OTg1MTI5ODQzMDIyNQ=="} +{"index":1524,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"1906.09453","reference_contexts":"The authors of [43] leverages robust classifiers learned with additional class labels.","reference_id":"MTQ5OTE4OTUwMDc4ODU1OTIwODc="} +{"index":1525,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"1503.03585","reference_contexts":"which leads to an evidence lower bound (ELBO) on the maximum likelihood objective [44]","reference_id":"MTc4MDI3MjcxNzE4NjQyNzQ2MTE="} +{"index":1526,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"1503.03585","reference_contexts":"DDRM is a denoising diffusion generative model [44, 19, 45] that gradually and stochastically denoises a sample to the desired output, conditioned on the measurements and the inverse problem","reference_id":"NDU4ODcyODAwMjY4NzYzNzM2OQ=="} +{"index":1527,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2111.08005","reference_contexts":"This setup is inherently general to all linear inverse problems, a property desired in many real-world applications such as medical imaging [46, 20]","reference_id":"MjMzNjI1NjQxMzE1OTYxMzcxNg=="} +{"index":1528,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2111.08005","reference_contexts":"More recently, denoising diffusion models were used to solve inverse problems in both supervised (i.e., degradation model is known during training) [42, 41, 13, 10, 54] and unsupervised settings [22, 26, 25, 21, 46, 47, 9]","reference_id":"MzM0MDQ4MTA2NzExNjkxNTUyMg=="} +{"index":1529,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2111.08005","reference_contexts":"However, real-world applications such as medical imaging often require flexibility to cope with multiple, possibly infinite, degradation models [46]","reference_id":"MTAzMTA1NzUzMDU2MjAyMjYyMDA="} +{"index":1530,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"2112.02475","reference_contexts":"More recently, denoising diffusion models were used to solve inverse problems in both supervised (i.e., degradation model is known during training) [42, 41, 13, 10, 54] and unsupervised settings [22, 26, 25, 21, 46, 47, 9]","reference_id":"MzM0MDQ4MTA2NzExNjkxNTUyMg=="} +{"index":1531,"citer_arxiv_id":"2201.11793","cited_arxiv_id":"1506.03365","reference_contexts":"We demonstrate our algorithm's capabilities using the diffusion models from [19] , which are trained on CelebA-HQ [23] , LSUN bedrooms, and LSUN cats [56] (all 256 × 256 pixels)","reference_id":"ODU1MTY2MDMxMTg2Mzg5Njk2MA=="} +{"index":1532,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1312.6114","reference_contexts":"Variational Auto-encoders (VAEs) [2, 22] are perhaps the most influential AE based models","reference_id":"Mjg1NzM2MjMyNTg3MTg1NTAyMw=="} +{"index":1533,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1410.8516","reference_contexts":"Normalizing flows [3, 4, 12] model p(x) as an invertible transformation from a simple distribution through a change of variables","reference_id":"NTI5MTcwNjc2NDU5NDI1MjY0MA=="} +{"index":1534,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1605.08803","reference_contexts":"GANs have exhibited impressive performance in generating high quality images [13] and other vision tasks [14, 15]","reference_id":"MTU0NTAyNjQ2ODgxOTkwOTg4NTU="} +{"index":1535,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1605.08803","reference_contexts":"Based on [4] , we adopt the following layers to form the flows used in our model.","reference_id":"NTg5MTA1MTIxNDg3MTU3Mjc2NA=="} +{"index":1536,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1707.05776","reference_contexts":"GLANN [7] learns a latent representation by GLO [5] and matches the densities of the latent variables with an implicit maximum likelihood estimator [6]","reference_id":"OTc0NjM0NzYxMDIwOTEwNzQ5MQ=="} +{"index":1537,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1809.09087","reference_contexts":"GLANN [7] learns a latent representation by GLO [5] and matches the densities of the latent variables with an implicit maximum likelihood estimator [6]","reference_id":"OTc0NjM0NzYxMDIwOTEwNzQ5MQ=="} +{"index":1538,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1903.05789","reference_contexts":"We also trained VAEs+flow prior with different choices of decoder variances (equivalently, different choices of β), plus one with learnable decoder variance as done in [10]","reference_id":"OTUxOTcyODU5NDQ4MjEzOTk5Mw=="} +{"index":1539,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1903.05789","reference_contexts":"For example, it is observed that VAEs tend to generate blurry images, an effect that is usually attributed to the failure to match the marginal distribution in the latent space [10, 46]","reference_id":"MTU0Mzg1NzQ5MDUyMzMwMzc2MzQ="} +{"index":1540,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1903.05789","reference_contexts":"Two-stage-VAE [10] introduces another VAE on the latent space defined by the first VAE to learn the distribution of its latent variables","reference_id":"MTM1MDE0MDk0NjgyOTYxNDI0NzA="} +{"index":1541,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1903.05789","reference_contexts":"In particular, empirically VAEs with flow posterior have been shown to improve neither the matching of q(z) and p(z) [46] , nor the generation quality [10]","reference_id":"MzgyODIxMTcyNjY0NzIwMzU0Ng=="} +{"index":1542,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1903.05789","reference_contexts":"We adopt the settings in the original paper [10]","reference_id":"MTY1Mjk0NzU3NzQwODU5NDQ3Mw=="} +{"index":1543,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1903.05789","reference_contexts":"For CelebA, one particular detail worth noting is that, [10] applies 128 × 128 center-crop before re-sizing on CelebA, while 160 × 160 center-crop is used in our evaluations","reference_id":"MTY4OTMzOTg5MTQxNzc4OTU3MTg="} +{"index":1544,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1511.06434","reference_contexts":"Implicit models are dominated by Generative Adversarial Networks (GANs) [1, 11]","reference_id":"MjAzNjEzOTY0ODI1OTMxNzM2MA=="} +{"index":1545,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1511.06434","reference_contexts":"For CelebA-HQ dataset, we adopt our AE network structure based on DCGAN [11]","reference_id":"MTAxMzcwOTY0NDA2MzgwODI0Mw=="} +{"index":1546,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1809.11096","reference_contexts":"GANs have exhibited impressive performance in generating high quality images [13] and other vision tasks [14, 15]","reference_id":"MTU0NTAyNjQ2ODgxOTkwOTg4NTU="} +{"index":1547,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1701.00160","reference_contexts":"It is well known that training GANs is unstable and extremely sensitive to hyper-parameter settings [16, 56] , and sometimes training leads to mode collapse [17] , where most of the samples share some common properties","reference_id":"NDMxOTk0NjI0NzE0NDE1OTA0NQ=="} +{"index":1548,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1802.05957","reference_contexts":"Although there have been multiple efforts to overcome the difficulties in training GANs, by modifying the objective functions or introducing normalization [20, 21, 18, 19] , researchers are also actively studying non-adversarial methods that are known to be less affected by these issues","reference_id":"NzIyMzQ3NzAxMjg2MjA4MjI3NQ=="} +{"index":1549,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1701.07875","reference_contexts":"Although there have been multiple efforts to overcome the difficulties in training GANs, by modifying the objective functions or introducing normalization [20, 21, 18, 19] , researchers are also actively studying non-adversarial methods that are known to be less affected by these issues","reference_id":"NzIyMzQ3NzAxMjg2MjA4MjI3NQ=="} +{"index":1550,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1611.02163","reference_contexts":"Although there have been multiple efforts to overcome the difficulties in training GANs, by modifying the objective functions or introducing normalization [20, 21, 18, 19] , researchers are also actively studying non-adversarial methods that are known to be less affected by these issues","reference_id":"NzIyMzQ3NzAxMjg2MjA4MjI3NQ=="} +{"index":1551,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1401.4082","reference_contexts":"Variational Auto-encoders (VAEs) [2, 22] are perhaps the most influential AE based models","reference_id":"Mjg1NzM2MjMyNTg3MTg1NTAyMw=="} +{"index":1552,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1511.05644","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1553,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1409.1556","reference_contexts":"The perceptual loss is obtained by feeding both training images and reconstructed images into a pre-trained network such as VGG [26] , and computing the L 1 distance between some of the intermediate layers' activation","reference_id":"NTY5NzQ5Mjc4MzIzMTY3NTk5OA=="} +{"index":1554,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1711.01558","reference_contexts":"Adversarial auto-encoders [52] and Wasserstein auto-encoders [30] use adversarial regularizer or MMD regularizers [55] to force the q(z) to be close to p(z)","reference_id":"NDgwNTQ3MTczOTI5Mzg1OTA1OQ=="} +{"index":1555,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1711.01558","reference_contexts":"We follow the settings introduced in the original WAE paper [30]","reference_id":"MTAwNDc3NTA4MzM5MTg4NTYxODY="} +{"index":1556,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1711.01558","reference_contexts":"Some other modifications have recently been proposed to better match the distribution of latent variables and the prior distribution [10, 30, 28, 7] , and they are shown to have the potential to generate high quality samples","reference_id":"NTkxNjIxMjYxMTUxNzU3OTYxOA=="} +{"index":1557,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1505.05770","reference_contexts":"There are also attempts to use normalizing flows as VAEs' posterior distributions [31, 32, 33]","reference_id":"NjE4MzYwOTQ4NDEzMzQ5MjkxMA=="} +{"index":1558,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1803.05649","reference_contexts":"There are also attempts to use normalizing flows as VAEs' posterior distributions [31, 32, 33]","reference_id":"NjE4MzYwOTQ4NDEzMzQ5MjkxMA=="} +{"index":1559,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1509.00519","reference_contexts":"Some modifications to VAEs [34, 31] improve the estimated test data likelihood","reference_id":"MTc2NTY5NjQ1MjcyOTg0MzkwNzQ="} +{"index":1560,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1708.07747","reference_contexts":"To demonstrate the performance of our method, we present both quantitative and qualitative evaluations on four commonly used datasets for generative models: MNIST [35] , Fashion MNIST [36] , CIFAR-10 [37] and CelebA [8]","reference_id":"MTI3NzMyNTcwNjM2NjczOTk4NzY="} +{"index":1561,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1711.10337","reference_contexts":"In Table 4 we combine our reported results of AE based models and the FID scores of GANs cited from [38] .","reference_id":"MTE1MTI5MzQ3MzU3NTk1NzUyOTA="} +{"index":1562,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1711.10337","reference_contexts":"For CelebA, we follow the same pre-processing as in [38] : center crop to 160 × 160 and then resize to 64 × 64.","reference_id":"MjkyNTQ3NTQ2MjY3OTM3ODQwOQ=="} +{"index":1563,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1711.10337","reference_contexts":"To compare our method with GANs, we also include the results from [38] in Appendix C","reference_id":"NjY2NDM5MTM3NzE0ODA1NzM1MQ=="} +{"index":1564,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1511.01844","reference_contexts":"However, it is known that higher likelihood is not directly related to better sample quality [40, 57]","reference_id":"OTc0MTE5NTA3NTA3MjQ2MTEwNw=="} +{"index":1565,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1706.08500","reference_contexts":"We use the Fréchet Inception Distance (FID) [41] as a metric for image generation quality","reference_id":"MTAyMzAyOTgzODAzNjEzMjY1MTk="} +{"index":1566,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1412.6980","reference_contexts":"For all datasets (except CelebA-HQ), we use batch size 256 and Adam [43] optimizer with initial learning rate 10 -3 for the parameters of both the AE and the flow","reference_id":"MzI4NDk3NzQ4NTgyNDY4MTI2OQ=="} +{"index":1567,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1601.06759","reference_contexts":"For example, auto-regressive models [44, 45] , which assume the data distribution can be expressed in an auto-regressive pattern, have a simple and stable training process, and currently give the best likelihood results; however, they cannot provide low-dimensional representations of images, and their sampling procedure is inefficient","reference_id":"NzUwMzc3MTg0MDUyNTUzNTkwOQ=="} +{"index":1568,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1802.06847","reference_contexts":"For example, it is observed that VAEs tend to generate blurry images, an effect that is usually attributed to the failure to match the marginal distribution in the latent space [10, 46]","reference_id":"MTU0Mzg1NzQ5MDUyMzMwMzc2MzQ="} +{"index":1569,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1802.06847","reference_contexts":"In particular, empirically VAEs with flow posterior have been shown to improve neither the matching of q(z) and p(z) [46] , nor the generation quality [10]","reference_id":"MzgyODIxMTcyNjY0NzIwMzU0Ng=="} +{"index":1570,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1705.07120","reference_contexts":"[48, 49, 29] propose different ways to approximate the aggregated posterior q(z) during training, and use the approximated q(z) as their VAEs' prior distributions","reference_id":"MTM5NzExMTUxNTkxMjU2ODEyMQ=="} +{"index":1571,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1802.05983","reference_contexts":"Several modifications to VAE's objective [47, 50, 51] , which are designed for the task of unsupervised disentanglement, further decompose the KL term in ELBO, put a stronger penalty specifically on the mismatch between q(z) and p(z)","reference_id":"Njc4NDI5NjI1NTY4NDk0MDMxNw=="} +{"index":1572,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1511.05644","reference_contexts":"Adversarial auto-encoders [52] and Wasserstein auto-encoders [30] use adversarial regularizer or MMD regularizers [55] to force the q(z) to be close to p(z)","reference_id":"NDgwNTQ3MTczOTI5Mzg1OTA1OQ=="} +{"index":1573,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1710.02248","reference_contexts":"VAEs with flow prior have attracted very little attention [54] , and they have only focused on improvement of the data likelihood","reference_id":"MTQ1ODIzMTQ2MzE1NDUzODI4MzY="} +{"index":1574,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1705.08868","reference_contexts":"However, it is known that higher likelihood is not directly related to better sample quality [40, 57]","reference_id":"OTc0MTE5NTA3NTA3MjQ2MTEwNw=="} +{"index":1575,"citer_arxiv_id":"1905.10485","cited_arxiv_id":"1710.10196","reference_contexts":"Besides samples of the datasets used for quantitative evaluation, samples of CelebA-HQ [59] with the larger size of 256 × 256 are also included in Figure 2e to show our method's potential of scaling up to images with higher resolution","reference_id":"MTYxMjg4OTI4OTc1MDg2MDA2MjE="} +{"index":1576,"citer_arxiv_id":"2009.09941","cited_arxiv_id":"1805.06085","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1577,"citer_arxiv_id":"2009.09941","cited_arxiv_id":"1708.04552","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1578,"citer_arxiv_id":"2009.09941","cited_arxiv_id":"1608.08710","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1579,"citer_arxiv_id":"2009.09941","cited_arxiv_id":"1710.09412","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1580,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1508.05326","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1581,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"2005.14165","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1582,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1909.05858","reference_contexts":"Pre-trained language models (PLMs) have substantially advanced the state-of-the-art across a variety of natural language processing (NLP) tasks (Peters et al., 2018; Devlin et al., 2019; Yang et al., 2019; Radford et al., 2019; Liu et al., 2019; Keskar et al., 2019; Shoeybi et al., 2019)","reference_id":"OTc4MDQ0OTAyMDE4NTg0MTQ3NA=="} +{"index":1583,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1909.05858","reference_contexts":"Connecting BERT & GPT-2 Two technical questions remain, when pre-training OPTIMUS from BERT & GPT-2: (i) How to represent sentences, since the two PLMs employ different tokenization schemes? (ii) How to adapt a pre-trained GPT-2 to arbitrary conditional input without retraining the model again? Controllable GPT-2 models have been studied in (Keskar et al., 2019; Zellers et al., 2019; Peng et al., 2020a,b) when prescribed control codes/tokens are provided, but it is still unknown how to ground GPT-2 to arbitrary conditional inputs.","reference_id":"MzU4NjgxNDEwMDE1OTY5Mjk4Mg=="} +{"index":1584,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1909.05858","reference_contexts":"Large-scale Transformer-based PLMs have recently achieved state-of-the-art performance on various natural language understanding and generation tasks (Devlin et al., 2019; Yang et al., 2019; Radford et al., 2019; Liu et al., 2019; Keskar et al., 2019)","reference_id":"MTI4MjQwNzY4NTk1NDc2NzA1NQ=="} +{"index":1585,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1910.13461","reference_contexts":"In a bid to combine language understanding and generation tasks in one unified framework, several model variants have been proposed, including UniLM (Dong et al., 2019) , BART (Lewis et al., 2019) , and T5 (Raffel et al., 2019)","reference_id":"MTYzOTU2NTEzMzU5NzA1NTY4MjU="} +{"index":1586,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1701.06547","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1587,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1710.03957","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1588,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1907.11692","reference_contexts":"Pre-trained language models (PLMs) have substantially advanced the state-of-the-art across a variety of natural language processing (NLP) tasks (Peters et al., 2018; Devlin et al., 2019; Yang et al., 2019; Radford et al., 2019; Liu et al., 2019; Keskar et al., 2019; Shoeybi et al., 2019)","reference_id":"OTc4MDQ0OTAyMDE4NTg0MTQ3NA=="} +{"index":1589,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1907.11692","reference_contexts":"Large-scale Transformer-based PLMs have recently achieved state-of-the-art performance on various natural language understanding and generation tasks (Devlin et al., 2019; Yang et al., 2019; Radford et al., 2019; Liu et al., 2019; Keskar et al., 2019)","reference_id":"MTI4MjQwNzY4NTk1NDc2NzA1NQ=="} +{"index":1590,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1411.1784","reference_contexts":"A conditional GAN (Mirza and Osindero, 2014) is trained on the fixed latent space","reference_id":"NTg1NDM5NTgwMzI1ODIxMjMx"} +{"index":1591,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1802.05365","reference_contexts":"Pre-trained language models (PLMs) have substantially advanced the state-of-the-art across a variety of natural language processing (NLP) tasks (Peters et al., 2018; Devlin et al., 2019; Yang et al., 2019; Radford et al., 2019; Liu et al., 2019; Keskar et al., 2019; Shoeybi et al., 2019)","reference_id":"OTc4MDQ0OTAyMDE4NTg0MTQ3NA=="} +{"index":1592,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1910.10683","reference_contexts":"In a bid to combine language understanding and generation tasks in one unified framework, several model variants have been proposed, including UniLM (Dong et al., 2019) , BART (Lewis et al., 2019) , and T5 (Raffel et al., 2019)","reference_id":"MTYzOTU2NTEzMzU5NzA1NTY4MjU="} +{"index":1593,"citer_arxiv_id":"2004.04092","cited_arxiv_id":"1909.08053","reference_contexts":"Pre-trained language models (PLMs) have substantially advanced the state-of-the-art across a variety of natural language processing (NLP) tasks (Peters et al., 2018; Devlin et al., 2019; Yang et al., 2019; Radford et al., 2019; Liu et al., 2019; Keskar et al., 2019; Shoeybi et al., 2019)","reference_id":"OTc4MDQ0OTAyMDE4NTg0MTQ3NA=="} +{"index":1594,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1908.03557","reference_contexts":"Two such prominent datasets -SBU [1] and Conceptual Captions [2] -are widely used for pre-training vision-and-language (V&L) representations [3] [4] [5] [6] [7] [8] [9] [10] [11] that transfer to a variety of downstream V&L tasks like visual question answering [12] [13] [14] , visual reasoning [15, 16] , and image captioning [17, 18]","reference_id":"MjEzMTc4MzYzMTYzMjUzMTMzNg=="} +{"index":1595,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1908.03557","reference_contexts":"Originally intended for image-text retrieval and image captioning, they are now widely used for training generic V&L representations [3] [4] [5] [6] [7] [8] [9] [10] [11] 92 ] that transfer to downstream tasks like visual question answering [12] [13] [14] , referring expressions [93] , and visual reasoning [15, 16]","reference_id":"MTQwMzM0NTMxMDU4Mzc1MTE2Njc="} +{"index":1596,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1909.11740","reference_contexts":"Two such prominent datasets -SBU [1] and Conceptual Captions [2] -are widely used for pre-training vision-and-language (V&L) representations [3] [4] [5] [6] [7] [8] [9] [10] [11] that transfer to a variety of downstream V&L tasks like visual question answering [12] [13] [14] , visual reasoning [15, 16] , and image captioning [17, 18]","reference_id":"MjEzMTc4MzYzMTYzMjUzMTMzNg=="} +{"index":1597,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1909.11740","reference_contexts":"Originally intended for image-text retrieval and image captioning, they are now widely used for training generic V&L representations [3] [4] [5] [6] [7] [8] [9] [10] [11] 92 ] that transfer to downstream tasks like visual question answering [12] [13] [14] , referring expressions [93] , and visual reasoning [15, 16]","reference_id":"MTQwMzM0NTMxMDU4Mzc1MTE2Njc="} +{"index":1598,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2004.00849","reference_contexts":"Two such prominent datasets -SBU [1] and Conceptual Captions [2] -are widely used for pre-training vision-and-language (V&L) representations [3] [4] [5] [6] [7] [8] [9] [10] [11] that transfer to a variety of downstream V&L tasks like visual question answering [12] [13] [14] , visual reasoning [15, 16] , and image captioning [17, 18]","reference_id":"MjEzMTc4MzYzMTYzMjUzMTMzNg=="} +{"index":1599,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2004.00849","reference_contexts":"Originally intended for image-text retrieval and image captioning, they are now widely used for training generic V&L representations [3] [4] [5] [6] [7] [8] [9] [10] [11] 92 ] that transfer to downstream tasks like visual question answering [12] [13] [14] , referring expressions [93] , and visual reasoning [15, 16]","reference_id":"MTQwMzM0NTMxMDU4Mzc1MTE2Njc="} +{"index":1600,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1504.00325","reference_contexts":"The base model from [19] used a ResNet-50 [21] visual backbone, and Transformers [64] in textual head that are L = 1 layers deep and H = 2048 dimensions wide, and was trained on COCO Captions [17] (118K images)","reference_id":"MTU3Mzg2MTg5MTYwMzQ0MjE4MjE="} +{"index":1601,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1504.00325","reference_contexts":"[78] trained visual N-gram models on YFCC-100M [95]; [19, 20] learn features from COCO Captions [17] that are competitive with supervised ImageNet training [21, 96] on many downstream tasks [22, 24, [26] [27] [28] , and [29, 30] scale up to very larger non-public datasets that are larger than RedCaps.","reference_id":"MTEzMjE5ODU1MjQzMTgwNDc1MTc="} +{"index":1602,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1504.00325","reference_contexts":"Two such prominent datasets -SBU [1] and Conceptual Captions [2] -are widely used for pre-training vision-and-language (V&L) representations [3] [4] [5] [6] [7] [8] [9] [10] [11] that transfer to a variety of downstream V&L tasks like visual question answering [12] [13] [14] , visual reasoning [15, 16] , and image captioning [17, 18]","reference_id":"MjEzMTc4MzYzMTYzMjUzMTMzNg=="} +{"index":1603,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1504.00325","reference_contexts":"Recent work [19, 20] also shows that image-text data from COCO [17] can be used to learn visual features that are competitive with supervised pretraining [21] on ImageNet [22, 23] when transfered to downstream tasks [24] [25] [26] [27] [28]","reference_id":"NDk3NjEzMzM3MDE3NjQyOTI3Nw=="} +{"index":1604,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"We lowercase captions and use ftfy [44] to remove character accents, emojis, and non-latin characters, following [29, 35, 36]","reference_id":"MTc4NjQ5Mjk0NTg1ODY3ODU0MDU="} +{"index":1605,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"Zero-shot image classification: Training with language supervision enables zero-shot transfer to downstream tasks without any task-specific training [29, 78]","reference_id":"ODAxNjA3NjE5MjE5NDgwMjA1Nw=="} +{"index":1606,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"[78] trained visual N-gram models on YFCC-100M [95]; [19, 20] learn features from COCO Captions [17] that are competitive with supervised ImageNet training [21, 96] on many downstream tasks [22, 24, [26] [27] [28] , and [29, 30] scale up to very larger non-public datasets that are larger than RedCaps.","reference_id":"MTEzMjE5ODU1MjQzMTgwNDc1MTc="} +{"index":1607,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"We evaluate the utility of different datasets for representation learning by comparing zero-shot performance on seven classification datasets: Oxford-IIIT Pets [79] , Food-101 [80] , Flowers-102 [81] , Stanford Cars [82] , Country-211 [29] , and SUN-397 [83] , and Birdsnap [84]","reference_id":"MTI5MzQ0Njk2MzM0MzI0MTg4NTI="} +{"index":1608,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"2 0 1 1 2 0 1 2 2 0 1 3 2 0 1 4 2 0 1 5 2 0 1 6 2 0 1 7 2 0 1 8 2 0 1 9 2 0 2 0 3M 6M 9M 12M 1.0M 860K 3.3M 2.9M 4.4M 6.1M 8.7M 12.0M RedCaps SBU Captions Conceptual Captions Datasets in 2021 # Instances Released RedCaps (ours) 12,011,111 CC-12M [31] 12,423,374 WIT-english [62] 5,500,746 CLIP [29] 400M × ALIGN [30] 1.8B × 10K 100K 1M pics 1.9M mildlyinteresting 731K cats 643K food 393K itookapicture 327K sneakers 314K earthporn 262K gardening 208K foodporn 202K succulents 201K houseplants 182K mechanicalkeyboards 156K rarepuppers 150K aquariums 139K crochet 125K dogpictures 120K baking 119K woodworking 112K rabbits 105K carporn 102K Figure 4: Instances per subreddit: Top 20 subreddits with most image-text pairs in RedCaps","reference_id":"NjM4NDk1Nzc3NjQ3MTQ4NTY5Mw=="} +{"index":1609,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"Our evaluation setup closely follows recent works on self-supervised learning [75] [76] [77] and language-supervised [19, 29] learning","reference_id":"MTU0MDE5NjY4Mzk2NDMwODczNjA="} +{"index":1610,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"While CLIP [29] and ALIGN [30] used orders of magnitude larger training datasets, they are not released for public use -RedCaps remains one of the largest public image-text datasets.","reference_id":"NzIyODMyOTg5NjQzNjg2MDUzMA=="} +{"index":1611,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"More recently, CLIP [29] and ALIGN [30] scale up to 400M and 1B+ web-curated image-text pairs, enabling zero-shot visual recognition.","reference_id":"MTU2Mjc5NTA2MzI1ODM0MzE4ODI="} +{"index":1612,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.00020","reference_contexts":"We also transfer the learned features to eleven different downstream datasets for tasks including image classification, object detection, instance segmentation, and fine-grained recognition using both fine-tuning and languagebased zero-shot classification [29]","reference_id":"MTQxMTA2OTMyNzYyOTA4ODMwNzQ="} +{"index":1613,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2102.05918","reference_contexts":"More recently, CLIP [29] and ALIGN [30] scale up to 400M and 1B+ web-curated image-text pairs, enabling zero-shot visual recognition.","reference_id":"MTU2Mjc5NTA2MzI1ODM0MzE4ODI="} +{"index":1614,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2102.05918","reference_contexts":"2 0 1 1 2 0 1 2 2 0 1 3 2 0 1 4 2 0 1 5 2 0 1 6 2 0 1 7 2 0 1 8 2 0 1 9 2 0 2 0 3M 6M 9M 12M 1.0M 860K 3.3M 2.9M 4.4M 6.1M 8.7M 12.0M RedCaps SBU Captions Conceptual Captions Datasets in 2021 # Instances Released RedCaps (ours) 12,011,111 CC-12M [31] 12,423,374 WIT-english [62] 5,500,746 CLIP [29] 400M × ALIGN [30] 1.8B × 10K 100K 1M pics 1.9M mildlyinteresting 731K cats 643K food 393K itookapicture 327K sneakers 314K earthporn 262K gardening 208K foodporn 202K succulents 201K houseplants 182K mechanicalkeyboards 156K rarepuppers 150K aquariums 139K crochet 125K dogpictures 120K baking 119K woodworking 112K rabbits 105K carporn 102K Figure 4: Instances per subreddit: Top 20 subreddits with most image-text pairs in RedCaps","reference_id":"NjM4NDk1Nzc3NjQ3MTQ4NTY5Mw=="} +{"index":1615,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2102.05918","reference_contexts":"While CLIP [29] and ALIGN [30] used orders of magnitude larger training datasets, they are not released for public use -RedCaps remains one of the largest public image-text datasets.","reference_id":"NzIyODMyOTg5NjQzNjg2MDUzMA=="} +{"index":1616,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2102.05918","reference_contexts":"[78] trained visual N-gram models on YFCC-100M [95]; [19, 20] learn features from COCO Captions [17] that are competitive with supervised ImageNet training [21, 96] on many downstream tasks [22, 24, [26] [27] [28] , and [29, 30] scale up to very larger non-public datasets that are larger than RedCaps.","reference_id":"MTEzMjE5ODU1MjQzMTgwNDc1MTc="} +{"index":1617,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2102.05918","reference_contexts":"\"alt img\" [30] )","reference_id":"MzEwMjc0NTU3Mjg5MzI4ODAyNg=="} +{"index":1618,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2005.14165","reference_contexts":"Textual data from Reddit is already used for pre-training massive language models [33] [34] [35] [36] in NLP","reference_id":"MTQ5NjU3NTcyNTI5Njk2NTA2NTY="} +{"index":1619,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2005.14165","reference_contexts":"We lowercase captions and use ftfy [44] to remove character accents, emojis, and non-latin characters, following [29, 35, 36]","reference_id":"MTc4NjQ5Mjk0NTg1ODY3ODU0MDU="} +{"index":1620,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1606.00372","reference_contexts":"Prior works in dialog modeling and text summarization have trained on Reddit comments [33, [37] [38] [39] [40]","reference_id":"MTYwNzYyMjI4NTE1MjQ3NDkwMjU="} +{"index":1621,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2001.08435","reference_contexts":"Image post filtering: We use Pushshift [41] and Reddit [42, 43] APIs to download all image posts submitted to our selected subreddits from 2008-2020","reference_id":"MTQ3NzU0NDI2MzU1NzQ3NTE1NzY="} +{"index":1622,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1803.09010","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1623,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2012.05345","reference_contexts":"Harmful Stereotypes: Another concern with Reddit data is that images or language may represent harmful stereotypes about gender, race, or other characteristics of people [48, 49, 51]","reference_id":"MTIyMzU5MzQ5NjQ4OTkyODI0MjA="} +{"index":1624,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.06191","reference_contexts":"(Filtered) 5K (%) 50K 12M Privacy: The individual who posts a given photo on Reddit may not be the person appearing in said photo; this can pose privacy risks for people who did not expect to appear in images online [49, 50]","reference_id":"NjA0MjAxMjk5MDA3MzMxNzM3MA=="} +{"index":1625,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2009.11462","reference_contexts":"This stands in contrast to less curated uses of Reddit data, such as GPT-2 [35] whose training data includes at least 63K documents from banned or quarantined subreddits which may contain toxic language [53]","reference_id":"MTcyNjQ0NDM0MTU1NTc2MzA2MTA="} +{"index":1626,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1707.09457","reference_contexts":"Subtler issues may also exist, such as imbalanced representation of demographic groups [59] or gender bias in object co-occurrence [60] or language [61]","reference_id":"ODgyNzAzNzQ0MjQ2MTIxMzcxMg=="} +{"index":1627,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.01913","reference_contexts":"LAIT [94] , Conceptual-12M [31] , and Wikipedia-ImageText [62]","reference_id":"MTgyOTc0MDY1NzI2OTU4NTAz"} +{"index":1628,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.01913","reference_contexts":"RedCaps is 2× larger than the English subset of multilingual Wikipedia image-text dataset [62] , and nearly as large as CC-12M [31]","reference_id":"NjY4OTEzMTE4MzA5ODI1NTQzMg=="} +{"index":1629,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2103.01913","reference_contexts":"2 0 1 1 2 0 1 2 2 0 1 3 2 0 1 4 2 0 1 5 2 0 1 6 2 0 1 7 2 0 1 8 2 0 1 9 2 0 2 0 3M 6M 9M 12M 1.0M 860K 3.3M 2.9M 4.4M 6.1M 8.7M 12.0M RedCaps SBU Captions Conceptual Captions Datasets in 2021 # Instances Released RedCaps (ours) 12,011,111 CC-12M [31] 12,423,374 WIT-english [62] 5,500,746 CLIP [29] 400M × ALIGN [30] 1.8B × 10K 100K 1M pics 1.9M mildlyinteresting 731K cats 643K food 393K itookapicture 327K sneakers 314K earthporn 262K gardening 208K foodporn 202K succulents 201K houseplants 182K mechanicalkeyboards 156K rarepuppers 150K aquariums 139K crochet 125K dogpictures 120K baking 119K woodworking 112K rabbits 105K carporn 102K Figure 4: Instances per subreddit: Top 20 subreddits with most image-text pairs in RedCaps","reference_id":"NjM4NDk1Nzc3NjQ3MTQ4NTY5Mw=="} +{"index":1630,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1607.06450","reference_contexts":"We use the recent Pre-LN Transformer variant [35, 65, 66] that is more stable to train large transformers [67] -LayerNorm [68] is moved inside the residual connection, and we add LayerNorm before the prediction layer.","reference_id":"NTYwNjE3NTYzOTU2MzU4ODEzOA=="} +{"index":1631,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1608.03983","reference_contexts":"-Training details: We use AdamW [72, 73] with weight decay 10 -2 and max learning rate 5 × 10 -4 with linear warmup for the first 10K iterations, followed by cosine decay [74] to zero","reference_id":"MTI4MDA3Njc2OTQxNDcxNzg2MzM="} +{"index":1632,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"2001.07966","reference_contexts":"LAIT [94] , Conceptual-12M [31] , and Wikipedia-ImageText [62]","reference_id":"MTgyOTc0MDY1NzI2OTU4NTAz"} +{"index":1633,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1708.02862","reference_contexts":"Related to this goal are efforts that learn from large quantities of noisy non-text labels for web images such as WebVision [97] , YFCC-100M [95], JFT-300M [98, 99] , and Instagram-3.5B [100] .","reference_id":"MjU2NDE4MTU4NDA3NzkwODIyOQ=="} +{"index":1634,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1503.02531","reference_contexts":"Related to this goal are efforts that learn from large quantities of noisy non-text labels for web images such as WebVision [97] , YFCC-100M [95], JFT-300M [98, 99] , and Instagram-3.5B [100] .","reference_id":"MjU2NDE4MTU4NDA3NzkwODIyOQ=="} +{"index":1635,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1610.02357","reference_contexts":"Related to this goal are efforts that learn from large quantities of noisy non-text labels for web images such as WebVision [97] , YFCC-100M [95], JFT-300M [98, 99] , and Instagram-3.5B [100] .","reference_id":"MjU2NDE4MTU4NDA3NzkwODIyOQ=="} +{"index":1636,"citer_arxiv_id":"2111.11431","cited_arxiv_id":"1910.04867","reference_contexts":"We perform end-to-end fine-tuning of pre-trained weights and follow the same training schedule for every dataset, highly similar to VTAB [102]","reference_id":"NDYyMzQ1OTc2NzMxNjQ3MzE3Mg=="} +{"index":1637,"citer_arxiv_id":"2106.02081","cited_arxiv_id":"2106.01357","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1638,"citer_arxiv_id":"2106.02081","cited_arxiv_id":"1312.6114","reference_contexts":"The coupling of φ, φ in these objective does not directly admit the application of modern methods in density estimation since it does not allow us to freely parametrise MLE estimators for the boundaries thus neural density estimators such as (Kingma and Welling 2013; Papamakarios, Pavlakou, and Murray 2017; Papamakarios 2019 ) cannot be taken advantage of to circumvent the computation of the partition function","reference_id":"MTI1NzMwMTM4NTMzODEyNTAzMTI="} +{"index":1639,"citer_arxiv_id":"2106.02081","cited_arxiv_id":"2107.00630","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1640,"citer_arxiv_id":"2106.02081","cited_arxiv_id":"1308.0215","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1641,"citer_arxiv_id":"2106.02081","cited_arxiv_id":"1705.07057","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1642,"citer_arxiv_id":"2106.02081","cited_arxiv_id":"2011.13456","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1643,"citer_arxiv_id":"2106.02081","cited_arxiv_id":"2106.10410","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1644,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1810.02419","reference_contexts":"We find that conditioning on the labels significantly eases the transformer modeling task and boosts the generation quality, improving IS from 57.63 to 79.28, which is close to the upper bound shown in the \"Real data\" row [1, 28]","reference_id":"NDgzMjczOTc0NzM2ODgxMDkwMw=="} +{"index":1645,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1810.02419","reference_contexts":"ProgressiveV-GAN [1] extends ProgressiveGAN [30] to video generation by simultaneously generating in the temporal direction progressively","reference_id":"MTMzNzI2MjMzNzY2OTQ2OTgzNjg="} +{"index":1646,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2206.03429","reference_contexts":"[5] uses a low-resolution long video generator and short-video super-resolution network to generate videos of dynamic scenes.","reference_id":"NzI1MTExMTA5MzIzNDM4MDg2Nw=="} +{"index":1647,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2106.02719","reference_contexts":"In addition, without further modifications, the length of videos generated by these models is limited by the GPU memory (e.g., at most 140 frames can be generated on 32GB V100 GPU by HVG [8] )","reference_id":"MTczMjQxNTAzMDM4MDY4NTk0MzU="} +{"index":1648,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2106.02719","reference_contexts":"One common strategy to improve the long-term capacity is to use a hierarchical model [16, 8] to reduce the chance of drifting away from the target theme","reference_id":"NjY0MjkxNjg5MTUyMTQ1NDM5NQ=="} +{"index":1649,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2106.02719","reference_contexts":"HVG [8] further introduces a hierarchical pipeline to interpolate and upsample low-resolution and low-frame-rate videos gradually","reference_id":"MTQ2ODExOTE3OTQ5NzAwMDQ1OTc="} +{"index":1650,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1907.06571","reference_contexts":"DVD-GAN [11] adopts a similar architecture as MoCoGAN with a focus on scaling up training","reference_id":"MTYyNTAyNjIyODIwMjQxNTkxNDY="} +{"index":1651,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1907.06571","reference_contexts":"GAN-based methods [70, 56, 11, 41] can generate plausible short videos, but extending them to longer videos requires prohibitively high memory and time cost of training and inference","reference_id":"NDM5MTQyMTAzNTA3ODUwODM5OQ=="} +{"index":1652,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1907.06571","reference_contexts":"Both 3D deconvolutionals [70] and additional RNN or LSTM [56, 67, 67, 11, 57, 46, 65] have been used","reference_id":"MzMxNjQ4NDM5MzY3MzE3ODQxOQ=="} +{"index":1653,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1907.06571","reference_contexts":"Our model is trained on short video clips, e.g., 16 frames, like the previous methods [70, 56, 67, 11, 57, 65]","reference_id":"MTMyOTUyODEwNTE1MDAxOTY4NzM="} +{"index":1654,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1907.06571","reference_contexts":"Similar to other GAN-based video generation models [67, 11] , we use two types of discriminators in our model -a spatial discriminator f Ds that takes in random reconstructed frames xi ∈ R H×W ×3 to encourage frame quality and a temporal discriminator f Dt that takes in the entire reconstruction x ∈ R T ×H×W ×3 to penalize implausible motions:","reference_id":"NDc3NDc1NTI1MDc5MjI4MTQ2"} +{"index":1655,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1907.06571","reference_contexts":"In addition, we follow the methods evaluated on UCF-101 [11, 37] to report the Inception Score (IS) [57] calculated by a trained C3D model [66]","reference_id":"MTc4NjQwMTY5OTcyOTg3ODI0MDc="} +{"index":1656,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1907.06571","reference_contexts":"The former has been the focus of many existing studies [70, 56, 11, 65, 80] working with tens of frames","reference_id":"MjkyODQ4ODA5OTA1MzcxNDQyNg=="} +{"index":1657,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2005.00341","reference_contexts":"We do not use the random start trick for codebook embeddings [14, 80]","reference_id":"MTY3NTUxNDQ2ODAwODMxNzM3OTY="} +{"index":1658,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2204.03458","reference_contexts":"TATS-base improves on IS by 76.2% over previous method with synthetic initial frames [37] and is competitive against concurrent works [61, 24, 22]","reference_id":"MTMxMTAwOTY0NTIyMTg0OTAyNDA="} +{"index":1659,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2204.03458","reference_contexts":"In parallel to our work, [22] proposes a gradient conditioning method for sampling longer videos with a diffusion model, [47] and [24] explore a hierarchical model with frame-level VQVAE embeddings or RGB frames","reference_id":"MTI5OTkxNjU5MDYzMzE1MTIxNTY="} +{"index":1660,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2205.15868","reference_contexts":"TATS-base improves on IS by 76.2% over previous method with synthetic initial frames [37] and is competitive against concurrent works [61, 24, 22]","reference_id":"MTMxMTAwOTY0NTIyMTg0OTAyNDA="} +{"index":1661,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2205.15868","reference_contexts":"In parallel to our work, [22] proposes a gradient conditioning method for sampling longer videos with a diffusion model, [47] and [24] explore a hierarchical model with frame-level VQVAE embeddings or RGB frames","reference_id":"MTI5OTkxNjU5MDYzMzE1MTIxNTY="} +{"index":1662,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2107.04589","reference_contexts":"First, we find that the axial-attention layers introduced in VideoGPT interact poorly with the GAN losses and exacerbate the collapse, which is also noticed in recent work on training ViT with GAN losses [38]","reference_id":"OTkxODcyNDE0MDQ4OTY5MzQyOA=="} +{"index":1663,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2003.04035","reference_contexts":"GAN-based methods [70, 56, 11, 41] can generate plausible short videos, but extending them to longer videos requires prohibitively high memory and time cost of training and inference","reference_id":"NDM5MTQyMTAzNTA3ODUwODM5OQ=="} +{"index":1664,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2203.09494","reference_contexts":"In parallel to our work, [22] proposes a gradient conditioning method for sampling longer videos with a diffusion model, [47] and [24] explore a hierarchical model with frame-level VQVAE embeddings or RGB frames","reference_id":"MTI5OTkxNjU5MDYzMzE1MTIxNTY="} +{"index":1665,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2006.10704","reference_contexts":"For example, RNNs and LSTMs generate temporal noise vectors for an image generator [56, 67, 11, 57, 46, 65] ; transformers either directly generate pixel values [29, 74] or indirectly predict latent tokens [48, 15, 53, 76, 37, 80]","reference_id":"NzM3NzcxMTU1MTYzMDIxMTcwMQ=="} +{"index":1666,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2102.12092","reference_contexts":"The VQVAE framework has been shown effective in conditional image generation [15, 54]","reference_id":"Nzg2NDgyODk5NjM4MTI0NDQ1OA=="} +{"index":1667,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1412.6604","reference_contexts":"Video prediction aims at modeling the transformation between frames and predicting future frames given real frames [55, 63, 17, 69, 36, 49]","reference_id":"NzY4OTA2OTMwMTk0ODY5MTkzMQ=="} +{"index":1668,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1409.1556","reference_contexts":"where f (i) D s/t /VGG denotes the i th layer of either a trained VGG network [59] or discriminators with a scaling factor p i","reference_id":"NDc5OTIwMzI3ODUyNDkyOTE2"} +{"index":1669,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2112.14683","reference_contexts":"Recent works use implicit neural representations to reduce cost by directly mapping temporal positions to either pixels [81] or StyleGAN [33] feature map values [61]","reference_id":"MTY4OTIwMDY3MDY2NDU3MzY0ODU="} +{"index":1670,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2112.14683","reference_contexts":"Implicit Neural Representations [60,64] represent continuous signals such as videos by mapping the coordinate space to RGB space [81, 61]","reference_id":"ODY4NTg3MDU5OTYxMDc1ODQyMA=="} +{"index":1671,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2112.14683","reference_contexts":"StyleGAN-v [61] maps a continuous temporal positional embedding to the input feature map of the StyleGAN.","reference_id":"MTMyODc2Njk3MzkyMTUxNTQzOTI="} +{"index":1672,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2112.14683","reference_contexts":"After removing all the padding, one needs to pad real frames to both ends of the input videos to obtain the desired output size [31, 61]","reference_id":"MTgwMjkxNDk5Mjg2ODY3NjUyMDA="} +{"index":1673,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2112.14683","reference_contexts":"TATS-base improves on IS by 76.2% over previous method with synthetic initial frames [37] and is competitive against concurrent works [61, 24, 22]","reference_id":"MTMxMTAwOTY0NTIyMTg0OTAyNDA="} +{"index":1674,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1212.0402","reference_contexts":"Our model achieves state-of-the-art short and long video generation results on the UCF-101 [62] , Sky Time-lapse [78] , Taichi-HD [58] , and AudioSet-Drum [18] datasets","reference_id":"MTAwMzE0Njg5NzEyNDIwNjEyODA="} +{"index":1675,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1212.0402","reference_contexts":"We validate our approach on the UCF-101 [62] , Sky Time-lapse [78] , Taichi-HD [58] , AudioSet-Drum [18] , and MUGEN [21] datasets","reference_id":"MTUzMjA0NzE0MDIzODk1Mjk2MA=="} +{"index":1676,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1212.0402","reference_contexts":"We show results on UCF-101 [62] , Sky Time-lapse [78] , and Taichi-HD [58] for unconditional or class-conditioned video generation with 128 × 128 resolution following [81] , AudioSet-Drum [18] for audio-conditioned video generation with 64 × 64 resolution following [37] , and MUGEN [21] with 256 × 256 resolution for text-conditioned video generation","reference_id":"Mzk3ODQ2NTcyODUyNzg1MjU2Mg=="} +{"index":1677,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1701.08435","reference_contexts":"Video prediction aims at modeling the transformation between frames and predicting future frames given real frames [55, 63, 17, 69, 36, 49]","reference_id":"NzY4OTA2OTMwMTk0ODY5MTkzMQ=="} +{"index":1678,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"1701.08435","reference_contexts":"For instance, [69] divides the frames into patches, calculates the affine transformation between temporally adjacent patches, and predicts such affine transformation to be applied to the most recent frame","reference_id":"OTIzMDQ2NTg2MTk4OTA4NzA0Ng=="} +{"index":1679,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.14806","reference_contexts":"We hence consider several kinds of conditional information that provide additional temporal information such as audio [9] , text [75] , and so on","reference_id":"MTE2NzM3MjIyMDkyNjM4MDU2MDg="} +{"index":1680,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.14806","reference_contexts":"GODIVA [75] relies on a pretrained VQVAE model to produce video frame representations and autoregressively predicts the video representations based on the input text representations and former frames","reference_id":"Mjc2MzE4NjIyNDE3NDA4MjIyNw=="} +{"index":1681,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2111.12417","reference_contexts":"For example, RNNs and LSTMs generate temporal noise vectors for an image generator [56, 67, 11, 57, 46, 65] ; transformers either directly generate pixel values [29, 74] or indirectly predict latent tokens [48, 15, 53, 76, 37, 80]","reference_id":"NzM3NzcxMTU1MTYzMDIxMTcwMQ=="} +{"index":1682,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.10157","reference_contexts":"We do not use the random start trick for codebook embeddings [14, 80]","reference_id":"MTY3NTUxNDQ2ODAwODMxNzM3OTY="} +{"index":1683,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.10157","reference_contexts":"compared with VQGAN [15] and VideoGPT [80]","reference_id":"NjA0ODM4NTE3ODcyNzcwMjQxNA=="} +{"index":1684,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.10157","reference_contexts":"For example, RNNs and LSTMs generate temporal noise vectors for an image generator [56, 67, 11, 57, 46, 65] ; transformers either directly generate pixel values [29, 74] or indirectly predict latent tokens [48, 15, 53, 76, 37, 80]","reference_id":"NzM3NzcxMTU1MTYzMDIxMTcwMQ=="} +{"index":1685,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.10157","reference_contexts":"VideoGPT [80] further leverages 3D convolution and axial self-attention in the VQVAE to encode the temporal dimension as well","reference_id":"MTM4MjQ5NzMxMDAzNzY4OTY5MDg="} +{"index":1686,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.10157","reference_contexts":"VideoGPT [80] ) or 3D-VQGAN fails to preserve the video quality beyond the training length, as shown in Figure 2","reference_id":"MTM0NjE2NjcyMTg1NzA0MzcwMzY="} +{"index":1687,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.10157","reference_contexts":"The former has been the focus of many existing studies [70, 56, 11, 65, 80] working with tens of frames","reference_id":"MjkyODQ4ODA5OTA1MzcxNDQyNg=="} +{"index":1688,"citer_arxiv_id":"2204.03638","cited_arxiv_id":"2104.10157","reference_contexts":"Directly applying GAN losses to VideoGPT [80] or 3D VQGAN [15] leads to training stability issues","reference_id":"NzQ4OTg4Njc1NzI2MTY5NjI1Nw=="} +{"index":1689,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2302.08113","reference_contexts":"However, despite their remarkable capabilities in synthesizing realistic images consistent with text prompts, most diffusion models usually struggle to accurately follow some complex prompts (Feng et al., 2022; Lian et al., 2023; Liu et al., 2022; Bar-Tal et al., 2023) , which require the model to compose objects with different attributes and relationships into a single image","reference_id":"MTQzODc0NDQyMzExOTgwNTQ3NzI="} +{"index":1690,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.08465","reference_contexts":"Despite some improvements achieved by these methods, there are still two main limitations in the context of compositional/complex image generation: (i) existing layoutbased or attention-based methods can only provide rough and suboptimal spatial guidance, and struggle to deal with overlapped objects (Cao et al., 2023; Hertz et al., 2022; Lian et al., 2023) ; (ii) feedback-based methods require to collect high-quality feedback and incur additional training costs.","reference_id":"MTYwNjExNDcxODEzOTc2NTc5NjU="} +{"index":1691,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.08465","reference_contexts":"Nevertheless, such manipulation-based methods can only make rough controls, and often lead to unsatisfied compositional generation results, especially when dealing with overlapped objects (Lian et al., 2023; Cao et al., 2023)","reference_id":"OTA1NTcyNzk4OTY1NDE2MzY0Mg=="} +{"index":1692,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.08465","reference_contexts":"Compared to traditional methods that utilize cross-attention map swap or replacement (Hertz et al., 2022; Cao et al., 2023) for editing, our mask-and-inpainting method powered by CoT planning enables more accurate and complex editing operations (i.e., addition, deletion and modification).","reference_id":"MTA2MDgxNjA0NTYwOTgzNDc1NDk="} +{"index":1693,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.08465","reference_contexts":"Qualitative Results In the qualitative comparison of textguided image editing, we choose some strong baseline methods, including Prompt2Prompt (Hertz et al., 2022) , InstructPix2Pix (Brooks et al., 2023) and MasaCtrl (Cao et al., 2023)","reference_id":"MTc0OTQ3MzI0MjQ5NTk0NjQwMTU="} +{"index":1694,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.08465","reference_contexts":"We outperform previous powerful methods including Prompt2Prompt (Hertz et al., 2022) , InstructPix2Pix (Brooks et al., 2023) and MasaCtrl (Cao et al., 2023) .","reference_id":"MTUwOTU0ODU5MDk4MTU4NzAyNQ=="} +{"index":1695,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.08465","reference_contexts":"Training-free methods aim to steer diffusion models through manipulating latent or cross-attention maps according to spatial or semantic constraints during inference stages (Feng et al., 2022; Liu et al., 2022; Hertz et al., 2022; Cao et al., 2023; Chen et al., 2024; Chefer et al., 2023)","reference_id":"MzUyMjMzODgxMDEyMDEwOTYyMA=="} +{"index":1696,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2310.00426","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1697,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2210.11416","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1698,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2309.11499","reference_contexts":"Further, Multimodal Large language Models (MLLMs), (Koh et al., 2023; Yu et al., 2023; Sun et al., 2023b; Dong et al., 2023; Fu et al., 2023; Pan et al., 2023; Wu et al., 2023a; Zou et al., 2023; Yang et al., 2023d; Gupta & Kembhavi, 2023; Surís et al., 2023) integrate LLMs with vision models to extend their impressive abilities from language tasks to vision tasks, including image understanding, reasoning and synthesis","reference_id":"MTU3MzkzMTI2NzkyMDE2MTQzOTQ="} +{"index":1699,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.15393","reference_contexts":"The collaboration between LLMs (ChatGPT, 2022; OpenAI, 2023) and diffusion models (Ramesh et al., 2022; Betker et al., 2023) can significantly improve the text-image alignment as well as the quality of generated images (Yu et al., 2023; Chen et al., 2023b; Dong et al., 2023; Wu et al., 2023b; Feng et al., 2023; Pan et al., 2023)","reference_id":"MTUwMzkwMTEyNzMxNzQ1ODM2MDA="} +{"index":1700,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2309.17102","reference_contexts":"Further, Multimodal Large language Models (MLLMs), (Koh et al., 2023; Yu et al., 2023; Sun et al., 2023b; Dong et al., 2023; Fu et al., 2023; Pan et al., 2023; Wu et al., 2023a; Zou et al., 2023; Yang et al., 2023d; Gupta & Kembhavi, 2023; Surís et al., 2023) integrate LLMs with vision models to extend their impressive abilities from language tasks to vision tasks, including image understanding, reasoning and synthesis","reference_id":"MTU3MzkzMTI2NzkyMDE2MTQzOTQ="} +{"index":1701,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2208.01626","reference_contexts":"Despite some improvements achieved by these methods, there are still two main limitations in the context of compositional/complex image generation: (i) existing layoutbased or attention-based methods can only provide rough and suboptimal spatial guidance, and struggle to deal with overlapped objects (Cao et al., 2023; Hertz et al., 2022; Lian et al., 2023) ; (ii) feedback-based methods require to collect high-quality feedback and incur additional training costs.","reference_id":"MTYwNjExNDcxODEzOTc2NTc5NjU="} +{"index":1702,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2208.01626","reference_contexts":"Compared to traditional methods that utilize cross-attention map swap or replacement (Hertz et al., 2022; Cao et al., 2023) for editing, our mask-and-inpainting method powered by CoT planning enables more accurate and complex editing operations (i.e., addition, deletion and modification).","reference_id":"MTA2MDgxNjA0NTYwOTgzNDc1NDk="} +{"index":1703,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2208.01626","reference_contexts":"Qualitative Results In the qualitative comparison of textguided image editing, we choose some strong baseline methods, including Prompt2Prompt (Hertz et al., 2022) , InstructPix2Pix (Brooks et al., 2023) and MasaCtrl (Cao et al., 2023)","reference_id":"MTc0OTQ3MzI0MjQ5NTk0NjQwMTU="} +{"index":1704,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2208.01626","reference_contexts":"We outperform previous powerful methods including Prompt2Prompt (Hertz et al., 2022) , InstructPix2Pix (Brooks et al., 2023) and MasaCtrl (Cao et al., 2023) .","reference_id":"MTUwOTU0ODU5MDk4MTU4NzAyNQ=="} +{"index":1705,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2208.01626","reference_contexts":"Training-free methods aim to steer diffusion models through manipulating latent or cross-attention maps according to spatial or semantic constraints during inference stages (Feng et al., 2022; Liu et al., 2022; Hertz et al., 2022; Cao et al., 2023; Chen et al., 2024; Chefer et al., 2023)","reference_id":"MzUyMjMzODgxMDEyMDEwOTYyMA=="} +{"index":1706,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2307.06350","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1707,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2302.09778","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1708,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2212.12017","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1709,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.02643","reference_contexts":"We locate and mask the target contour associated with the editing instruction (Kirillov et al., 2023) , and apply diffusion-based inpainting (Rombach et al., 2022) to edit the target contour region according to the planned operation list Plan(y tar , x)","reference_id":"MTA3NTgwMTYwMDQyNTA2NDIz"} +{"index":1710,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.17216","reference_contexts":"Further, Multimodal Large language Models (MLLMs), (Koh et al., 2023; Yu et al., 2023; Sun et al., 2023b; Dong et al., 2023; Fu et al., 2023; Pan et al., 2023; Wu et al., 2023a; Zou et al., 2023; Yang et al., 2023d; Gupta & Kembhavi, 2023; Surís et al., 2023) integrate LLMs with vision models to extend their impressive abilities from language tasks to vision tasks, including image understanding, reasoning and synthesis","reference_id":"MTU3MzkzMTI2NzkyMDE2MTQzOTQ="} +{"index":1711,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.17216","reference_contexts":"For instance, GILL (Koh et al., 2023) can condition on arbitrarily interleaved image and text inputs to synthesize coherent image outputs, and Emu (Sun et al., 2023b) stands out as a generalist multimodal interface for both image-to-text and text-to-image tasks","reference_id":"MTM0NDMzMjY4ODk4Mjg5NDIwNDk="} +{"index":1712,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2301.12597","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1713,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2308.10253","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1714,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13655","reference_contexts":"Despite some improvements achieved by these methods, there are still two main limitations in the context of compositional/complex image generation: (i) existing layoutbased or attention-based methods can only provide rough and suboptimal spatial guidance, and struggle to deal with overlapped objects (Cao et al., 2023; Hertz et al., 2022; Lian et al., 2023) ; (ii) feedback-based methods require to collect high-quality feedback and incur additional training costs.","reference_id":"MTYwNjExNDcxODEzOTc2NTc5NjU="} +{"index":1715,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13655","reference_contexts":"Nevertheless, such manipulation-based methods can only make rough controls, and often lead to unsatisfied compositional generation results, especially when dealing with overlapped objects (Lian et al., 2023; Cao et al., 2023)","reference_id":"OTA1NTcyNzk4OTY1NDE2MzY0Mg=="} +{"index":1716,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13655","reference_contexts":"Some works introduce additional layouts/boxes (Li et al., 2023b; Xie et al., 2023; Yang et al., 2023e; Qu et al., 2023; Chen et al., 2024; Wu et al., 2023b; Lian et al., 2023) as conditions or leveraging prompt-aware attention guidance (Feng et al., 2022; Chefer et al., 2023; Wang et al., 2023) to improve compositional text-to-image synthesis","reference_id":"NDQzNDcwMTc0NTQ0NTY3MzQ1"} +{"index":1717,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13655","reference_contexts":"We compare with previous SOTA text-toimage models DALL-E 3 (Betker et al., 2023) , SDXL and LMD+ (Lian et al., 2023) in three main compositional scenarios: (i) Attribute Binding","reference_id":"MzM5ODM4NDU4NjAzNTIyOTgw"} +{"index":1718,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13655","reference_contexts":"Recently, LMD (Lian et al., 2023) utilizes LLMs to enhance the compositional generation of diffusion models by generating images grounded on bounding box layouts from the LLM (Li et al., 2023b)","reference_id":"MTc0Nzg3MTY4MDQxNTE3MjQ5OQ=="} +{"index":1719,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13655","reference_contexts":"However, despite their remarkable capabilities in synthesizing realistic images consistent with text prompts, most diffusion models usually struggle to accurately follow some complex prompts (Feng et al., 2022; Lian et al., 2023; Liu et al., 2022; Bar-Tal et al., 2023) , which require the model to compose objects with different attributes and relationships into a single image","reference_id":"MTQzODc0NDQyMzExOTgwNTQ3NzI="} +{"index":1720,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13655","reference_contexts":"Compared to LMD+ (Lian et al., 2023) , a LLMgrounded layout-based text-to-image diffusion model, our RPG demonstrates both enhanced semantic expression capabilities and image fidelity","reference_id":"MTYzNzg3NTg4MjU0NzYyMTQwMDY="} +{"index":1721,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.08485","reference_contexts":"Leading examples like ChatGPT (ChatGPT, 2022) have showcased the advanced language comprehension and reasoning skills through techniques such as instruction tuning (Ouyang et al., 2022; Li et al., 2023c; Zhang et al., 2023c; Liu et al., 2023)","reference_id":"NTkyMjA1MjUxNjA3MTM5ODM0OQ=="} +{"index":1722,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2302.08453","reference_contexts":"Some approaches mainly introduce additional modules into diffusion models in training (Li et al., 2023b; Avrahami et al., 2023; Zhang et al., 2023a; Mou et al., 2023; Yang et al., 2023e; Huang et al., 2023b; a)","reference_id":"MTYzNTgxNzgxMTg1NDIzMjIxMzI="} +{"index":1723,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2302.08453","reference_contexts":"T2I-Adapter and ControlNet (Zhang et al., 2023a; Mou et al., 2023) specify some high-level features of images for controlling semantic structures (Zhang et al., 2023b)","reference_id":"MjIxMjgzNjIwNzIzOTY2NTE1Mg=="} +{"index":1724,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2211.01786","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1725,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2112.10741","reference_contexts":"GLIDE (Nichol et al., 2021) and Imagen (Saharia et al., 2022) focus on the text-guided image synthesis, leveraging pre-trained CLIP model (Radford et al., 2021; Raffel et al., 2020) in the image sampling process to improve the semantic alignment between text prompt and generated image","reference_id":"ODYyMTgyNTg4ODY1NTI4MzIyNg=="} +{"index":1726,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2310.02992","reference_contexts":"Further, Multimodal Large language Models (MLLMs), (Koh et al., 2023; Yu et al., 2023; Sun et al., 2023b; Dong et al., 2023; Fu et al., 2023; Pan et al., 2023; Wu et al., 2023a; Zou et al., 2023; Yang et al., 2023d; Gupta & Kembhavi, 2023; Surís et al., 2023) integrate LLMs with vision models to extend their impressive abilities from language tasks to vision tasks, including image understanding, reasoning and synthesis","reference_id":"MTU3MzkzMTI2NzkyMDE2MTQzOTQ="} +{"index":1727,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2310.02992","reference_contexts":"The collaboration between LLMs (ChatGPT, 2022; OpenAI, 2023) and diffusion models (Ramesh et al., 2022; Betker et al., 2023) can significantly improve the text-image alignment as well as the quality of generated images (Yu et al., 2023; Chen et al., 2023b; Dong et al., 2023; Wu et al., 2023b; Feng et al., 2023; Pan et al., 2023)","reference_id":"MTUwMzkwMTEyNzMxNzQ1ODM2MDA="} +{"index":1728,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2307.01952","reference_contexts":"Recent advancements in diffusion models (Sohl-Dickstein et al., 2015; Dhariwal & Nichol, 2021; Song et al., 2020; Yang et al., 2023c) have significantly improve the synthesis results of text-to-image models, such as Imagen (Saharia et al., 2022) , DALL-E 2/3 (Ramesh et al., 2022; Betker et al., 2023) and SDXL (Podell et al., 2023)","reference_id":"OTIzMDE4MzM4NjU4MDE4OTk3Nw=="} +{"index":1729,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2307.01952","reference_contexts":"Compared to SDXL (Podell et al., 2023) and DALL-E 3 (Betker et al., 2023) , our proposed RPG exhibits a superior ability to convey intricate and compositional text prompts within generated images (colored text denotes critical part).","reference_id":"NzcyMzA2NDc5NzE1NTE0MjI="} +{"index":1730,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2307.01952","reference_contexts":"It is worth noting that complementary regional diffusion can generalize to arbitrary diffusion backbones including SDXL (Podell et al., 2023) , ConPreDiff (Yang et al., 2023b) and ControlNet (Zhang et al., 2023a ), which will be evaluated in Section 3.1.","reference_id":"MjE5OTIxNjYyNzU4OTQxNjk5Nw=="} +{"index":1731,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2307.01952","reference_contexts":"Recent advancements in text-to-image diffusion models , such as SDXL (Podell et al., 2023) , ContextDiff (Yang et al., 2024b) and DALL-E 3 (Betker et al., 2023) , further improve both quality and alignment from different perspectives","reference_id":"MTgyNTIzMDcxNzc3OTI3NjcyNTc="} +{"index":1732,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2204.06125","reference_contexts":"Recent advancements in diffusion models (Sohl-Dickstein et al., 2015; Dhariwal & Nichol, 2021; Song et al., 2020; Yang et al., 2023c) have significantly improve the synthesis results of text-to-image models, such as Imagen (Saharia et al., 2022) , DALL-E 2/3 (Ramesh et al., 2022; Betker et al., 2023) and SDXL (Podell et al., 2023)","reference_id":"OTIzMDE4MzM4NjU4MDE4OTk3Nw=="} +{"index":1733,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2204.06125","reference_contexts":"The collaboration between LLMs (ChatGPT, 2022; OpenAI, 2023) and diffusion models (Ramesh et al., 2022; Betker et al., 2023) can significantly improve the text-image alignment as well as the quality of generated images (Yu et al., 2023; Chen et al., 2023b; Dong et al., 2023; Wu et al., 2023b; Feng et al., 2023; Pan et al., 2023)","reference_id":"MTUwMzkwMTEyNzMxNzQ1ODM2MDA="} +{"index":1734,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2306.08877","reference_contexts":"Other methods apply Gaussian kernels (Chefer et al., 2023) or incorporate linguistic features (Feng et al., 2022; Rassin et al., 2023) to manipulate the cross-attention maps","reference_id":"MTcwNjk2NjA2Njk3Mjc0OTYxNDY="} +{"index":1735,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2011.13456","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1736,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2311.17946","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1737,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2307.05222","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1738,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2303.08128","reference_contexts":"Further, Multimodal Large language Models (MLLMs), (Koh et al., 2023; Yu et al., 2023; Sun et al., 2023b; Dong et al., 2023; Fu et al., 2023; Pan et al., 2023; Wu et al., 2023a; Zou et al., 2023; Yang et al., 2023d; Gupta & Kembhavi, 2023; Surís et al., 2023) integrate LLMs with vision models to extend their impressive abilities from language tasks to vision tasks, including image understanding, reasoning and synthesis","reference_id":"MTU3MzkzMTI2NzkyMDE2MTQzOTQ="} +{"index":1739,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2211.09085","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1740,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2312.11805","reference_contexts":"Then we utilize MLLMs (e.g., GPT4 (OpenAI, 2023), Gemini Pro (Team et al., 2023) ) to check the differences between","reference_id":"ODY5NTQzODgzMTI1MDI5NDc2NA=="} +{"index":1741,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2302.13971","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1742,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2307.09288","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1743,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13921","reference_contexts":"Some works introduce additional layouts/boxes (Li et al., 2023b; Xie et al., 2023; Yang et al., 2023e; Qu et al., 2023; Chen et al., 2024; Wu et al., 2023b; Lian et al., 2023) as conditions or leveraging prompt-aware attention guidance (Feng et al., 2022; Chefer et al., 2023; Wang et al., 2023) to improve compositional text-to-image synthesis","reference_id":"NDQzNDcwMTc0NTQ0NTY3MzQ1"} +{"index":1744,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.13921","reference_contexts":"Complementary Regional Diffusion Recent works (Liu et al., 2022; Wang et al., 2023; Chefer et al., 2023; Feng et al., 2022) have adjusted cross-attention masks or layouts to facilitate compositional generation","reference_id":"MTEzOTgyMDg3MjY5MjEyNjY1MDY="} +{"index":1745,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2211.05100","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1746,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2303.04671","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1747,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.05977","reference_contexts":"Inspired by the reinforcement learning from human feedback (RLHF) (Ouyang et al., 2022; Stiennon et al., 2020) in natural language processing, ImageReward (Xu et al., 2023) builds a general-purpose reward model to improve text-to-image models in aligning with human preference.","reference_id":"MTE4MjA5ODUyMDUzNzIwMTAxMzU="} +{"index":1748,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.05977","reference_contexts":"Another potential solution is to leverage image understanding feedback (Huang et al., 2023a; Xu et al., 2023; Sun et al., 2023a; Fang et al., 2023) for refining diffusion generation","reference_id":"MTgxOTkyMzE4Njc5MzYxNjUzOTI="} +{"index":1749,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2309.10305","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1750,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2303.11381","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1751,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2309.02591","reference_contexts":"Further, Multimodal Large language Models (MLLMs), (Koh et al., 2023; Yu et al., 2023; Sun et al., 2023b; Dong et al., 2023; Fu et al., 2023; Pan et al., 2023; Wu et al., 2023a; Zou et al., 2023; Yang et al., 2023d; Gupta & Kembhavi, 2023; Surís et al., 2023) integrate LLMs with vision models to extend their impressive abilities from language tasks to vision tasks, including image understanding, reasoning and synthesis","reference_id":"MTU3MzkzMTI2NzkyMDE2MTQzOTQ="} +{"index":1752,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2309.02591","reference_contexts":"The collaboration between LLMs (ChatGPT, 2022; OpenAI, 2023) and diffusion models (Ramesh et al., 2022; Betker et al., 2023) can significantly improve the text-image alignment as well as the quality of generated images (Yu et al., 2023; Chen et al., 2023b; Dong et al., 2023; Wu et al., 2023b; Feng et al., 2023; Pan et al., 2023)","reference_id":"MTUwMzkwMTEyNzMxNzQ1ODM2MDA="} +{"index":1753,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2210.02414","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1754,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2205.01068","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1755,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2305.18583","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1756,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2302.00923","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1757,"citer_arxiv_id":"2401.11708","cited_arxiv_id":"2304.10592","reference_contexts":"Multimodal LLMs for Image Generation Large Language Models (LLMs) (ChatGPT, 2022; Chung et al., 2022; Zhang et al., 2022; Iyer et al., 2022; Workshop et al., 2022; Muennighoff et al., 2022; Zeng et al., 2022; Taylor et al., 2022; Chowdhery et al., 2023; Chen et al., 2023b; Zhu et al., 2023; Touvron et al., 2023a; Yang et al., 2023a; Li et al., 2023a) have profoundly impacted the AI community","reference_id":"MTg5ODU5ODgzMzQzNjUxNTU2Mw=="} +{"index":1758,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2306.14544","reference_contexts":"Several studies (Agarwal et al., 2023; Chefer et al., 2023) have attempted to resolve issues related to attribute binding by manipulating the attention mechanism within diffusion models","reference_id":"MzM3MjY3MTQ3MTIwMjIwNTA4Nw=="} +{"index":1759,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2301.00704","reference_contexts":"To address these issues, researchers have investigated the use of Transformer-based auto-regressive models for text-to-image generation (Ramesh et al., 2021; Ding et al., 2021; Esser et al., 2021a; Ding et al., 2022; Zhang et al., 2021; Lee et al., 2022; Chang et al., 2023) , combined with a discrete VAE (Van Den Oord et al., 2017; Razavi et al., 2019; Esser et al., 2021b) for image tokenization and Transformers (Vaswani et al., 2017) for modeling the joint distribution of textual and image tokens","reference_id":"MTcyNDQyNDQxMDU4OTI1MDYyOTM="} +{"index":1760,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2306.15195","reference_contexts":"There are also many inspiring LVLM works for exploring different tuning strategies (Ye et al., 2023) , collecting more diverse datasets (Chen et al., 2023) , and incorporating multi-modality (Guo et al., 2023) .","reference_id":"MTc4OTUzOTg4Nzg3MTgyNDI4MDU="} +{"index":1761,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2204.14217","reference_contexts":"To address these issues, researchers have investigated the use of Transformer-based auto-regressive models for text-to-image generation (Ramesh et al., 2021; Ding et al., 2021; Esser et al., 2021a; Ding et al., 2022; Zhang et al., 2021; Lee et al., 2022; Chang et al., 2023) , combined with a discrete VAE (Van Den Oord et al., 2017; Razavi et al., 2019; Esser et al., 2021b) for image tokenization and Transformers (Vaswani et al., 2017) for modeling the joint distribution of textual and image tokens","reference_id":"MTcyNDQyNDQxMDU4OTI1MDYyOTM="} +{"index":1762,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2304.06767","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1763,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2304.15010","reference_contexts":"Therein, LLaMA-Adapter series (Zhang et al., 2023a; Gao et al., 2023; Han et al., 2023) introduce zero-initialized attention mechanisms, and conduct multi-modal parameter-efficient fine-tuning","reference_id":"MTMxNTE0NzI0NTY1MDQ4MzE4Ng=="} +{"index":1764,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2309.00615","reference_contexts":"There are also many inspiring LVLM works for exploring different tuning strategies (Ye et al., 2023) , collecting more diverse datasets (Chen et al., 2023) , and incorporating multi-modality (Guo et al., 2023) .","reference_id":"MTc4OTUzOTg4Nzg3MTgyNDI4MDU="} +{"index":1765,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2309.03905","reference_contexts":"Therein, LLaMA-Adapter series (Zhang et al., 2023a; Gao et al., 2023; Han et al., 2023) introduce zero-initialized attention mechanisms, and conduct multi-modal parameter-efficient fine-tuning","reference_id":"MTMxNTE0NzI0NTY1MDQ4MzE4Ng=="} +{"index":1766,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2212.09611","reference_contexts":"Efforts have been made to align text-to-image generation models with human preferences and aesthetic standards (Hao et al., 2022; Lee et al., 2023; Wu et al., 2023; Xu et al., 2023; Dong et al., 2023a; Fang et al., 2023)","reference_id":"MTQ2NDk5MDg0NjA3MDE3NTQ2NjY="} +{"index":1767,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2303.11897","reference_contexts":"To evaluate the alignment of generated images in compositional text-to-image synthesis, we integrate Large Vision-Language Models (LVLMs) into the Text-to-Image Faithfulness evaluation with Question Answering (TIFA (Hu et al., 2023) ) framework","reference_id":"NzM1MjkwNDEwOTU4Nzk3ODk1Nw=="} +{"index":1768,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2307.06350","reference_contexts":"While there are methodologies like T2I-CompBench (Huang et al., 2023) that introduce compositional evaluation methods, these are primarily applied to finetune diffusion models during the training phase","reference_id":"MzA1MDQ2NDk0MDI1ODA5NjQxMg=="} +{"index":1769,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2303.05511","reference_contexts":"(2016) , and numerous subsequent GANbased models sought to enhance visual fidelity and caption congruence (Zhang et al., 2017; 2018; Xu et al., 2018; Li et al., 2019; Dong et al., 2017; Zhu et al., 2019; Tao et al., 2020; Ye et al., 2021; Kang et al., 2023; Sauer et al., 2023)","reference_id":"MTQ0NzczMzMzMTcyNTYxMjEyODY="} +{"index":1770,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2302.12192","reference_contexts":"Efforts have been made to align text-to-image generation models with human preferences and aesthetic standards (Hao et al., 2022; Lee et al., 2023; Wu et al., 2023; Xu et al., 2023; Dong et al., 2023a; Fang et al., 2023)","reference_id":"MTQ2NDk5MDg0NjA3MDE3NTQ2NjY="} +{"index":1771,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2301.12597","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1772,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2110.05208","reference_contexts":"Early efforts, such as CLIP (Radford et al., 2021) , ALIGN (Jia et al., 2021) , and follow-up works (Li et al., 2021; Dong et al., 2023b; Zhang et al., 2022) , adopt vision-language contrastive pre-training on extensive web-scale data, emerging superior generalization performance for zero-shot evaluation","reference_id":"MTU2MDk3NTUxOTc5Mzg4MDU3NzU="} +{"index":1773,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2304.08485","reference_contexts":"LLaVA (Liu et al., 2023) introduces a high-quality visual instruction dataset to fully fine-tune the entire LLaMA, while MiniGPT-4 (Zhu et al., 2023) only adopts a projection layer for vision-language alignment","reference_id":"MTE5NzkyNzM5MjIyMDgzMjE4MzA="} +{"index":1774,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2112.10741","reference_contexts":"For instance, GLIDE (Nichol et al., 2021) innovatively conditions the diffusion model on an input caption, building upon earlier works in the diffusion model sphere (Dhariwal & Nichol, 2021; Ho & Salimans, 2022)","reference_id":"MzE5NzM2NjA0Nzk1MTI4NDU3Mg=="} +{"index":1775,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2112.10741","reference_contexts":"Diffusion models have also been embraced for text-to-image generation (Nichol et al., 2021; Ho et al., 2022; Ramesh et al., 2022; Saharia et al., 2022; Rombach et al., 2022; Xu et al., 2022; Zhang et al., 2023b)","reference_id":"MTQ3NDc4ODQ5MTUxMjM1NTQyNzA="} +{"index":1776,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2303.08774","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1777,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2204.06125","reference_contexts":"Moreover, DALL-E 2 (Ramesh et al., 2022) enhances the GLIDE model by conditioning on a supplemental CLIP image embedding for heightened diversity","reference_id":"MTAwNjg5NTk3OTIxMTMyMTAxNDI="} +{"index":1778,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2204.06125","reference_contexts":"Diffusion models have also been embraced for text-to-image generation (Nichol et al., 2021; Ho et al., 2022; Ramesh et al., 2022; Saharia et al., 2022; Rombach et al., 2022; Xu et al., 2022; Zhang et al., 2023b)","reference_id":"MTQ3NDc4ODQ5MTUxMjM1NTQyNzA="} +{"index":1779,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2301.09515","reference_contexts":"(2016) , and numerous subsequent GANbased models sought to enhance visual fidelity and caption congruence (Zhang et al., 2017; 2018; Xu et al., 2018; Li et al., 2019; Dong et al., 2017; Zhu et al., 2019; Tao et al., 2020; Ye et al., 2021; Kang et al., 2023; Sauer et al., 2023)","reference_id":"MTQ0NzczMzMzMTcyNTYxMjEyODY="} +{"index":1780,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2008.05865","reference_contexts":"(2016) , and numerous subsequent GANbased models sought to enhance visual fidelity and caption congruence (Zhang et al., 2017; 2018; Xu et al., 2018; Li et al., 2019; Dong et al., 2017; Zhu et al., 2019; Tao et al., 2020; Ye et al., 2021; Kang et al., 2023; Sauer et al., 2023)","reference_id":"MTQ0NzczMzMzMTcyNTYxMjEyODY="} +{"index":1781,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2302.13971","reference_contexts":"In contrast to the powerful but close-source GPT-4 (OpenAI, 2023b) and Bard (Google, 2023) , a new branch of LVLMs is based on the open-source LLaMA (Touvron et al., 2023) , and endows it with image understanding ability by visual instruction tuning","reference_id":"MTMxNzM3ODE5ODM5NjIwMDg1ODg="} +{"index":1782,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2303.14420","reference_contexts":"Efforts have been made to align text-to-image generation models with human preferences and aesthetic standards (Hao et al., 2022; Lee et al., 2023; Wu et al., 2023; Xu et al., 2023; Dong et al., 2023a; Fang et al., 2023)","reference_id":"MTQ2NDk5MDg0NjA3MDE3NTQ2NjY="} +{"index":1783,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2211.08332","reference_contexts":"Diffusion models have also been embraced for text-to-image generation (Nichol et al., 2021; Ho et al., 2022; Ramesh et al., 2022; Saharia et al., 2022; Rombach et al., 2022; Xu et al., 2022; Zhang et al., 2023b)","reference_id":"MTQ3NDc4ODQ5MTUxMjM1NTQyNzA="} +{"index":1784,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2107.02423","reference_contexts":"(2016) , and numerous subsequent GANbased models sought to enhance visual fidelity and caption congruence (Zhang et al., 2017; 2018; Xu et al., 2018; Li et al., 2019; Dong et al., 2017; Zhu et al., 2019; Tao et al., 2020; Ye et al., 2021; Kang et al., 2023; Sauer et al., 2023)","reference_id":"MTQ0NzczMzMzMTcyNTYxMjEyODY="} +{"index":1785,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2304.14178","reference_contexts":"There are also many inspiring LVLM works for exploring different tuning strategies (Ye et al., 2023) , collecting more diverse datasets (Chen et al., 2023) , and incorporating multi-modality (Guo et al., 2023) .","reference_id":"MTc4OTUzOTg4Nzg3MTgyNDI4MDU="} +{"index":1786,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2303.16199","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1787,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2305.03048","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1788,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2105.14211","reference_contexts":"To address these issues, researchers have investigated the use of Transformer-based auto-regressive models for text-to-image generation (Ramesh et al., 2021; Ding et al., 2021; Esser et al., 2021a; Ding et al., 2022; Zhang et al., 2021; Lee et al., 2022; Chang et al., 2023) , combined with a discrete VAE (Van Den Oord et al., 2017; Razavi et al., 2019; Esser et al., 2021b) for image tokenization and Transformers (Vaswani et al., 2017) for modeling the joint distribution of textual and image tokens","reference_id":"MTcyNDQyNDQxMDU4OTI1MDYyOTM="} +{"index":1789,"citer_arxiv_id":"2310.06311","cited_arxiv_id":"2304.10592","reference_contexts":"LLaVA (Liu et al., 2023) introduces a high-quality visual instruction dataset to fully fine-tune the entire LLaMA, while MiniGPT-4 (Zhu et al., 2023) only adopts a projection layer for vision-language alignment","reference_id":"MTE5NzkyNzM5MjIyMDgzMjE4MzA="} +{"index":1790,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2305.10403","reference_contexts":"Finetuning generative models with predicted scores To generate the training prompt set, we provide five handcrafted seed prompts as examples and then ask PaLM 2 [1] to generate similar textual prompts","reference_id":"ODg1OTYzNjU5NDA0MDA2NDUyNw=="} +{"index":1791,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2305.10403","reference_contexts":"The score predictor consists of four convolutional layers with layer norm and ReLU activation, and the filter size, kernel size, and strides are [768, 384, 128, 64] , [2, 2, 2, 2], [1, 1, 1, 1], respectively","reference_id":"MTYzOTg2OTc0ODUzNjMyNzk2MjM="} +{"index":1792,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2305.10403","reference_contexts":"First, we generate eight images for each of the 12,564 prompts (the prompt set is created via PaLM 2 [1, 11] with some seed prompts) using the pre-trained Muse model","reference_id":"Mjk1NDA4MTY2NzUwMTY3NTkxNQ=="} +{"index":1793,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"1809.11096","reference_contexts":"The heatmap predictor consists of two convolution layers with filter size, kernel size, and stride as [768, 384], [3, 3] , [1, 1] , respectively","reference_id":"MTgyMTExMTU2NzgyMTY0NzMwNzE="} +{"index":1794,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"1809.11096","reference_contexts":"An early work is the Generative Adversarial Network (GAN) [3, 16, 27] , which trains a generator for image generation and a discriminator to distinguish between real and generated images, in parallel (also see [33, 39, 48, 56, 61, 63] among others)","reference_id":"MTgwMDc4MzEwNDU0MDMyMzA0MDk="} +{"index":1795,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2301.00704","reference_contexts":"Muse inpainting [6] is applied within the masked region to generate new images that match the text prompt","reference_id":"MTU4MzAyOTUyMTU0MjM2NjI0MTg="} +{"index":1796,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2301.00704","reference_contexts":"To ensure that the gains from our RAHF model generalize across generative model families, we mainly use Muse [6] as our target model to improve, which is based on a masked transformer architecture and thus different from the Stable Diffusion model variants in our RichHF-18K dataset.","reference_id":"MTgyMjI0MjMwNjY4Mzc3MzQ5NzM="} +{"index":1797,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2301.00704","reference_contexts":"We further demonstrate the usefulness of the predicted rich human feedback by RAHF to improve image generation: (i) by using the predicted heatmaps as masks to inpaint problematic image regions and (ii) by using the predicted scores to help finetune image generation models (like Muse [6] ), e.g., via selecting/filtering finetuning data, or as reward guidance","reference_id":"MTY5NDA2MTk3NzIzMTM3NDg5NTE="} +{"index":1798,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2310.18235","reference_contexts":"In terms of prompt-image alignment, there are also seminal single-score metrics such as CLIP-Score [19] and more recent question generation and answering pipelines [8, 10, 25, 58]","reference_id":"NTY4ODM1NTA3NjcyOTQ2MDYxNw=="} +{"index":1799,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2204.02311","reference_contexts":"First, we generate eight images for each of the 12,564 prompts (the prompt set is created via PaLM 2 [1, 11] with some seed prompts) using the pre-trained Muse model","reference_id":"Mjk1NDA4MTY2NzUwMTY3NTkxNQ=="} +{"index":1800,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2010.11929","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1801,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2210.02303","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1802,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2303.11897","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1803,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2307.06350","reference_contexts":"[26] proposed a benchmark called T2I-CompBench for evaluating text-toimage models, which consists of 6,000 text prompts describing attribute binding, object relationships, and complex compositions","reference_id":"MTI4MDYwODQ0NDg4ODMyMTE4NTI="} +{"index":1804,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2307.06350","reference_contexts":"Text-to-image evaluation and reward models There has been much recent work on evaluation of text-to-image models along many dimensions [9, 26, 31, 32, 38, 52, 53, 55]","reference_id":"NTY2MDQxMjYwNjU5MDIzMDg3NA=="} +{"index":1805,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"1312.6114","reference_contexts":"Another category of generation models develops from variational auto-encoders (VAEs) [21, 30, 49] , which optimize evidence lower bound (ELBO) for the likelihood of the image data.","reference_id":"MTc3MjUwMDQ4NDMzOTY1MzU4NTA="} +{"index":1806,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2305.01569","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1807,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2112.10741","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1808,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2203.17189","reference_contexts":"We adopt a vision-language model based on ViT [14] and T5X [41] models, inspired by the Spotlight model architecture [34], but modifying both the model and pretraining datasets to better suit our tasks","reference_id":"MTYwMzY4Nzg4MTQwODkxOTQ2MjM="} +{"index":1809,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2306.09341","reference_contexts":"[52, 53] collected a large scale dataset of human choices on generated images and utilized the dataset to train a classifier that outputs a Human Preference Score (HPS)","reference_id":"MjIwOTM0ODAyNzA3OTMyNTgyMA=="} +{"index":1810,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2306.09341","reference_contexts":"Text-to-image evaluation and reward models There has been much recent work on evaluation of text-to-image models along many dimensions [9, 26, 31, 32, 38, 52, 53, 55]","reference_id":"NTY2MDQxMjYwNjU5MDIzMDg3NA=="} +{"index":1811,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2310.10647","reference_contexts":"Text-to-image (T2I) generation models [12, 17, 42, 43, 57, 59, 60] are rapidly becoming a key to content creation in various domains, including entertainment, art, design, and advertising, and are also being generalized to image editing [4, 28, 45, 51] , video generation [23, 36, 54] , among many other applications","reference_id":"MTM2ODc1MTc3NjI0MjIwNDM5NzU="} +{"index":1812,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2206.10789","reference_contexts":"Existing datasets are 1) relatively small (e.g., TIFA [24] has 4k prompts, Davidsonian Scene Graph (DSG) [10] has only 1k prompts), or 2) containing prompts that are simple and not diverse enough, for example, only measuring single objects in Parti benchmark [59]","reference_id":"MzU5NjI5Mzg2MjcyOTAxMzMxNw=="} +{"index":1813,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2206.10789","reference_contexts":"Text-to-image (T2I) generation models [12, 17, 42, 43, 57, 59, 60] are rapidly becoming a key to content creation in various domains, including entertainment, art, design, and advertising, and are also being generalized to image editing [4, 28, 45, 51] , video generation [23, 36, 54] , among many other applications","reference_id":"MTM2ODc1MTc3NjI0MjIwNDM5NzU="} +{"index":1814,"citer_arxiv_id":"2312.10240","cited_arxiv_id":"2303.07909","reference_contexts":"Text-to-image (T2I) generation models [12, 17, 42, 43, 57, 59, 60] are rapidly becoming a key to content creation in various domains, including entertainment, art, design, and advertising, and are also being generalized to image editing [4, 28, 45, 51] , video generation [23, 36, 54] , among many other applications","reference_id":"MTM2ODc1MTc3NjI0MjIwNDM5NzU="} +{"index":1815,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.05814","reference_contexts":"For the Seg-to-3D task, we train two additional parsing networks [99] with labels provided by Pix2Pix3D [19] , where the segmentation ground-truth of cats is obtained via clustering the DINO feature as proposed by [3]","reference_id":"MTgzMDEwNDg5MzA5MTQxNDcxMzg="} +{"index":1816,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2302.07121","reference_contexts":"Prior research studied guidance with a variety of classifiers, constraints or auxiliary loss functions [6, 17, 20, 24, 27, 33, 44]","reference_id":"ODg4NDY2OTUyODgzMDU0NjIxNg=="} +{"index":1817,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"Model Our method is largely based on EG3D [12] † which adopts tri-plane representations to achieve efficient rendering process","reference_id":"NTQxNTgwOTE2NzA0Mzk0NDU0MA=="} +{"index":1818,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"To address this limitation, we propose a framework called Control3Diff, which links diffusion models to generative adversarial networks (GANs) [23] and takes advantage of the success of GANs in 3D-aware image synthesis [11, 12, 25, 63, 64, 81, 86]","reference_id":"NjUwODQzMDMyMDYyODAzNzU3Mw=="} +{"index":1819,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"Recently, 3D-GANs [12, 13, 25, 62, 81, 97] have been developed by integrating 3D representation and rendering into GANs","reference_id":"OTg1OTAwNzE5NDg5MzA5MDM0OA=="} +{"index":1820,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"We directly adopted the downsampled, realigned version provided by EG3D [12] , which re-cropped the face and estimate the camera poses.","reference_id":"NDM0ODU3NDI1MDYwNjg4NTEyNg=="} +{"index":1821,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"To validate the proposed framework, we use a variant of the recently proposed EG3D [12] that learns an efficient tri-plane representation as the basis for Control3Diff","reference_id":"OTIwMjYxNDIxMDU0OTM5MDE4MQ=="} +{"index":1822,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"f z is parameterized as MLPs [57] or tri-planes [12, 15] with upsamplers [25, 63]","reference_id":"MTQzNTI3ODA2ODc4Mzk0NjM1NzA="} +{"index":1823,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"Following EG3D [12] , each image is associated with its camera pose","reference_id":"MTMzNjc4OTY4NjQ5MTQ5NjQ0OTI="} +{"index":1824,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"In this paper, considering its state-of-the-art performance, we build Control3Diff based on EG3D [12]","reference_id":"MTM2MTQzNTMzNTQ1NTQzMTE4OTA="} +{"index":1825,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"We perform exactly the same pre-processing procedure as EG3D [12] over these synthetic images, which re-centers the faces and estimates the camera positions.","reference_id":"MTU3ODU3NjYwNDg4OTc3ODUzMjc="} +{"index":1826,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"* https://github.com/sxyu/pixel-nerf † https://github.com/NVlabs/eg3d.git Training We follow similar recipes [12] for training EG3D models on four datasets","reference_id":"MTg0MTk2NzQ4NzcwMTAwNDE2ODE="} +{"index":1827,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2112.07945","reference_contexts":"We adhere to the same pre-processing procedure utilized by EG3D [12] for these synthetic images","reference_id":"MzY0ODEyNzkyODc1MDYyMjY0NA=="} +{"index":1828,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2203.09517","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1829,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2209.14687","reference_contexts":"Prior research studied guidance with a variety of classifiers, constraints or auxiliary loss functions [6, 17, 20, 24, 27, 33, 44]","reference_id":"ODg4NDY2OTUyODgzMDU0NjIxNg=="} +{"index":1830,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2206.09012","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1831,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2110.08985","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1832,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2302.10109","reference_contexts":"For example, NerfDiff [27] applies a test-time optimization by distilling 2D diffusion priors into NeRF for single-view reconstruction","reference_id":"MTU2OTg3MjM4Mjg3MjQyMTY1NjA="} +{"index":1833,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2302.10109","reference_contexts":"Another line of work [18, 27, 67, 93, 104] applies 2D diffusion priors to the sparse-view reconstruction or text-to-3D generation tasks","reference_id":"MTM5NDI4NzAzMTg0Mzc0NDg1ODU="} +{"index":1834,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2302.10109","reference_contexts":"Additionally, similar to [27, 96] , we include a cross-attention layer between each self-attention outputs of the encoder and denoiser to strengthen the conditional modeling","reference_id":"MTMxNjU3MzAzODQ2MzQ4OTQ5MjY="} +{"index":1835,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2302.10109","reference_contexts":"Prior research studied guidance with a variety of classifiers, constraints or auxiliary loss functions [6, 17, 20, 24, 27, 33, 44]","reference_id":"ODg4NDY2OTUyODgzMDU0NjIxNg=="} +{"index":1836,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2302.10109","reference_contexts":"We implemented all our models based on the standard U-Net architectures [20] where for conditional diffusion models, an U-Net-based encoder is adopted to encode the input image similar to [27] , see Fig","reference_id":"NzE3MTk5ODcwNjg4ODY2NDQ3OA=="} +{"index":1837,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2004.02546","reference_contexts":"Follow-up works enable controllability by either adding conditioning input along with the sampled vectors as input (named \"Conditional GAN\") [35, 65] or manipulating the sampled vectors [28, 82, 105]","reference_id":"MTU4Mzg3NTAyOTMzNTg2MDE2NTI="} +{"index":1838,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2207.12598","reference_contexts":"Finally, by applying the guidance techniques [20, 30] in 2D diffusion models, we are able to learn controllable 3D generation with a single loss function for all conditional tasks","reference_id":"MTc1NjExMDI2MzkzODE3Mjc0MDg="} +{"index":1839,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2207.12598","reference_contexts":"Other works explored learning conditional diffusion models with parallel data (e.g., class labels [30] , text prompts [73], aligned image maps [100] )","reference_id":"OTYwMDc3MzcxNTQ2Mzk3ODAyMA=="} +{"index":1840,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2207.12598","reference_contexts":"No classifier-free guidance [30] is applied","reference_id":"NDI5MTE2OTQ2MTA3MzM5NTA4"} +{"index":1841,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2210.02303","reference_contexts":"Recently, significant progress has been made in the field of 2D image generation through the use of diffusionbased generative models [20, 31, 87, 89] , which learn the prior and have achieved remarkable success in various conditional applications such as super-resolution [26, 49, 75] , in-painting [54] , image translation [76, 100] and text-guided synthesis [32, 70, 73, 77]","reference_id":"MTAwNzYyMjMwNjQ0NTE3NDY0NDk="} +{"index":1842,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2204.03458","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1843,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2301.11093","reference_contexts":"To encourage our high-resolution denosier to learn sufficiently on noisy tri-planes, we adopt a shifted cosine schedule (256 2 → 64 2 ) inspired by [34]","reference_id":"MTQyOTc3MTI1OTA0NzI5NDUxNTM="} +{"index":1844,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2201.11793","reference_contexts":"Prior research studied guidance with a variety of classifiers, constraints or auxiliary loss functions [6, 17, 20, 24, 27, 33, 44]","reference_id":"ODg4NDY2OTUyODgzMDU0NjIxNg=="} +{"index":1845,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1846,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"1711.05101","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1847,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"1610.03483","reference_contexts":"Although it is efficient to sample high-quality tri-planes z, GANs are implicit generative models [58] and do not model the likelihood in the learned latent space","reference_id":"MTEzNzE2MDA2MTgxODA4MDYzNzc="} +{"index":1848,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2212.01206","reference_contexts":"Different from [60, 95] , we do not need any auxiliary loss or architectural change","reference_id":"MTEyMDk2NjI1MDg2NTgyNDA2MDM="} +{"index":1849,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2212.01206","reference_contexts":"More precisely, targeting on Challenge # 1, prior arts [7, 60, 83, 95] first reconstruct the latent z from dense multi-view images of each scene","reference_id":"MTc3MDE5NTA2NTY5MDkzMTEyODE="} +{"index":1850,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2212.09748","reference_contexts":"In the exploration stage, we also tried different architectures such as Transformers [66] , however, we did not notice significant difference on generation, and keep UNet as the basic backbone","reference_id":"MTQxMTU1Mjc0MTUzNDUxMTI3MTQ="} +{"index":1851,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2212.09748","reference_contexts":"Typically, θ is parameterized as a U-Net [31, 74] or ViT [66] .","reference_id":"MTQwMzY2MjU0ODkxODEzNjQ0NjM="} +{"index":1852,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2209.14988","reference_contexts":"We also conduct experiments with text feature guidance in (d), where we directly apply a pre-trained 2D diffusion model as a score function similar to DreamFusion [67] , and guide the generation of an unconditional 3D diffusion model.","reference_id":"MTAzMjg0ODMzNzUwNTEzMzM4NTg="} +{"index":1853,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2209.14988","reference_contexts":"Another line of work [18, 27, 67, 93, 104] applies 2D diffusion priors to the sparse-view reconstruction or text-to-3D generation tasks","reference_id":"MTM5NDI4NzAzMTg0Mzc0NDg1ODU="} +{"index":1854,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2204.06125","reference_contexts":"Importantly, classifier-free guidance [30] , which enables generation with a balance between sampling controlled quality and diversity, has become a basic building block for large-scale diffusion models [70, 77] .","reference_id":"NzUxMDMxNDczODExOTcwNTEwNQ=="} +{"index":1855,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2204.06125","reference_contexts":"Recently, significant progress has been made in the field of 2D image generation through the use of diffusionbased generative models [20, 31, 87, 89] , which learn the prior and have achieved remarkable success in various conditional applications such as super-resolution [26, 49, 75] , in-painting [54] , image translation [76, 100] and text-guided synthesis [32, 70, 73, 77]","reference_id":"MTAwNzYyMjMwNjQ0NTE3NDY0NDk="} +{"index":1856,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2204.06125","reference_contexts":"While it is commonly known that the text and image spaces of CLIP are not fully aligned [70] , we find the conditioning is effective as long as both features are normalized before diffusion.","reference_id":"MTkwNDUzOTczMDQzMDc1OTAyNA=="} +{"index":1857,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2104.07636","reference_contexts":"Recently, significant progress has been made in the field of 2D image generation through the use of diffusionbased generative models [20, 31, 87, 89] , which learn the prior and have achieved remarkable success in various conditional applications such as super-resolution [26, 49, 75] , in-painting [54] , image translation [76, 100] and text-guided synthesis [32, 70, 73, 77]","reference_id":"MTAwNzYyMjMwNjQ0NTE3NDY0NDk="} +{"index":1858,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2205.11487","reference_contexts":"Importantly, classifier-free guidance [30] , which enables generation with a balance between sampling controlled quality and diversity, has become a basic building block for large-scale diffusion models [70, 77] .","reference_id":"NzUxMDMxNDczODExOTcwNTEwNQ=="} +{"index":1859,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2205.11487","reference_contexts":"Recently, significant progress has been made in the field of 2D image generation through the use of diffusionbased generative models [20, 31, 87, 89] , which learn the prior and have achieved remarkable success in various conditional applications such as super-resolution [26, 49, 75] , in-painting [54] , image translation [76, 100] and text-guided synthesis [32, 70, 73, 77]","reference_id":"MTAwNzYyMjMwNjQ0NTE3NDY0NDk="} +{"index":1860,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2202.00512","reference_contexts":"However, our methods can be easily integrated with existing works for speed-up diffusion models [79]","reference_id":"MTQ1OTgzOTY5OTYxNTk2NTAxODc="} +{"index":1861,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2202.00512","reference_contexts":"Our initial experiments showed that predicting the noise ϵ (default setting as suggested by DDPM [31] ) or the velocity v [79] tend to produce noticeable high-frequency artifacts on the generated tri-planes","reference_id":"MTEwOTI5MjA0MDA1ODYwMzEzMTY="} +{"index":1862,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2210.08402","reference_contexts":"Conventionally, learning such conditional models requires labeling parallel corpus, e.g., large-scale text-image pairs [80] for the Text-to-Image task","reference_id":"NDYxMjk0NDMxMTI0OTQ0MTAxMw=="} +{"index":1863,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2007.06600","reference_contexts":"Follow-up works enable controllability by either adding conditioning input along with the sampled vectors as input (named \"Conditional GAN\") [35, 65] or manipulating the sampled vectors [28, 82, 105]","reference_id":"MTU4Mzg3NTAyOTMzNTg2MDE2NTI="} +{"index":1864,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2211.16677","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1865,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2010.02502","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1866,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2011.13456","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1867,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2102.02766","reference_contexts":"Baselines We choose the standard optimization-based and encoder-based [47, 92] approaches for image-to-3D inversion, and the recent Pix2Pix3D [19] as the major baseline to compare on the Seg-to-3D task","reference_id":"MjkyNDM0MjQ3ODYyNjYzMDgwOQ=="} +{"index":1868,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2212.06135","reference_contexts":"Different from [60, 95] , we do not need any auxiliary loss or architectural change","reference_id":"MTEyMDk2NjI1MDg2NTgyNDA2MDM="} +{"index":1869,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2212.06135","reference_contexts":"More precisely, targeting on Challenge # 1, prior arts [7, 60, 83, 95] first reconstruct the latent z from dense multi-view images of each scene","reference_id":"MTc3MDE5NTA2NTY5MDkzMTEyODE="} +{"index":1870,"citer_arxiv_id":"2304.06700","cited_arxiv_id":"2210.04628","reference_contexts":"Additionally, similar to [27, 96] , we include a cross-attention layer between each self-attention outputs of the encoder and denoiser to strengthen the conditional modeling","reference_id":"MTMxNjU3MzAzODQ2MzQ4OTQ5MjY="} +{"index":1871,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2211.01324","reference_contexts":"To further transfer the style from one image to the synthesized video, we choose the image embedding as the stylistic guidance, following [4, 26] .","reference_id":"NjI1NjkxNTE1OTM5OTE3MTA1Nw=="} +{"index":1872,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2304.14404","reference_contexts":"Both MCDiff [9] and LaMD [25] target motion-guided video generation; the former focuses on generating human action videos and encodes the dynamics by tracking the keypoints and reference points, while the latter employs a learnable motion latent to improve quality","reference_id":"OTM3NzQwMzgyMTU0MjI1OTYyMQ=="} +{"index":1873,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2210.11427","reference_contexts":"DiffEdit [11] introduces automatically generated masks to assist text-driven image editing","reference_id":"MTYyNDcyMDQ0OTgxNDY5NjA4NjM="} +{"index":1874,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"1410.8516","reference_contexts":"Recently, research efforts on image synthesis have shifted from utilizing GANs [17] , VAEs [29] , and flow models [13] to diffusion models [21, 50, 54, 73] due to more stable training, enhanced sample quality, and increased flexibility in a conditional generation","reference_id":"MTE0OTc2NDc5Mjc3MTg0Njc4NjI="} +{"index":1875,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.03011","reference_contexts":"A subsequent work, Gen-1 [15] , integrates depth maps alongside texts using cross-attention mechanisms to provide structural guidance","reference_id":"MTM5OTQ0MzE5MzMwNjc4MzYxMTk="} +{"index":1876,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.03011","reference_contexts":"We utilize two metrics to evaluate VideoComposer: i) To evaluate video continuity, we follow Gen-1 [15] to compute the average CLIP cosine similarity of two consecutive frames, serving as a frame consistency metric; ii) To evaluate motion controllability, we adopt end-point-error [52, 67] as a motion control metric, which measures the Euclidean distance between the predicted and the ground truth optical flow for each pixel.","reference_id":"MTE0NDE0MjU0MDQ2MzM4NzU4OA=="} +{"index":1877,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2205.11495","reference_contexts":"Recent research has demonstrated the potential of employing diffusion models for video synthesis [6, 18, 23, 28, 32, 68]","reference_id":"NDI3ODMwODM2MzUwMzg2ODU2NQ=="} +{"index":1878,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2208.01626","reference_contexts":"Regarding image editing, pix2pix-zero [38] and prompt-to-prompt editing [19] follow instructional texts by manipulating cross-attention maps","reference_id":"NDYzNjA4MTAyNzQyMjMyOTMxNg=="} +{"index":1879,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2207.12598","reference_contexts":"We incorporate classifier-free guidance [20] to ensure that the generative results adhere to specified conditions","reference_id":"MTM2NDY0MjcwMzkwNTA4NzQzMDI="} +{"index":1880,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2210.02303","reference_contexts":"Notably, ImagenVideo [22] and Make-A-Video [49] both model the video distribution in pixel space, which limits their applicability due to high computational demands","reference_id":"MzQ3Mjg4NjA1MDgxODcxMDEyNQ=="} +{"index":1881,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2210.02303","reference_contexts":"Driven by the advances in computation, data scaling and architectural design, current visual generative models, especially diffusion-based models, have made remarkable strides in automating content creation, empowering designers to generate realistic images or videos from a textual prompt as input [22, 44, 49]","reference_id":"Mzk0Mjc3MDUzNjM4MjQ2NDUxOQ=="} +{"index":1882,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2210.02303","reference_contexts":"As with prior generative foundation models, such as Imagen Video [22] and Make-A-Video [49] , VideoComposer inherits the implicit knowledge embedded within the pre-trained model (i.e., StableDiffusion) and the pre-trained dataset (i.e., WebVid and LAION)","reference_id":"MTYyODgyMjI2NjAzOTU5NjYyODM="} +{"index":1883,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2204.03458","reference_contexts":"Recent research has demonstrated the potential of employing diffusion models for video synthesis [6, 18, 23, 28, 32, 68]","reference_id":"NDI3ODMwODM2MzUwMzg2ODU2NQ=="} +{"index":1884,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2204.03458","reference_contexts":"To exploit the inductive bias of locality and temporal inductive bias of sequentiality during denoising, we instantiate ϵ θ (•, •, t) as a 3D UNet augmented with temporal convolution and cross-attention mechanism following [1, 23, 45] .","reference_id":"MTM2NTU3MDU5OTA5ODIwNzg0NDY="} +{"index":1885,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2205.15868","reference_contexts":"Method Zero-shot FVD ↓ CLIPSIM ↑ GODIVA [64] No -0.2402 Nüwa [65] No -0.2439 CogVideo (Chinese) [24] Yes -0.2614 CogVideo (English) [24] Yes 1294 0.2631 MagicVideo [75] Yes 1290 -Make-A-Video [49] Yes -0.3049 Video LDM [ 6] Yes -0.2929 Text-to-video pre-training (First stage) Yes 803 0.2876 VideoComposer Yes 580 0.2932 In Fig","reference_id":"MTYwODc3NTM2NzUwNDAyNTM2MDU="} +{"index":1886,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2205.15868","reference_contexts":"8 , we present visualization comparing CogVideo [24] and VideoComposer","reference_id":"MzYwNTU0MTA4NjEzOTI1NDY1MA=="} +{"index":1887,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.09778","reference_contexts":"During the second stage pre-training, we adhere to [26] , using a probability of 0.1 to keep all conditions, a probability of 0.1 to discard all conditions, and an independent probability of 0.5 to keep or discard a specific condition","reference_id":"MTE5NTY1ODU3ODAzNzEwNTM3MzE="} +{"index":1888,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.09778","reference_contexts":"Expanding upon this idea, Composer [26] proposes a new generative paradigm centered on the concept of compositionality, which is capable of composing an image with various input conditions, leading to remarkable flexibility","reference_id":"OTkxODkyMDgyMjMyNjQ4MDk5OQ=="} +{"index":1889,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.09778","reference_contexts":"One milestone, Composer [26] , trains a multi-condition diffusion model that broadly expands the control space and displays remarkable results","reference_id":"NjM4MzcyMjg3ODc1OTc2MjIxNw=="} +{"index":1890,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.09778","reference_contexts":"To further transfer the style from one image to the synthesized video, we choose the image embedding as the stylistic guidance, following [4, 26] .","reference_id":"NjI1NjkxNTE1OTM5OTE3MTA1Nw=="} +{"index":1891,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2210.09276","reference_contexts":"Imagic [27] interpolates between an optimized embedding and the target embedding derived from text instructions to manipulate images","reference_id":"Mzk5MTc0NzI1NTIxNzU3ODk5Nw=="} +{"index":1892,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2303.13439","reference_contexts":"Recent research has demonstrated the potential of employing diffusion models for video synthesis [6, 18, 23, 28, 32, 68]","reference_id":"NDI3ODMwODM2MzUwMzg2ODU2NQ=="} +{"index":1893,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"1312.6114","reference_contexts":"Recently, research efforts on image synthesis have shifted from utilizing GANs [17] , VAEs [29] , and flow models [13] to diffusion models [21, 50, 54, 73] due to more stable training, enhanced sample quality, and increased flexibility in a conditional generation","reference_id":"MTE0OTc2NDc5Mjc3MTg0Njc4NjI="} +{"index":1894,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2303.04761","reference_contexts":"With the goal of editing videos guided by texts, VideoP2P [30] and vid2vid-zero [60] manipulate the crossattention map, while Dreamix [33] proposes an image-video mixed fine-tuning strategy","reference_id":"NjU1MTYwNjI5MTI1NDUwMjk5Mg=="} +{"index":1895,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"1711.05101","reference_contexts":"We adopt AdamW [31] as the default optimizer with a learning rate set to 5 × 10 -5","reference_id":"MTE3NjA5NDAxMTMxMjYyNjk5MjM="} +{"index":1896,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.01329","reference_contexts":"With the goal of editing videos guided by texts, VideoP2P [30] and vid2vid-zero [60] manipulate the crossattention map, while Dreamix [33] proposes an image-video mixed fine-tuning strategy","reference_id":"NjU1MTYwNjI5MTI1NDUwMjk5Mg=="} +{"index":1897,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.08453","reference_contexts":"To enable conditional synthesis with flexible input, ControlNet [72] and T2I-Adapter [34] incorporate a specific spatial condition into the model, providing more fine-grained control","reference_id":"MTA2Mzk4MzgyNDQ3MDg1MTA1MDg="} +{"index":1898,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.08453","reference_contexts":"Most existing methods typically achieve controllable generation mainly by introducing new conditions, such as segmentation maps [44, 59] , inpainting masks [66] or sketches [34, 72] , in addition to texts","reference_id":"MTQwMDEyOTY2NzY5MjQ0MzA0ODI="} +{"index":1899,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2112.10741","reference_contexts":"Intuitively, adapting image diffusion models that process in the pixel space [36, 42] to the video domain impedes the scaling of VideoComposer to web-scale data","reference_id":"ODM1Njg1NDEyODk3MTU1MTQ2Ng=="} +{"index":1900,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2112.10741","reference_contexts":"Regarding image generation, notable works such as DALL-E 2 [42] and GLIDE [36] employ diffusion models for text-to-image generation by conducting the diffusion process in pixel space, guided by CLIP [40] or classifier-free approaches","reference_id":"MjQ5NDM0NjIxNzUzNzU5NjMyMQ=="} +{"index":1901,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.03027","reference_contexts":"Regarding image editing, pix2pix-zero [38] and prompt-to-prompt editing [19] follow instructional texts by manipulating cross-attention maps","reference_id":"NDYzNjA4MTAyNzQyMjMyOTMxNg=="} +{"index":1902,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2204.06125","reference_contexts":"Intuitively, adapting image diffusion models that process in the pixel space [36, 42] to the video domain impedes the scaling of VideoComposer to web-scale data","reference_id":"ODM1Njg1NDEyODk3MTU1MTQ2Ng=="} +{"index":1903,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2204.06125","reference_contexts":"Regarding image generation, notable works such as DALL-E 2 [42] and GLIDE [36] employ diffusion models for text-to-image generation by conducting the diffusion process in pixel space, guided by CLIP [40] or classifier-free approaches","reference_id":"MjQ5NDM0NjIxNzUzNzU5NjMyMQ=="} +{"index":1904,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2111.02114","reference_contexts":"These approaches typically train a powerful diffusion model [44] conditioned by text [21] on large-scale video-text and image-text datasets [3, 47] , reaching unprecedented levels of fidelity and diversity","reference_id":"MTA0MDM3Njk5NDQ0NzM1MjU5MDk="} +{"index":1905,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2111.02114","reference_contexts":"To optimize VideoComposer, we leverage two widely recognized and publicly accessible datasets: WebVid10M [3] and LAION-400M [47]","reference_id":"MzE5MDE4NTYwNzU2MzUxMjUzNA=="} +{"index":1906,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2209.14792","reference_contexts":"Notably, ImagenVideo [22] and Make-A-Video [49] both model the video distribution in pixel space, which limits their applicability due to high computational demands","reference_id":"MzQ3Mjg4NjA1MDgxODcxMDEyNQ=="} +{"index":1907,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2209.14792","reference_contexts":"Driven by the advances in computation, data scaling and architectural design, current visual generative models, especially diffusion-based models, have made remarkable strides in automating content creation, empowering designers to generate realistic images or videos from a textual prompt as input [22, 44, 49]","reference_id":"Mzk0Mjc3MDUzNjM4MjQ2NDUxOQ=="} +{"index":1908,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2209.14792","reference_contexts":"As with prior generative foundation models, such as Imagen Video [22] and Make-A-Video [49] , VideoComposer inherits the implicit knowledge embedded within the pre-trained model (i.e., StableDiffusion) and the pre-trained dataset (i.e., WebVid and LAION)","reference_id":"MTYyODgyMjI2NjAzOTU5NjYyODM="} +{"index":1909,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2209.14792","reference_contexts":"Method Zero-shot FVD ↓ CLIPSIM ↑ GODIVA [64] No -0.2402 Nüwa [65] No -0.2439 CogVideo (Chinese) [24] Yes -0.2614 CogVideo (English) [24] Yes 1294 0.2631 MagicVideo [75] Yes 1290 -Make-A-Video [49] Yes -0.3049 Video LDM [ 6] Yes -0.2929 Text-to-video pre-training (First stage) Yes 803 0.2876 VideoComposer Yes 580 0.2932 In Fig","reference_id":"MTYwODc3NTM2NzUwNDAyNTM2MDU="} +{"index":1910,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2205.12952","reference_contexts":"Most existing methods typically achieve controllable generation mainly by introducing new conditions, such as segmentation maps [44, 59] , inpainting masks [66] or sketches [34, 72] , in addition to texts","reference_id":"MTQwMDEyOTY2NzY5MjQ0MzA0ODI="} +{"index":1911,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2303.17599","reference_contexts":"With the goal of editing videos guided by texts, VideoP2P [30] and vid2vid-zero [60] manipulate the crossattention map, while Dreamix [33] proposes an image-video mixed fine-tuning strategy","reference_id":"NjU1MTYwNjI5MTI1NDUwMjk5Mg=="} +{"index":1912,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2104.14806","reference_contexts":"Method Zero-shot FVD ↓ CLIPSIM ↑ GODIVA [64] No -0.2402 Nüwa [65] No -0.2439 CogVideo (Chinese) [24] Yes -0.2614 CogVideo (English) [24] Yes 1294 0.2631 MagicVideo [75] Yes 1290 -Make-A-Video [49] Yes -0.3049 Video LDM [ 6] Yes -0.2929 Text-to-video pre-training (First stage) Yes 803 0.2876 VideoComposer Yes 580 0.2932 In Fig","reference_id":"MTYwODc3NTM2NzUwNDAyNTM2MDU="} +{"index":1913,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2212.05034","reference_contexts":"Most existing methods typically achieve controllable generation mainly by introducing new conditions, such as segmentation maps [44, 59] , inpainting masks [66] or sketches [34, 72] , in addition to texts","reference_id":"MTQwMDEyOTY2NzY5MjQ0MzA0ODI="} +{"index":1914,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2203.09481","reference_contexts":"Recent research has demonstrated the potential of employing diffusion models for video synthesis [6, 18, 23, 28, 32, 68]","reference_id":"NDI3ODMwODM2MzUwMzg2ODU2NQ=="} +{"index":1915,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.05543","reference_contexts":"To enable conditional synthesis with flexible input, ControlNet [72] and T2I-Adapter [34] incorporate a specific spatial condition into the model, providing more fine-grained control","reference_id":"MTA2Mzk4MzgyNDQ3MDg1MTA1MDg="} +{"index":1916,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2302.05543","reference_contexts":"Most existing methods typically achieve controllable generation mainly by introducing new conditions, such as segmentation maps [44, 59] , inpainting masks [66] or sketches [34, 72] , in addition to texts","reference_id":"MTQwMDEyOTY2NzY5MjQ0MzA0ODI="} +{"index":1917,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2206.05564","reference_contexts":"During inference, DDIM [73] is employed to enhance the sample quality and improve inference efficiency","reference_id":"MzE4OTg3ODY0NzY0Nzg1ODc4NA=="} +{"index":1918,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2206.05564","reference_contexts":"Recently, research efforts on image synthesis have shifted from utilizing GANs [17] , VAEs [29] , and flow models [13] to diffusion models [21, 50, 54, 73] due to more stable training, enhanced sample quality, and increased flexibility in a conditional generation","reference_id":"MTE0OTc2NDc5Mjc3MTg0Njc4NjI="} +{"index":1919,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2211.11018","reference_contexts":"Method Zero-shot FVD ↓ CLIPSIM ↑ GODIVA [64] No -0.2402 Nüwa [65] No -0.2439 CogVideo (Chinese) [24] Yes -0.2614 CogVideo (English) [24] Yes 1294 0.2631 MagicVideo [75] Yes 1290 -Make-A-Video [49] Yes -0.3049 Video LDM [ 6] Yes -0.2929 Text-to-video pre-training (First stage) Yes 803 0.2876 VideoComposer Yes 580 0.2932 In Fig","reference_id":"MTYwODc3NTM2NzUwNDAyNTM2MDU="} +{"index":1920,"citer_arxiv_id":"2306.02018","cited_arxiv_id":"2211.11018","reference_contexts":"In contrast, MagicVideo [75] models the video distribution in the latent space, following the paradigm of LDMs [44] , significantly reducing computational overhead","reference_id":"ODY3OTk5MjM3MTUwNjIzMjU2NA=="} +{"index":1921,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2303.08774","reference_contexts":"Large language models (LLMs) have demonstrated remarkable capabilities in learning diverse knowledge and generalizing to new scenarios [1, 7, 8, 69, 91]","reference_id":"OTkzMzk4NjE1MjQ1NDA2NzMyMg=="} +{"index":1922,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2303.08774","reference_contexts":"After being trained on large-scale datasets, LLMs exhibit the ability to generalize across various tasks and follow diverse instructions [1, 8, 69]","reference_id":"MjY2OTc5ODczMTM0OTU3MDI2OQ=="} +{"index":1923,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2308.12966","reference_contexts":"In multimodal comprehension benchmarks, JanusFlow attains scores of 74.9, 70.5 and 60.3 on MMBench [63] , SeedBench [46] , and GQA [35] , respectively, exceeding specialized models such as LLaVA-v1.5 [56] and Qwen-VL-Chat [4]","reference_id":"NDE0NjYyMjUyNjk3MTgxMTI2Nw=="} +{"index":1924,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2308.12966","reference_contexts":"By leveraging large-scale multimodal datasets and increasingly powerful LLMs, this architecture has facilitated the development of advanced multimodal models capable of addressing a wide range of vision-language tasks [4, 47, 56, 64] .","reference_id":"NjQzOTE1NDIxNTM1MTg3MzMxMA=="} +{"index":1925,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2308.12966","reference_contexts":"We evaluate JanusFlow's multimodal understanding abilities across a diverse set of vision-language benchmarks for general understanding capabilities, including POPE [52] , MME [24] , MMBench [63] , SEEDBench [46] , VQAv2 [29] , GQA [35] , MM-Vet [104] , MMMU [105] , ChartQA [70] and TextVQA [81] Image Generation Performances","reference_id":"MTI2NTkxNjk0OTAzMTc3ODc2NjA="} +{"index":1926,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2401.02954","reference_contexts":"We directly use the text corpus of DeepSeek-LLM [7] .","reference_id":"MjEzMzI2NDk1MDIxMjIwNjcwMw=="} +{"index":1927,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2401.02954","reference_contexts":"Large language models (LLMs) have demonstrated remarkable capabilities in learning diverse knowledge and generalizing to new scenarios [1, 7, 8, 69, 91]","reference_id":"OTkzMzk4NjE1MjQ1NDA2NzMyMg=="} +{"index":1928,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2401.02954","reference_contexts":"Building upon powerful LLMs [7, 91, 92] , recent MLLMs [2, 15, 49, 56, 58, 64] have demonstrated exceptional multimodal understanding capabilities","reference_id":"ODU3NDc4NDgyNDQ5NTQ2NzI0NQ=="} +{"index":1929,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2401.02954","reference_contexts":"Our framework builds upon an enhanced version foot_0 of DeepSeek-LLM (1.3B) [7, 64]","reference_id":"Mjg4MjE1MTg2OTQ2Mjk1MDAyOQ=="} +{"index":1930,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2303.12712","reference_contexts":"Large language models (LLMs) have demonstrated remarkable capabilities in learning diverse knowledge and generalizing to new scenarios [1, 7, 8, 69, 91]","reference_id":"OTkzMzk4NjE1MjQ1NDA2NzMyMg=="} +{"index":1931,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2303.12712","reference_contexts":"After being trained on large-scale datasets, LLMs exhibit the ability to generalize across various tasks and follow diverse instructions [1, 8, 69]","reference_id":"MjY2OTc5ODczMTM0OTU3MDI2OQ=="} +{"index":1932,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2310.00426","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1933,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.04692","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1934,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2311.12793","reference_contexts":"The datasets used in the pre-training stage for understanding include DetailedCaption [20], SAM [41] , arXivQA [50] , DenseFusion-1M [51] , MMSci [53] , PixelProse [82] , re-captioned LAION-Aesthetics [16] , re-captioned Open Images V4 [43] , ShareGPT4V [11] , WikiHow [42] and WIT [84]","reference_id":"MTM5MDA4ODIzMzU0NjgzNTE3OTk="} +{"index":1935,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2312.16886","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1936,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2402.03766","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1937,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2306.13394","reference_contexts":"We evaluate JanusFlow's multimodal understanding abilities across a diverse set of vision-language benchmarks for general understanding capabilities, including POPE [52] , MME [24] , MMBench [63] , SEEDBench [46] , VQAv2 [29] , GQA [35] , MM-Vet [104] , MMMU [105] , ChartQA [70] and TextVQA [81] Image Generation Performances","reference_id":"MTI2NTkxNjk0OTAzMTc3ODc2NjA="} +{"index":1938,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2307.08041","reference_contexts":"This separation often results in suboptimal performance compared to standalone diffusion models [25, 87]","reference_id":"MTM3MDA1MDI5NjUxOTgyODE5MTM="} +{"index":1939,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2307.08041","reference_contexts":"One prominent direction involves utilizing pre-trained text-to-image models for high-quality generation while training LLMs to generate conditions for these models [19, [25] [26] [27] 87]","reference_id":"MTAwOTMyMTQ3NjExNDY1MzU4ODI="} +{"index":1940,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2310.01218","reference_contexts":"One prominent direction involves utilizing pre-trained text-to-image models for high-quality generation while training LLMs to generate conditions for these models [19, [25] [26] [27] 87]","reference_id":"MTAwOTMyMTQ3NjExNDY1MzU4ODI="} +{"index":1941,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2404.14396","reference_contexts":"One prominent direction involves utilizing pre-trained text-to-image models for high-quality generation while training LLMs to generate conditions for these models [19, [25] [26] [27] 87]","reference_id":"MTAwOTMyMTQ3NjExNDY1MzU4ODI="} +{"index":1942,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2209.08199","reference_contexts":"We leverage the instruction tuning datasets from [29, 33, 35, 47, 65, 80 ]","reference_id":"ODcxMjY2NTUxMjU1NzE4NjE="} +{"index":1943,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05135","reference_contexts":"Specifically, on text-to-image generation benchmarks, MJHQ FID-30k [48] , GenEval [28] and DPG-Bench [34] , JanusFlow achieves scores of 9.51, 0.63 and 80.09%, surpassing established text-to-image models including SDv1.5 [77] and SDXL [73]","reference_id":"MTEzODMzMzg3MjYwNDk2ODE2NTA="} +{"index":1944,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05135","reference_contexts":"To evaluate semantic accuracy, we utilize two specialized frameworks: GenEval [28] and DPG-Bench [34]","reference_id":"MTM0OTU2MDU5OTQzMTI0MDkwMjk="} +{"index":1945,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05135","reference_contexts":"The generation capabilities are evaluated using GenEval [28] , DPG-Benchmark [34] , and MJHQ FID-30k [48] , with results shown in Tab","reference_id":"MTA1NTMxODk5NzA3ODI5NzkyODc="} +{"index":1946,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"1810.09305","reference_contexts":"The datasets used in the pre-training stage for understanding include DetailedCaption [20], SAM [41] , arXivQA [50] , DenseFusion-1M [51] , MMSci [53] , PixelProse [82] , re-captioned LAION-Aesthetics [16] , re-captioned Open Images V4 [43] , ShareGPT4V [11] , WikiHow [42] and WIT [84]","reference_id":"MTM5MDA4ODIzMzU0NjgzNTE3OTk="} +{"index":1947,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"1810.09305","reference_contexts":"This sub-category is sourced from [42, 84]","reference_id":"OTkyNzcxNTg1OTMzMTIwMTE5MA=="} +{"index":1948,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2307.16125","reference_contexts":"In multimodal comprehension benchmarks, JanusFlow attains scores of 74.9, 70.5 and 60.3 on MMBench [63] , SeedBench [46] , and GQA [35] , respectively, exceeding specialized models such as LLaVA-v1.5 [56] and Qwen-VL-Chat [4]","reference_id":"NDE0NjYyMjUyNjk3MTgxMTI2Nw=="} +{"index":1949,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2307.16125","reference_contexts":"This version, trained on an expanded text corpus compared to the one in Janus [97] , has been demonstrated to possess better performance on multiple-choice benchmarks (e.g., MMBench [63] and SEED Bench [46] )","reference_id":"NDc4NzIzMDI5NDA4MTYzMjc4Ng=="} +{"index":1950,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2307.16125","reference_contexts":"We evaluate JanusFlow's multimodal understanding abilities across a diverse set of vision-language benchmarks for general understanding capabilities, including POPE [52] , MME [24] , MMBench [63] , SEEDBench [46] , VQAv2 [29] , GQA [35] , MM-Vet [104] , MMMU [105] , ChartQA [70] and TextVQA [81] Image Generation Performances","reference_id":"MTI2NTkxNjk0OTAzMTc3ODc2NjA="} +{"index":1951,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.03326","reference_contexts":"We directly incorporate the text-only data from [47] .","reference_id":"NTEzNTI4NjI4NDA5NTEyNzg4OA=="} +{"index":1952,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.03326","reference_contexts":"By leveraging large-scale multimodal datasets and increasingly powerful LLMs, this architecture has facilitated the development of advanced multimodal models capable of addressing a wide range of vision-language tasks [4, 47, 56, 64] .","reference_id":"NjQzOTE1NDIxNTM1MTg3MzMxMA=="} +{"index":1953,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.03326","reference_contexts":"We leverage the instruction tuning datasets from [29, 33, 35, 47, 65, 80 ]","reference_id":"ODcxMjY2NTUxMjU1NzE4NjE="} +{"index":1954,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.03326","reference_contexts":"Leveraging these capabilities, researchers have developed sophisticated models specialized in image comprehension [2, 15, 47, 49, 56, 58 ] and text-to-image generation [23, 73, 76, 79] .","reference_id":"MjAwMjY3MjAyNzk2ODk4Mjk4MA=="} +{"index":1955,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2402.17245","reference_contexts":"Specifically, on text-to-image generation benchmarks, MJHQ FID-30k [48] , GenEval [28] and DPG-Bench [34] , JanusFlow achieves scores of 9.51, 0.63 and 80.09%, surpassing established text-to-image models including SDv1.5 [77] and SDXL [73]","reference_id":"MTEzODMzMzg3MjYwNDk2ODE2NTA="} +{"index":1956,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2402.17245","reference_contexts":"The generation capabilities are evaluated using GenEval [28] , DPG-Benchmark [34] , and MJHQ FID-30k [48] , with results shown in Tab","reference_id":"MTA1NTMxODk5NzA3ODI5NzkyODc="} +{"index":1957,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2402.17245","reference_contexts":"For visual quality assessment, we employ the Fréchet Inception Distance [30] (FID) metric and compute FID between 30,000 generated images and their corresponding reference images from the MJHQ dataset [48]","reference_id":"NDIwMzM5Njg1Nzk3ODk3MTE4Ng=="} +{"index":1958,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2405.08748","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1959,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2402.08268","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1960,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2209.14577","reference_contexts":"Our work demonstrates that rectified flow [60] [61] [62] can be effectively integrated into LLMs, creating unified models that excel in both understanding and generation tasks.","reference_id":"MTQ2Mzk4MTc1MDA0NjkzNzE3MTc="} +{"index":1961,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05525","reference_contexts":"Following the adaptation stage, we train the entire model except for the visual encoder, consistent with previous approaches [58, 64]","reference_id":"MTQzNTkwNjA5NDY4MTI4MzUwNTk="} +{"index":1962,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05525","reference_contexts":"During this stage, we also unfreeze the SigLIP encoder parameters [64, 90, 97]","reference_id":"MTI1OTEyNTE3Mjk4NzkxMDgxOTc="} +{"index":1963,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05525","reference_contexts":"Building upon powerful LLMs [7, 91, 92] , recent MLLMs [2, 15, 49, 56, 58, 64] have demonstrated exceptional multimodal understanding capabilities","reference_id":"ODU3NDc4NDgyNDQ5NTQ2NzI0NQ=="} +{"index":1964,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05525","reference_contexts":"By leveraging large-scale multimodal datasets and increasingly powerful LLMs, this architecture has facilitated the development of advanced multimodal models capable of addressing a wide range of vision-language tasks [4, 47, 56, 64] .","reference_id":"NjQzOTE1NDIxNTM1MTg3MzMxMA=="} +{"index":1965,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2403.05525","reference_contexts":"Our framework builds upon an enhanced version foot_0 of DeepSeek-LLM (1.3B) [7, 64]","reference_id":"Mjg4MjE1MTg2OTQ2Mjk1MDAyOQ=="} +{"index":1966,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2401.08740","reference_contexts":"Importantly, our architecture differs from previous studies [66, 72] examined in [103] due to our incorporation of LLM and an additional skip connection between 𝑔 𝑒𝑛𝑐 and 𝑔 𝑑𝑒𝑐","reference_id":"MTM1MDY2MjQxMjg1NzIzNjQ0Njk="} +{"index":1967,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2303.09319","reference_contexts":"Recent years have witnessed remarkable progress in visual generation through diffusion models [32, 83] , leading to impressive models like [67, 73, [76] [77] [78] [79]","reference_id":"MjQzNjA2OTEwNDgyOTExMDAwNw=="} +{"index":1968,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2005.14165","reference_contexts":"Large language models (LLMs) have demonstrated remarkable capabilities in learning diverse knowledge and generalizing to new scenarios [1, 7, 8, 69, 91]","reference_id":"OTkzMzk4NjE1MjQ1NDA2NzMyMg=="} +{"index":1969,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2005.14165","reference_contexts":"After being trained on large-scale datasets, LLMs exhibit the ability to generalize across various tasks and follow diverse instructions [1, 8, 69]","reference_id":"MjY2OTc5ODczMTM0OTU3MDI2OQ=="} +{"index":1970,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2204.06125","reference_contexts":"JanusFlow achieves an overall score of 0.63, surpassing the previous unified framework and several generation specific models including SDXL [73] and DALL-E 2 [76]","reference_id":"NDU5MjI2NzQ0NDUxMzI1MzA1"} +{"index":1971,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2204.06125","reference_contexts":"Leveraging these capabilities, researchers have developed sophisticated models specialized in image comprehension [2, 15, 47, 49, 56, 58 ] and text-to-image generation [23, 73, 76, 79] .","reference_id":"MjAwMjY3MjAyNzk2ODk4Mjk4MA=="} +{"index":1972,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2204.06125","reference_contexts":"Recent years have witnessed remarkable progress in visual generation through diffusion models [32, 83] , leading to impressive models like [67, 73, [76] [77] [78] [79]","reference_id":"MjQzNjA2OTEwNDgyOTExMDAwNw=="} +{"index":1973,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.10328","reference_contexts":"We reformat the high-quality text-image pairs from [16, 82, 85] into an instruction format: \"User:\\n\\n Assistant:\".","reference_id":"MTAyOTQ2OTkzNjg4NTMzODc5MzE="} +{"index":1974,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.10328","reference_contexts":"Our image generation dataset combines high-quality images from [16, 21, 41, 43, 68, 71, 82, 85","reference_id":"MTQ4ODIzOTYzNTc3MzEwNzI3NTQ="} +{"index":1975,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.10328","reference_contexts":"The datasets used in the pre-training stage for understanding include DetailedCaption [20], SAM [41] , arXivQA [50] , DenseFusion-1M [51] , MMSci [53] , PixelProse [82] , re-captioned LAION-Aesthetics [16] , re-captioned Open Images V4 [43] , ShareGPT4V [11] , WikiHow [42] and WIT [84]","reference_id":"MTM5MDA4ODIzMzU0NjgzNTE3OTk="} +{"index":1976,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.10328","reference_contexts":"We filter the images in [16, 82] with aspect ratios and aesthetic scores, retaining approximately 20% of the original datasets","reference_id":"ODQ3MDk3MTMzNDg4MjY5NDUwMA=="} +{"index":1977,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.06525","reference_contexts":"Alternative approaches [88, 97, 99, 100, 108] propose training a single LLM for both tasks, typically incorporating either diffusion models [32, 83] or vector-quantized autoregressive models [22, 86] .","reference_id":"MTM1NjI1ODIwNjM4NzYxODkwMzg="} +{"index":1978,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.06525","reference_contexts":"Many of these methods employ vector-quantization [22, 86] to convert images into discrete tokens, enabling unified autoregressive processing [88, 97]","reference_id":"Mzk2MzEyNDc0OTc2OTUwNDc2"} +{"index":1979,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2405.09818","reference_contexts":"Alternative approaches [88, 97, 99, 100, 108] propose training a single LLM for both tasks, typically incorporating either diffusion models [32, 83] or vector-quantized autoregressive models [22, 86] .","reference_id":"MTM1NjI1ODIwNjM4NzYxODkwMzg="} +{"index":1980,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2405.09818","reference_contexts":"Another line of work [88, 97, 99, 100, 108] aim to train a single LLM for both tasks","reference_id":"MTA0ODQyNjA2NTc3Nzc3MDgzOA=="} +{"index":1981,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2405.09818","reference_contexts":"According to common practice [88, 97, 100] , we add special token |BOI| before the image and |EOI| after the image to help the model locate the image embeddings in the sequence.","reference_id":"MzMyNjQ1MzI5NDcyMDEwNjU3NQ=="} +{"index":1982,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2312.11805","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1983,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.16860","reference_contexts":"During this stage, we also unfreeze the SigLIP encoder parameters [64, 90, 97]","reference_id":"MTI1OTEyNTE3Mjk4NzkxMDgxOTc="} +{"index":1984,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2302.13971","reference_contexts":"Large language models (LLMs) have demonstrated remarkable capabilities in learning diverse knowledge and generalizing to new scenarios [1, 7, 8, 69, 91]","reference_id":"OTkzMzk4NjE1MjQ1NDA2NzMyMg=="} +{"index":1985,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2302.13971","reference_contexts":"Building upon powerful LLMs [7, 91, 92] , recent MLLMs [2, 15, 49, 56, 58, 64] have demonstrated exceptional multimodal understanding capabilities","reference_id":"ODU3NDc4NDgyNDQ5NTQ2NzI0NQ=="} +{"index":1986,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2307.09288","reference_contexts":"Building upon powerful LLMs [7, 91, 92] , recent MLLMs [2, 15, 49, 56, 58, 64] have demonstrated exceptional multimodal understanding capabilities","reference_id":"ODU3NDc4NDgyNDQ5NTQ2NzI0NQ=="} +{"index":1987,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2409.12191","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1988,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2409.18869","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":1989,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"During this stage, we also unfreeze the SigLIP encoder parameters [64, 90, 97]","reference_id":"MTI1OTEyNTE3Mjk4NzkxMDgxOTc="} +{"index":1990,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"However, recent work on unified autoregressive models has shown this shared encoder design to be suboptimal [97] , particularly in models that generate images through autoregression on vector-quantized tokens","reference_id":"ODUxOTAyNDE4ODIyNTM1NjQ3Mw=="} +{"index":1991,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"The results prove that the rectified flow is able to improve the quality of generated images over autoregressive models such as Janus [97] .","reference_id":"OTc1MjMwNDAzMzAxODM4NzQxNA=="} +{"index":1992,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"Another line of work [88, 97, 99, 100, 108] aim to train a single LLM for both tasks","reference_id":"MTA0ODQyNjA2NTc3Nzc3MDgzOA=="} +{"index":1993,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"According to common practice [88, 97, 100] , we add special token |BOI| before the image and |EOI| after the image to help the model locate the image embeddings in the sequence.","reference_id":"MzMyNjQ1MzI5NDcyMDEwNjU3NQ=="} +{"index":1994,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"We follow Janus [97] to construct the training data","reference_id":"Mzc0OTAxOTA2MTU0OTIzNzA2Mg=="} +{"index":1995,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"Alternative approaches [88, 97, 99, 100, 108] propose training a single LLM for both tasks, typically incorporating either diffusion models [32, 83] or vector-quantized autoregressive models [22, 86] .","reference_id":"MTM1NjI1ODIwNjM4NzYxODkwMzg="} +{"index":1996,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"Many of these methods employ vector-quantization [22, 86] to convert images into discrete tokens, enabling unified autoregressive processing [88, 97]","reference_id":"Mzk2MzEyNDc0OTc2OTUwNDc2"} +{"index":1997,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2410.13848","reference_contexts":"This version, trained on an expanded text corpus compared to the one in Janus [97] , has been demonstrated to possess better performance on multiple-choice benchmarks (e.g., MMBench [63] and SEED Bench [46] )","reference_id":"NDc4NzIzMDI5NDA4MTYzMjc4Ng=="} +{"index":1998,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2409.04429","reference_contexts":"Alternative approaches [88, 97, 99, 100, 108] propose training a single LLM for both tasks, typically incorporating either diffusion models [32, 83] or vector-quantized autoregressive models [22, 86] .","reference_id":"MTM1NjI1ODIwNjM4NzYxODkwMzg="} +{"index":1999,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2409.04429","reference_contexts":"Another line of work [88, 97, 99, 100, 108] aim to train a single LLM for both tasks","reference_id":"MTA0ODQyNjA2NTc3Nzc3MDgzOA=="} +{"index":2000,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.12528","reference_contexts":"Previous approaches that unify autoregressive generation and diffusion models within a joint LLM training framework [100, 108] employ identical encoders ( 𝑓 𝑒𝑛𝑐 and 𝑔 𝑒𝑛𝑐 ) for both understanding and generation tasks","reference_id":"NzU5NjQ4MTk3ODEwMDcxMjA5OQ=="} +{"index":2001,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.12528","reference_contexts":"[100] leverages MAGVIT-v2 [102] to encode image patches into discrete tokens for both tasks.","reference_id":"MzU3MzU2MzM4NjkyNjcyMzQ3"} +{"index":2002,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.12528","reference_contexts":"Another line of work [88, 97, 99, 100, 108] aim to train a single LLM for both tasks","reference_id":"MTA0ODQyNjA2NTc3Nzc3MDgzOA=="} +{"index":2003,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.12528","reference_contexts":"According to common practice [88, 97, 100] , we add special token |BOI| before the image and |EOI| after the image to help the model locate the image embeddings in the sequence.","reference_id":"MzMyNjQ1MzI5NDcyMDEwNjU3NQ=="} +{"index":2004,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.12528","reference_contexts":"Unlike previous approaches that employ various attention masking strategies [100, 108] , we found that causal attention suffices, as our preliminary experiments showed no performance benefits from alternative masking schemes","reference_id":"OTEyMzYwMTk5MTI2MzE0NTY3OA=="} +{"index":2005,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.12528","reference_contexts":"Alternative approaches [88, 97, 99, 100, 108] propose training a single LLM for both tasks, typically incorporating either diffusion models [32, 83] or vector-quantized autoregressive models [22, 86] .","reference_id":"MTM1NjI1ODIwNjM4NzYxODkwMzg="} +{"index":2006,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.12528","reference_contexts":"Compared to similar approaches [100, 107, 108] , JanusFlow offers three key advantages: (i) a simple yet effective generation process using rectified flow, (ii) enhanced performance through decoupled vision encoders that resolve inter-task conflicts, and (iii) improved generation quality through representation alignment regularization, enabled by our decoupled encoder design.","reference_id":"MTY3OTYyNzI5NjY2MDU0NTQzODU="} +{"index":2007,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2405.19335","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2008,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2409.16280","reference_contexts":"Compared to similar approaches [100, 107, 108] , JanusFlow offers three key advantages: (i) a simple yet effective generation process using rectified flow, (ii) enhanced performance through decoupled vision encoders that resolve inter-task conflicts, and (iii) improved generation quality through representation alignment regularization, enabled by our decoupled encoder design.","reference_id":"MTY3OTYyNzI5NjY2MDU0NTQzODU="} +{"index":2009,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.11039","reference_contexts":"Previous approaches that unify autoregressive generation and diffusion models within a joint LLM training framework [100, 108] employ identical encoders ( 𝑓 𝑒𝑛𝑐 and 𝑔 𝑒𝑛𝑐 ) for both understanding and generation tasks","reference_id":"NzU5NjQ4MTk3ODEwMDcxMjA5OQ=="} +{"index":2010,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.11039","reference_contexts":"B, following a design similar to Transfusion [108] , we implement shared ConvNeXt blocks in the SDXL-VAE latent space for both understanding and generation encoders","reference_id":"MTA0NjQwNjMzNTk5MDA1MDE0MTY="} +{"index":2011,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.11039","reference_contexts":"Another line of work [88, 97, 99, 100, 108] aim to train a single LLM for both tasks","reference_id":"MTA0ODQyNjA2NTc3Nzc3MDgzOA=="} +{"index":2012,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.11039","reference_contexts":"Unlike previous approaches that employ various attention masking strategies [100, 108] , we found that causal attention suffices, as our preliminary experiments showed no performance benefits from alternative masking schemes","reference_id":"OTEyMzYwMTk5MTI2MzE0NTY3OA=="} +{"index":2013,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.11039","reference_contexts":"Alternative approaches [88, 97, 99, 100, 108] propose training a single LLM for both tasks, typically incorporating either diffusion models [32, 83] or vector-quantized autoregressive models [22, 86] .","reference_id":"MTM1NjI1ODIwNjM4NzYxODkwMzg="} +{"index":2014,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.11039","reference_contexts":"[108] performs both tasks in the same VAE latent space using a shared U-Net or linear encoder, while Xie et al","reference_id":"NDk2ODU2MzA3MTg1NDU1NDQxNQ=="} +{"index":2015,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2408.11039","reference_contexts":"Compared to similar approaches [100, 107, 108] , JanusFlow offers three key advantages: (i) a simple yet effective generation process using rectified flow, (ii) enhanced performance through decoupled vision encoders that resolve inter-task conflicts, and (iii) improved generation quality through representation alignment regularization, enabled by our decoupled encoder design.","reference_id":"MTY3OTYyNzI5NjY2MDU0NTQzODU="} +{"index":2016,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2401.02330","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2017,"citer_arxiv_id":"2411.07975","cited_arxiv_id":"2406.18583","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2018,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2404.14219","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2019,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2406.11704","reference_contexts":"For comparison, we included baseline models such as Nemotron-4-340B-Reward (Adler et al., 2024) , Llama-3.1-Nemotron-70B-Reward (Wang et al., 2024c) , and Athene-RM-70B (Frick et al., 2024a)","reference_id":"MTU5MDc0NTUyMDk3NTc0MDgwNzc="} +{"index":2020,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2311.16867","reference_contexts":"While previous studies (Dubey et al., 2024; Almazrouei et al., 2023; Hoffmann et al., 2022) primarily used scaling laws to determine optimal model sizes given compute budgets, we leverage them to identify optimal hyperparameters across model architectures","reference_id":"MTc4MTYwMDkzMjI1MzYyODM2NjQ="} +{"index":2021,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2108.07732","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2022,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2309.16609","reference_contexts":"For tokenization, we utilize Qwen's tokenizer (Bai et al., 2023) , which implements byte-level byte-pair encoding (BBPE, Brown et al., 2020; Wang et al., 2020; Sennrich et al., 2016) with a vocabulary of 151,643 regular tokens","reference_id":"MTQ5MTM0MDk4NTc0Mzg4OTMyNTI="} +{"index":2023,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2309.16609","reference_contexts":"Besides the fast development of model capabilities, the recent two years have witnessed a burst of open (open-weight) large language models in the LLM community, for example, the Llama series (Touvron et al., 2023a; b; Dubey et al., 2024) , Mistral series (Jiang et al., 2023a; 2024a) , and our Qwen series (Bai et al., 2023; Yang et al., 2024a; Qwen Team, 2024a; Hui et al., 2024; Qwen Team, 2024c; Yang et al., 2024b)","reference_id":"MjY0Njg4MTM3NjcxOTM0NTI5Nw=="} +{"index":2024,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2309.16609","reference_contexts":"The sparks of artificial general intelligence (AGI) are increasingly visible through the fast development of large foundation models, notably large language models (LLMs) (Brown et al., 2020; OpenAI, 2023; 2024a; Gemini Team, 2024; Anthropic, 2023a; b; 2024; Bai et al., 2023; Yang et al., 2024a; Touvron et al., 2023a; b; Dubey et al., 2024)","reference_id":"NTEzMzc5MzU3NzMxMTk1NTkyNw=="} +{"index":2025,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2107.03374","reference_contexts":"It demonstrates significant improvements across various tasks, achieving 74.2 on general benchmarks like MMLU (Hendrycks et al., 2021a) , 49.8 on math challenges such as MATH (Hendrycks et al., 2021b) , and 57.9 on coding tasks like HumanEval (Chen et al., 2021) .","reference_id":"NjYwOTA4ODE1MjA0NzM1ODA4NA=="} +{"index":2026,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"1803.05457","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2027,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2110.14168","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2028,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2401.06066","reference_contexts":"Following the approaches demonstrated in Qwen1.5-MoE (Yang et al., 2024a) , we implement fine-grained expert segmentation (Dai et al., 2024) and shared experts routing (Rajbhandari et al., 2022; Dai et al., 2024)","reference_id":"MTQxMTMyODg4NjQ3NTEwMTg4NTY="} +{"index":2029,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2407.21783","reference_contexts":"Besides the fast development of model capabilities, the recent two years have witnessed a burst of open (open-weight) large language models in the LLM community, for example, the Llama series (Touvron et al., 2023a; b; Dubey et al., 2024) , Mistral series (Jiang et al., 2023a; 2024a) , and our Qwen series (Bai et al., 2023; Yang et al., 2024a; Qwen Team, 2024a; Hui et al., 2024; Qwen Team, 2024c; Yang et al., 2024b)","reference_id":"MjY0Njg4MTM3NjcxOTM0NTI5Nw=="} +{"index":2030,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2407.21783","reference_contexts":"Qwen2.5-7B For 7B-level models, we focus on comparing Qwen2.5-7B with other leading 7B+ models, including Mistral-7B (Jiang et al., 2023a) , Llama3-8B (Dubey et al., 2024) , Gemma2-9B (Gemma Team et al., 2024) , and our predecessor, Qwen2-7B (Yang et al., 2024a)","reference_id":"MzA2NzE0NjUyMDQ1NjExNDQwNw=="} +{"index":2031,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2407.21783","reference_contexts":"While previous studies (Dubey et al., 2024; Almazrouei et al., 2023; Hoffmann et al., 2022) primarily used scaling laws to determine optimal model sizes given compute budgets, we leverage them to identify optimal hyperparameters across model architectures","reference_id":"MTc4MTYwMDkzMjI1MzYyODM2NjQ="} +{"index":2032,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2407.21783","reference_contexts":"Datasets Gemma2-2B Phi3.5-Mini MiniCPM3-4B Qwen2.5-3B Non-Emb Params 2.0B 3.6B 4.0B 2.8B General Tasks MMLU-Pro 26.7 47.5 43.0 43.7 MMLU-redux 51.9 67.7 59.9 64.4 LiveBench 0831 20.1 27.4 27.6 26.8 Mathematics & Science Tasks GPQA 29.3 27.2 31.3 30.3 MATH 26.6 48.5 46.6 65.9 GSM8K 63.2 86.2 81.1 86.7 Coding Tasks HumanEval 68.9 72.6 74.4 74.4 MBPP 74.9 63.2 72.5 72.7 MultiPL-E 30.5 47.2 49.1 60.2 LiveCodeBench 5.8 15.8 23.8 19.9 Alignment Tasks IFEval 51.0 52.1 68.4 58.2 et al., 2024), Llama3.1-405B-Instruct (Dubey et al., 2024) , and our previous 72B version, Qwen2-72B-Instruct (Yang et al., 2024a)","reference_id":"MTA2NDQ4OTgwOTcyOTEyMzk0ODk="} +{"index":2033,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2407.21783","reference_contexts":"The sparks of artificial general intelligence (AGI) are increasingly visible through the fast development of large foundation models, notably large language models (LLMs) (Brown et al., 2020; OpenAI, 2023; 2024a; Gemini Team, 2024; Anthropic, 2023a; b; 2024; Bai et al., 2023; Yang et al., 2024a; Touvron et al., 2023a; b; Dubey et al., 2024)","reference_id":"NTEzMzc5MzU3NzMxMTk1NTkyNw=="} +{"index":2034,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2406.04127","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2035,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2408.00118","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2036,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2203.15556","reference_contexts":"While previous studies (Dubey et al., 2024; Almazrouei et al., 2023; Hoffmann et al., 2022) primarily used scaling laws to determine optimal model sizes given compute budgets, we leverage them to identify optimal hyperparameters across model architectures","reference_id":"MTc4MTYwMDkzMjI1MzYyODM2NjQ="} +{"index":2037,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2203.15556","reference_contexts":"We develop scaling laws for hyper-parameter based on the pre-training data of Qwen2.5 (Hoffmann et al., 2022; Kaplan et al., 2020)","reference_id":"MTI3Mzg4ODU1ODYwNjQ0NTUyMDk="} +{"index":2038,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2404.06395","reference_contexts":"5-mini-instruct (Abdin et al., 2024) and MiniCPM3-4B-Instruct (Hu et al., 2024 ) models, surpasses them in mathematics and coding tasks, as shown in Table 9","reference_id":"ODU4MDU0MjIxMjk0OTg4OTMwMw=="} +{"index":2039,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2403.07974","reference_contexts":"Specifically, we evaluate on MMLU-Pro, MMLU-redux and LiveBench 0831 (White et al., 2024) for general evaluation, GPQA, GSM8K and MATH for science and mathematics, HumanEval, MBPP, MultiPL-E and LiveCodeBench 2305-2409 (Jain et al., 2024) for coding, IFEval (Zhou et al., 2023) foot_1 for instruction following","reference_id":"MTI5MDIzMzA2ODUyNjgyOTAwMTE="} +{"index":2040,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2310.06825","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2041,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2401.04088","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2042,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2001.08361","reference_contexts":"We develop scaling laws for hyper-parameter based on the pre-training data of Qwen2.5 (Hoffmann et al., 2022; Kaplan et al., 2020)","reference_id":"MTI3Mzg4ODU1ODYwNjQ0NTUyMDk="} +{"index":2043,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2403.13787","reference_contexts":"Our assessment benchmarks encompass Reward Bench (Lambert et al., 2024) , RMB (Zhou et al., 2024) , PPE (Frick et al., 2024b) , and an internally collected out-of-domain Chinese human preference benchmark (Human-Preference-Chinese) to provide a comprehensive analysis","reference_id":"ODg5NzQ1MzkzNDI5OTIzNjEwMA=="} +{"index":2044,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2006.16668","reference_contexts":"Additionally, we also release the proprietary models of Mixture-of-Experts (MoE, Lepikhin et al., 2020; Fedus et al., 2022; Zoph et al., 2022) , namely Qwen2.5-Turbo and Qwen2.5-Plus foot_0 , which performs competitively against GPT-4o-mini and GPT-4o respectively.","reference_id":"MzA0MDc4NTcwNjY0MzI0NTk4NA=="} +{"index":2045,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2406.11939","reference_contexts":"Additionally, we assess the performance of human preference alignment and instruction following by evaluating on benchmarks including MT-Bench (Zheng et al., 2023) and Arena-Hard (Li et al., 2024)","reference_id":"NTg4OTUxNDg3NjAxMDA4MDIzMA=="} +{"index":2046,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2303.08774","reference_contexts":"The sparks of artificial general intelligence (AGI) are increasingly visible through the fast development of large foundation models, notably large language models (LLMs) (Brown et al., 2020; OpenAI, 2023; 2024a; Gemini Team, 2024; Anthropic, 2023a; b; 2024; Bai et al., 2023; Yang et al., 2024a; Touvron et al., 2023a; b; Dubey et al., 2024)","reference_id":"NTEzMzc5MzU3NzMxMTk1NTkyNw=="} +{"index":2047,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2309.00071","reference_contexts":"To enhance our models' ability to process longer sequences during inference, we implement two key strategies: YARN (Peng et al., 2023) and Dual Chunk Attention (DCA, An et al., 2024)","reference_id":"NDQ5MDczMzM4Nzk2OTM5MTQ1"} +{"index":2048,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2311.12022","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2049,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2402.03300","reference_contexts":"In our online reinforcement learning (RL) framework, we employ Group Relative Policy Optimization (GRPO, Shao et al., 2024)","reference_id":"MzgzOTY2NzM5MjE4MjczMjE5MQ=="} +{"index":2050,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2402.03300","reference_contexts":"The post-training data amounts to 1 million examples, across the stage of supervised finetuning (SFT, Ouyang et al., 2022) , direct preference optimization (DPO, Rafailov et al., 2023) , and group relative policy optimization (GRPO, Shao et al., 2024) .","reference_id":"NzkzNTM1Mjc1MzU0MzIzMjMxNQ=="} +{"index":2051,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2302.13971","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2052,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2307.09288","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2053,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2406.01574","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2054,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2406.19314","reference_contexts":"Specifically, we evaluate on MMLU-Pro, MMLU-redux and LiveBench 0831 (White et al., 2024) for general evaluation, GPQA, GSM8K and MATH for science and mathematics, HumanEval, MBPP, MultiPL-E and LiveCodeBench 2305-2409 (Jain et al., 2024) for coding, IFEval (Zhou et al., 2023) foot_1 for instruction following","reference_id":"MTI5MDIzMzA2ODUyNjgyOTAwMTE="} +{"index":2055,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2309.16039","reference_contexts":"Concurrently, we increase the base frequency of RoPEfrom 10,000 to 1,000,000 using the ABF technique (Xiong et al., 2023) .","reference_id":"MTI1NzUyMTY3ODkyNjA5MzU2MTY="} +{"index":2056,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2407.10671","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2057,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2403.04652","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2058,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2402.05136","reference_contexts":"We utilize three benchmarks to evaluate long context capabilities of Qwen2.5 models: RULER (Hsieh et al., 2024) , LV-Eval (Yuan et al., 2024) , and Longbench-Chat (Bai et al., 2024)","reference_id":"MTE2Nzc4MDQ2NjkzMzM0MzU4NTY="} +{"index":2059,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2308.01825","reference_contexts":"Additionally, we assess the performance of human preference alignment and instruction following by evaluating on benchmarks including MT-Bench (Zheng et al., 2023) and Arena-Hard (Li et al., 2024)","reference_id":"NTg4OTUxNDg3NjAxMDA4MDIzMA=="} +{"index":2060,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2308.01825","reference_contexts":"To ensure high-quality reasoning, we employ rejection sampling (Yuan et al., 2023) along with reward modeling and annotated answers for guidance, producing step-by-step reasoning process.","reference_id":"MTIwNzU2NDM5MjYxNjQwMzU0MjM="} +{"index":2061,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2311.07911","reference_contexts":"Specifically, we evaluate on MMLU-Pro, MMLU-redux and LiveBench 0831 (White et al., 2024) for general evaluation, GPQA, GSM8K and MATH for science and mathematics, HumanEval, MBPP, MultiPL-E and LiveCodeBench 2305-2409 (Jain et al., 2024) for coding, IFEval (Zhou et al., 2023) foot_1 for instruction following","reference_id":"MTI5MDIzMzA2ODUyNjgyOTAwMTE="} +{"index":2062,"citer_arxiv_id":"2412.15115","cited_arxiv_id":"2202.08906","reference_contexts":"Additionally, we also release the proprietary models of Mixture-of-Experts (MoE, Lepikhin et al., 2020; Fedus et al., 2022; Zoph et al., 2022) , namely Qwen2.5-Turbo and Qwen2.5-Plus foot_0 , which performs competitively against GPT-4o-mini and GPT-4o respectively.","reference_id":"MzA0MDc4NTcwNjY0MzI0NTk4NA=="} +{"index":2063,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2206.02779","reference_contexts":"Similar work has also been proposed by Blended Diffusion [2, 3] , which combines the features of each step for image blending","reference_id":"MTMxNTUyMjcxNzAxMjg5ODMyMzk="} +{"index":2064,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2206.02779","reference_contexts":"4 , we obtain a binary mask M t by thresholding the cross-attention map of the edited words during inversion by a constant τ [2, 3]","reference_id":"NDUxNjExNzY1MzgxNTc4NDUwMg=="} +{"index":2065,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2206.02779","reference_contexts":"Previous or concurrent diffusion-based editing methods [2, 3, 6, 16, 37, 47] majorly work on images","reference_id":"NTU2NTk4NzQwNzE0ODc5NTI5OA=="} +{"index":2066,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2206.02779","reference_contexts":"In this paper, we propose FateZero, a simple yet effective method for zero-shot video editing since we do not need to train for each target prompt individually [4, 24, 28] and have no user-specific mask [2, 3]","reference_id":"NTUwNTg2NjEyOTc4NjY5NzQ3Ng=="} +{"index":2067,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2210.11427","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2068,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2302.03011","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2069,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.13221","reference_contexts":"Diffusion-based models [19] can generate diverse and high-quality images [39, 41, 43] and videos [15, 18, 44, 55] through text prompts","reference_id":"MTYxNTgwOTA2NDIzNzYxODQ3MjY="} +{"index":2070,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.13221","reference_contexts":"Inspired by the casual self-attention [15, 20, 49, 50] and recent one-shot video generation method [51] , we reshape the original self-attention to Spatial-Temporal Self-Attention without changing pretrained weights","reference_id":"MTU0MDA0MzkzMDI2MTkzMjQyOTI="} +{"index":2071,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"Firstly, instead of solely relying on inversion and generation [16, 34, 47] , we adopt a different approach by storing all the self and cross-attention maps at every step of the inversion process","reference_id":"NTg2NzAwMDc4NjAxNTQyODg2Ng=="} +{"index":2072,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"Recently, image editing methods [16, 34, 37, 47]","reference_id":"NjQzMTI0NTY1Nzg1MTMwNjcyMA=="} +{"index":2073,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"In contrast, if we fuse full-resolution self-attention as in the previous work [16] , the shape editing ability of the framework can be severely degraded so that the geometry of generated car resembles the input video, especially in the last few frames","reference_id":"MTgwNjMyNjYzMDA3MjgwNjE5NDA="} +{"index":2074,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"(2) The Neural Layered Atlas [24] (NLA) based method is combined with keyframeediting via state-of-the-art image editing methods [16, 34] .","reference_id":"MTY3ODM3NzM5ODU2NjM3MzA2NDg="} +{"index":2075,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"(3) Frame-wise Null-text optimization [34] and then edit by prompt2prompt [16]","reference_id":"MTAyNjkyNjU5MjYwODY3MTgyMjI="} +{"index":2076,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"Previous or concurrent diffusion-based editing methods [2, 3, 6, 16, 37, 47] majorly work on images","reference_id":"NTU2NTk4NzQwNzE0ODc5NTI5OA=="} +{"index":2077,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"Prompt-to-Prompt [16] proves that images can be edited via reweighting the cross-attention map of different prompts","reference_id":"MTYwOTc2NDY3MDE3MjEyOTE3MjU="} +{"index":2078,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"In function ATT-FUSION, we inject the cross-attention maps of the unchanged part of the prompt similar to Prompt-to-Prompt [16] .","reference_id":"MTUwODY2MjY3ODEwMjU4OTIyMTg="} +{"index":2079,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"Therefore, previous methods require optimization of text-embedding [16] or other regularization [37] .","reference_id":"NTcyMTE4NzQ1Nzg1NjY4NjE1Mg=="} +{"index":2080,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"Inspired by the fact that the cross-attention map provides the semantic layout of the image [16] , as visualized in the second row of Fig","reference_id":"MTE4NjQ1NjM1NTQ4OTk0ODY4NzA="} +{"index":2081,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2208.01626","reference_contexts":"3 shows a toy comparison example between our attention fusion method and the typical method with simply inversion and then generation as in [16, 34] for image editing","reference_id":"MTMxNjk5ODE1MzMzMzM4MjA4MzY="} +{"index":2082,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2210.02303","reference_contexts":"Diffusion-based models [19] can generate diverse and high-quality images [39, 41, 43] and videos [15, 18, 44, 55] through text prompts","reference_id":"MTYxNTgwOTA2NDIzNzYxODQ3MjY="} +{"index":2083,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2210.02303","reference_contexts":"Current methods rely on the larger cascaded models [18, 44] and dataset","reference_id":"NzQ1NDA2MDY4MjE2OTk2MTY0Mg=="} +{"index":2084,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2210.02303","reference_contexts":"Dreamix [35] finetunes a stronger diffusion-based video model [18] for editing with stronger generative priors","reference_id":"MTIyNTQxMzMxMzM1MDQ2NzQwMDA="} +{"index":2085,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2210.02303","reference_contexts":"First, there are no publicly available generic text-to-video models [18, 44]","reference_id":"MTY1MTk1MjAxOTE5Njk0MDcxODM="} +{"index":2086,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2204.03458","reference_contexts":"Inspired by the casual self-attention [15, 20, 49, 50] and recent one-shot video generation method [51] , we reshape the original self-attention to Spatial-Temporal Self-Attention without changing pretrained weights","reference_id":"MTU0MDA0MzkzMDI2MTkzMjQyOTI="} +{"index":2087,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2210.09276","reference_contexts":"There are also some methods to achieve better editing ability via optimization [25, 34]","reference_id":"NjY1NDMzNDQ4NzM5MTU1NzA0MQ=="} +{"index":2088,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2210.09276","reference_contexts":"Based on this pipeline, several methods have been proposed in terms of cross-attention guidance [37] , plug-and-play feature [47] , and optimization [25, 34] .","reference_id":"NTIzMDc3OTIzMTIxMzEzMzQ3MQ=="} +{"index":2089,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"1312.6114","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2090,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2301.13173","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2091,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2108.01073","reference_contexts":"SDEdit [32] generates content for a new prompt by adding noise to the image first","reference_id":"MjYwMDkxOTEwOTk0MDM1NTgwMQ=="} +{"index":2092,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2108.01073","reference_contexts":"Directly applying the image editing methods [32, 34] to the video will show obverse flickering","reference_id":"NjAxNDMzMDMzMzMzMDEzNjgyNQ=="} +{"index":2093,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2108.01073","reference_contexts":"The editing result of framewise SDEdit [32] can not be localized and varies a lot among different frames","reference_id":"MTc2MTg0MDE1MDQzMTQ3MDYyNDU="} +{"index":2094,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2108.01073","reference_contexts":"( 4 ) Frame-wise zero-shot editing using SDEdit [32]","reference_id":"MzIyMTcxMDI2NTU4NzUzNjM5NA=="} +{"index":2095,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"Firstly, instead of solely relying on inversion and generation [16, 34, 47] , we adopt a different approach by storing all the self and cross-attention maps at every step of the inversion process","reference_id":"NTg2NzAwMDc4NjAxNTQyODg2Ng=="} +{"index":2096,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"Directly applying the image editing methods [32, 34] to the video will show obverse flickering","reference_id":"NjAxNDMzMDMzMzMzMDEzNjgyNQ=="} +{"index":2097,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"Recently, image editing methods [16, 34, 37, 47]","reference_id":"NjQzMTI0NTY1Nzg1MTMwNjcyMA=="} +{"index":2098,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"1, the proposed zero-shot method achieves the best temporal consistency against baselines and shows a comparable frame-wise editing accuracy as the pre-frame optimization method [34]","reference_id":"MTk2ODAyMzExMjUwODI0MDYzOQ=="} +{"index":2099,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"(2) The Neural Layered Atlas [24] (NLA) based method is combined with keyframeediting via state-of-the-art image editing methods [16, 34] .","reference_id":"MTY3ODM3NzM5ODU2NjM3MzA2NDg="} +{"index":2100,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"Based on this pipeline, several methods have been proposed in terms of cross-attention guidance [37] , plug-and-play feature [47] , and optimization [25, 34] .","reference_id":"NTIzMDc3OTIzMTIxMzEzMzQ3MQ=="} +{"index":2101,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"(3) Frame-wise Null-text optimization [34] and then edit by prompt2prompt [16]","reference_id":"MTAyNjkyNjU5MjYwODY3MTgyMjI="} +{"index":2102,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"There are also some methods to achieve better editing ability via optimization [25, 34]","reference_id":"NjY1NDMzNDQ4NzM5MTU1NzA0MQ=="} +{"index":2103,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"For CLIP-Score, we achieve the best temporal consistency and comparable framewise editing accuracy against an optimization-based image editing method [34] .","reference_id":"MTA0MzEwMDY0Mzc5OTQzNjExMjI="} +{"index":2104,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.09794","reference_contexts":"3 shows a toy comparison example between our attention fusion method and the typical method with simply inversion and then generation as in [16, 34] for image editing","reference_id":"MTMxNjk5ODE1MzMzMzM4MjA4MzY="} +{"index":2105,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2302.01329","reference_contexts":"Thus, a framework based on image models can be more valuable than on video ones [35] , thanks to the various open-sourced image models in the community [1, 36, 41, 53]","reference_id":"NDYwODc3MjM3MTk5NjE4MjAyMg=="} +{"index":2106,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2302.01329","reference_contexts":"Dreamix [35] finetunes a stronger diffusion-based video model [18] for editing with stronger generative priors","reference_id":"MTIyNTQxMzMxMzM1MDQ2NzQwMDA="} +{"index":2107,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2302.08453","reference_contexts":"Thus, a framework based on image models can be more valuable than on video ones [35] , thanks to the various open-sourced image models in the community [1, 36, 41, 53]","reference_id":"NDYwODc3MjM3MTk5NjE4MjAyMg=="} +{"index":2108,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2302.03027","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2109,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2204.06125","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2110,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2205.11487","reference_contexts":"Diffusion-based models [19] can generate diverse and high-quality images [39, 41, 43] and videos [15, 18, 44, 55] through text prompts","reference_id":"MTYxNTgwOTA2NDIzNzYxODQ3MjY="} +{"index":2111,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2205.11487","reference_contexts":"Recent models [41, 43] benefit from the stability of training diffusion-based model [19]","reference_id":"Njg3NTMzNDcwMzI0ODIzODAzMA=="} +{"index":2112,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2209.14792","reference_contexts":"Diffusion-based models [19] can generate diverse and high-quality images [39, 41, 43] and videos [15, 18, 44, 55] through text prompts","reference_id":"MTYxNTgwOTA2NDIzNzYxODQ3MjY="} +{"index":2113,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2209.14792","reference_contexts":"Current methods rely on the larger cascaded models [18, 44] and dataset","reference_id":"NzQ1NDA2MDY4MjE2OTk2MTY0Mg=="} +{"index":2114,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2209.14792","reference_contexts":"First, there are no publicly available generic text-to-video models [18, 44]","reference_id":"MTY1MTk1MjAxOTE5Njk0MDcxODM="} +{"index":2115,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"DDIM Inversion [45]","reference_id":"NjA0NjMwNzM5MTk2MDY5ODgzNQ=="} +{"index":2116,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"where α t is a parameter for noise scheduling [19, 45] Based on the ODE limit analysis of the diffusion process, DDIM inversion [7, 45] is proposed to map a clean latent z 0 back to a noised latent ẑT in revered steps t : 1 → T : Such that the inverted latent ẑT can reconstruct a latent ẑ0 (p src ) = DDIM(ẑ T , p src ) similar to the clean latent z 0 at classifier-free guidance scale s cf g = 1","reference_id":"MTAxNzY4NTYwMjM2MDE2MzMwNTY="} +{"index":2117,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"5 ) ability in test-time than simple DDIM inversion [45]","reference_id":"MTM4NTg3Mjk4ODUxMTM4MjE1NDU="} +{"index":2118,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"In this case, we compare our editing method with simple DDIM inversion [45] , where our method also achieves better performance in terms of editing ability, motion consistency, and temporal consistency","reference_id":"NDg1MDE2NjQwMzczMTE1MDY3NA=="} +{"index":2119,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"To edit real images, their methods utilize deterministic DDIM [45] for the image-to-noise inversion, and then, the inverted noise gradually generates the edited images under the condition of the target prompt","reference_id":"MzI5MjAzMjA0OTU5ODQ5NDQyMQ=="} +{"index":2120,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"(3) only holds in the limit of small steps [45, 46]","reference_id":"ODgyNzQ3ODg2NDY4NjYzMDA4Nw=="} +{"index":2121,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"While Tune-A-Video [51] with DDIM [45] ranks second in editing quality and image fidelity of Tab","reference_id":"OTQxNTc0NDAwOTU5NDcyOTA5OQ=="} +{"index":2122,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2010.02502","reference_contexts":"Inspired by the image editing methods and tune-a-video, our method can edit the style of the real-world video and images using the trained text-to-image model [41] and shows better object replacing performance than the one-shot finetuned video diffusion model [51] with simple DDIM inversion [45] in real videos (Fig","reference_id":"NTYwNjUxMDk3ODAwNTAxMzMyMw=="} +{"index":2123,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.12572","reference_contexts":"Firstly, instead of solely relying on inversion and generation [16, 34, 47] , we adopt a different approach by storing all the self and cross-attention maps at every step of the inversion process","reference_id":"NTg2NzAwMDc4NjAxNTQyODg2Ng=="} +{"index":2124,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.12572","reference_contexts":"Recently, image editing methods [16, 34, 37, 47]","reference_id":"NjQzMTI0NTY1Nzg1MTMwNjcyMA=="} +{"index":2125,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.12572","reference_contexts":"Based on this pipeline, several methods have been proposed in terms of cross-attention guidance [37] , plug-and-play feature [47] , and optimization [25, 34] .","reference_id":"NTIzMDc3OTIzMTIxMzEzMzQ3MQ=="} +{"index":2126,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.12572","reference_contexts":"Previous or concurrent diffusion-based editing methods [2, 3, 6, 16, 37, 47] majorly work on images","reference_id":"NTU2NTk4NzQwNzE0ODc5NTI5OA=="} +{"index":2127,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.12572","reference_contexts":"Plugand-play [47] gets the inversion noise and applies the denois-ing for feature reconstruction","reference_id":"MTAxNzk4MjIwNjM2NTkzNzg0Mg=="} +{"index":2128,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"(1) Tune-A-Video [51] overfits an inflated diffusion model on a single video to generate similar content","reference_id":"MTA5NjQ3MDUzOTA1MDkzNTExMjQ="} +{"index":2129,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"In this paper, we majorly use Stable Diffusion [41] and the video generation model based on Stable Diffusion (Tune-A-Video [51] ) for its popularity and generalization ability.","reference_id":"MTI5ODc3MDA5NzI3MjE5NjM5NTQ="} +{"index":2130,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"A train traveling down tracks next to a forest and a man on the side of the track → ..., Makoto Shinkai style Since there is no publicly-available generic video diffusion model, we perform the editing on the one-shot video diffusion model [51] instead","reference_id":"NjAwODE5NzYyODUwOTgyNzMyMQ=="} +{"index":2131,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"Recently, Tune-A-Video [51] overfits a single video for text-based video generation","reference_id":"MzY1MTQ0ODkwODU4ODQwMjY3Nw=="} +{"index":2132,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"1 present the result of difficult object shape editing, with a pretrained video model [51]","reference_id":"MTgyMDAxNzk4NTE5NTMxMzY4MjE="} +{"index":2133,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"Inspired by the casual self-attention [15, 20, 49, 50] and recent one-shot video generation method [51] , we reshape the original self-attention to Spatial-Temporal Self-Attention without changing pretrained weights","reference_id":"MTU0MDA0MzkzMDI2MTkzMjQyOTI="} +{"index":2134,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"Furthermore, we reform the self-attention blocks to the spatial-temporal attention blocks as in [51] to make the appearance more consistent","reference_id":"MTExNTY3NDg1MDk3MjM4MDg3NzI="} +{"index":2135,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"During shape editing, since the motion is produced by the one-shot video diffusion model [51] , it is difficult to generate totally new motion (e.g.,'swim'-→'fly' ) or very different shape (e.g.,'swan' -→'pterosaur')","reference_id":"MTcxNTY3NTQyMTc5ODQwMzE1NDc="} +{"index":2136,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"This is because the spatial-temporal self-attention maps represent the correspondence between frames and the temporal modeling ability of existing video diffusion model [51] is not satisfactory","reference_id":"NDE4MjY3NDM1OTg5NDA5NDY2Mw=="} +{"index":2137,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"For shape editing, we utilize the pretrained model of the specific video [51] at 100 iterations and fuse the attention at DDIM timestep t ∈ [0.5 × T, T ], giving more freedom for new shape generation","reference_id":"MTI3MjMxNjEzNzgzNDk3ODQ5OTA="} +{"index":2138,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"While Tune-A-Video [51] with DDIM [45] ranks second in editing quality and image fidelity of Tab","reference_id":"OTQxNTc0NDAwOTU5NDcyOTA5OQ=="} +{"index":2139,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2212.11565","reference_contexts":"We conduct the quantitative evaluation using the trained CLIP [33] model as previous methods [9, 37, 51]","reference_id":"MjI2MzUzNjQ0OTA5NTgwMTE1MQ=="} +{"index":2140,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.11018","reference_contexts":"Diffusion-based models [19] can generate diverse and high-quality images [39, 41, 43] and videos [15, 18, 44, 55] through text prompts","reference_id":"MTYxNTgwOTA2NDIzNzYxODQ3MjY="} +{"index":2141,"citer_arxiv_id":"2303.09535","cited_arxiv_id":"2211.11018","reference_contexts":"Differently, magic-video [55] and gen1 [9] initialize the model from textto-image [41] and generate the continuous contents through extra time-aware layers","reference_id":"MTU5MjI2MTU1MzM0Mjg0OTgyNzE="} +{"index":2142,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2311.15127","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2143,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2412.07774","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2144,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2208.01618","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2145,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2307.04725","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2146,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2410.00086","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2147,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2408.05939","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2148,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2104.08718","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2149,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2205.15868","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2150,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2106.09685","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2151,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2410.23775","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2152,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2409.17920","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2153,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2411.16318","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2154,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2307.01952","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2155,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2209.14988","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2156,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2306.17843","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2157,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2308.16512","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2158,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2411.06686","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2159,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2209.14792","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2160,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2010.02502","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2161,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2411.15098","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2162,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2401.07519","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2163,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2409.11340","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2164,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2408.06072","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2165,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2308.06721","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2166,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2311.15127","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2167,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2412.07774","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2168,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2208.01618","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2169,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2307.04725","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2170,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2410.00086","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2171,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2408.05939","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2172,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2104.08718","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2173,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2205.15868","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2174,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2106.09685","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2175,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2410.23775","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2176,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2409.17920","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2177,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2411.16318","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2178,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2307.01952","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2179,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2209.14988","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2180,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2306.17843","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2181,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2308.16512","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2182,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2411.06686","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2183,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2209.14792","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2184,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2010.02502","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2185,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2411.15098","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2186,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2401.07519","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2187,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2409.11340","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2188,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2408.06072","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2189,"citer_arxiv_id":"2503.07027","cited_arxiv_id":"2308.06721","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2190,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2112.10741","reference_contexts":"We observe that the DA-Score shows a significantly higher correlation with human evaluation ratings as opposed to prior works across varying number of subjects N ∈ [2, 5] in the input prompt","reference_id":"MzA5MzkxNTIwNDczODU1NDg2Nw=="} +{"index":2191,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2112.10741","reference_contexts":"To enforce, heavier reliance of generated outputs on the provided text, classifier-free guidance methods [2, 3, 23] have been proposed","reference_id":"MTc2NDA3ODMzOTAzODA4NDg2MDU="} +{"index":2192,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2112.10741","reference_contexts":"Most prior works [2, 3, 5, 9] typically use the cosine similarity between the text and image embedding from large-scale multi-modal models [9, [36] [37] [38] such as CLIP [9] , BLIP [10] , BLIP-2 [11] for evaluating the alignment scores","reference_id":"MTEyMzgyOTg3MDU5MDM3NjQwNDA="} +{"index":2193,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2112.10741","reference_contexts":"The field of text-to-image generation has made significant advancements with the recent advent of large-scale language-image (LLI) models [1] [2] [3] [4] [5]","reference_id":"OTk3NTIwMDIxMjUwNjMyOTc2MQ=="} +{"index":2194,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2205.11487","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2195,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2204.06125","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2196,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2206.10789","reference_contexts":"Similarly, [5, 19] explore the use of autoregressive models for zero-shot text to image generation","reference_id":"ODM3ODU5MjI3NjQ0NTU2MDQxNg=="} +{"index":2197,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2206.10789","reference_contexts":"Most prior works [2, 3, 5, 9] typically use the cosine similarity between the text and image embedding from large-scale multi-modal models [9, [36] [37] [38] such as CLIP [9] , BLIP [10] , BLIP-2 [11] for evaluating the alignment scores","reference_id":"MTEyMzgyOTg3MDU5MDM3NjQwNDA="} +{"index":2198,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2206.10789","reference_contexts":"We observe that the DA-Score shows a significantly higher correlation with human evaluation ratings as opposed to prior works across varying number of subjects N ∈ [2, 5] in the input prompt","reference_id":"MzA5MzkxNTIwNDczODU1NDg2Nw=="} +{"index":2199,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2206.10789","reference_contexts":"The field of text-to-image generation has made significant advancements with the recent advent of large-scale language-image (LLI) models [1] [2] [3] [4] [5]","reference_id":"OTk3NTIwMDIxMjUwNjMyOTc2MQ=="} +{"index":2200,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2206.10789","reference_contexts":"In particular, at any iteration k ∈ [1, 5] during the refinement process, we first compute both overall text-image similarity score Ω(I k , P) and assertion-level alignment scores u i (I k , P)","reference_id":"MjE3ODQyOTMwMTg1MjQwNjYw"} +{"index":2201,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Similar to [7, 8] , we use the official Stable Diffusion v1.4 model as the underlying pretrained text-to-image generation model while reporting results with all methods [6] [7] [8] (including the iterative refinement approach proposed in the main paper)","reference_id":"NDY5MzY3MzYyOTMwNzEzMzMx"} +{"index":2202,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Finally, we report results on the recently proposed text-to-text (T2T) similarity metric [7] which computes image-text similarity as the average cosine similarity between input prompt and captions generated (using BLIP) from the input image.","reference_id":"MTU2MzEyNzc2NzMxMjMyNzMz"} +{"index":2203,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"While the proposed iterative refinement approach shows better performance than previous works [6] [7] [8] , it still has some limitations","reference_id":"MTM0NjE4NDQ5NDg0NTg5MDM0ODk="} +{"index":2204,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"We also provide additional results comparing our method performance with Attend-and-Excite [7] in Sec","reference_id":"MTYxOTkxNDQyMjczMjczMjg1NTI="} +{"index":2205,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"The dataset consists an overall of 24960 human annotations on images generated using all methods [1, 6, 7] (including ours) across a diverse set of 4160 input prompts","reference_id":"MTA3NDE3NzM2NjMyNTg3NjY0NzQ="} +{"index":2206,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"In particular, for all methods [6, 7] and ours, we first collect a set of 4-5 human generated examples describing the desired outputs (e.g., main subjects for [7] ) for the input prompts","reference_id":"MTc2MTk3OTc0NDI2NzgyNjE3ODE="} +{"index":2207,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Furthermore, unlike prior works [7] which predominantly analyse the performance on relatively simple prompts with two subjects (e.g","reference_id":"NzY0NjYzMTMzMzQ1NjI0MTYwNA=="} +{"index":2208,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Through both qualitative and quantitative analysis, we find that the proposed iterative refinement process allows for generation of better aligned image outputs over prior works [6] [7] [8] while on average showing comparable inference times (Sec","reference_id":"OTg3MTczNjg5NDAyOTY2NzA4NQ=="} +{"index":2209,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"We also note that training-free methods such as ours or prior works [6] [7] [8] , can in-turn help improve the performance of human-feedback based methods by providing a better quality dataset for determining both aesthetics and content alignment of the generated images.","reference_id":"MjkwODM1NzMzNzcwNzE5MzE4MQ=="} +{"index":2210,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"9 ), we find that the our approach with only prompt-weighting is able to consistently perform on-par or better than Attend-and-Excite [7]","reference_id":"OTI0NjQ1OTcwODExNTMzNTQwMg=="} +{"index":2211,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Furthermore, unlike prior works [7] which predominantly analyse the performance on relatively simple prompts with two subjects (e.g., object a and object b), we construct a systematically diverse pool of input prompts for better understanding Algorithm 2 Iterative Refinement: Improving Text-to-Image Alignment Input: Text prompt P, subprompts p i , disjoint assertions in question format a q i","reference_id":"NDcwNTU0Nzk4NTc3ODQ5NzIzMg=="} +{"index":2212,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Note that while it is possible to explore relatively simpler methods e.g., extracting noun phrases for subject extraction in [7] , it can lead to errors as the complexity of the input prompt increases","reference_id":"NjAzNTE3MDkxNjY3NTY4MDEyMQ=="} +{"index":2213,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"While remarkable, generating images which align perfectly with the input text-prompt remains a challenging problem [6] [7] [8] 22 ]","reference_id":"MjE5MDIzNjcxMjAxNzg3MTM2Mw=="} +{"index":2214,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"We also note that while the recently proposed T2T similarity score [7] shows comparable correlation with ours for N = 2, its performance significantly drops as the number of subjects in the input prompt increases.","reference_id":"MzE4MjAzOTc5NjA4NjI1MjI0NA=="} +{"index":2215,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"To facilitate the reliable use of current text-to-image generation models for practical applications, it is essential to answer two key questions: 1) Can we detect such fine-grain misalignments between the input text and the generated output in a robust manner? and 2) Once detected, can we improve the text-to-image alignment for failure cases? While several metrics for evaluating text-to-image alignment (e.g., CLIP [9] , BLIP [10] , BLIP2 [11] ) exist, it has been observed [7, 12] that a high score with these metrics can be achieved even if the image does not fully correspond with input prompt","reference_id":"MTM3NzExNTc5MTMxNjQ2MjM5OTA="} +{"index":2216,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Furthermore, to reduce the inference time for each iteration as compared to [7] , cross-attention updates are only applied for first 20 steps of the reverse diffusion process","reference_id":"NTU1NzcxMDIyNDU3NjQ0MjA0Mw=="} +{"index":2217,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"In this section, we provide further details for the implementation of our approach as well as other baselines [1, [6] [7] [8] used while reporting results in the main paper","reference_id":"MTcyODQ4MzE0MDIyMjM0OTA1MTU="} +{"index":2218,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Instead of relying on human-inputs as in [6, 7] foot_0 , we leverage the in-context","reference_id":"MTQwNzk5MjY5MzE4NzAwMTc5OTU="} +{"index":2219,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"[7] require the user to provide a set of entities / subjects in the prompt, over which the cross-attention optimization should be performed","reference_id":"NjAzNTUxNTI0MjMzMDYzMTM1NQ=="} +{"index":2220,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"However, while remarkable, it has been observed [6] [7] [8] that as the complexity of the input text increases, the generated images do not always accurately align with the semantic meaning of the textual prompt.","reference_id":"MTM1NDI4MjIwODY1MjQ3NDIxNg=="} +{"index":2221,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"1 shows comparison for the average inference time (per image) for our approach with prior works [6] [7] [8]","reference_id":"MTIyMjY5ODMzODQzMDk3NDIzMzI="} +{"index":2222,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Attend-and-Excite [7] shows much better performance in addressing missing objects (e.g., telescope in example-1 and umbrella in example-4)","reference_id":"MTQwNjY0ODE4NTM2NjQ1NzI2NA=="} +{"index":2223,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Since Attend-and-Excite [7] is limited to applying the cross-attention update w.r.t the least dominant subject, as the complexity of input prompt increases, it may miss some objects (e.g., umbrella, beach, sunny day) during the generation process","reference_id":"MzQ2MTczODU5NTIzMTk1MzIxMQ=="} +{"index":2224,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Recently several works [6] [7] [8] have been proposed to explore the problem of improving image-text alignment in a training free manner","reference_id":"MTA4MzQzMzQwNjE2NjcyMzg5ODQ="} +{"index":2225,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"In particular, we show comparisons with 1) Stable Diffusion [1] , 2) Composable Diffusion [6] 3) StructureDiffusion [8] and 4) Attendand-Excite [7]","reference_id":"MTE2NjI2NzUxMDgzMjYzODgxMDk="} +{"index":2226,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"[7] require the user to provide a set of entities / subjects in the prompt, over which their optimization should be performed.","reference_id":"OTI0NDM5NTYyMzI4MzIyODY5MQ=="} +{"index":2227,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.13826","reference_contexts":"Similar to [7] , we also explore the idea of modifying the noise latents z t during the reverse diffusion process, to increase the cross-attention strength of the main nounsubject for each sub-assertion a i","reference_id":"MTI5MzY5MjY1MDg3OTY1Njkx"} +{"index":2228,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"While remarkable, generating images which align perfectly with the input text-prompt remains a challenging problem [6] [7] [8] 22 ]","reference_id":"MjE5MDIzNjcxMjAxNzg3MTM2Mw=="} +{"index":2229,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"Similar to [7, 8] , we use the official Stable Diffusion v1.4 model as the underlying pretrained text-to-image generation model while reporting results with all methods [6] [7] [8] (including the iterative refinement approach proposed in the main paper)","reference_id":"NDY5MzY3MzYyOTMwNzEzMzMx"} +{"index":2230,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"Through both qualitative and quantitative analysis, we find that the proposed iterative refinement process allows for generation of better aligned image outputs over prior works [6] [7] [8] while on average showing comparable inference times (Sec","reference_id":"OTg3MTczNjg5NDAyOTY2NzA4NQ=="} +{"index":2231,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"Recently several works [6] [7] [8] have been proposed to explore the problem of improving image-text alignment in a training free manner","reference_id":"MTA4MzQzMzQwNjE2NjcyMzg5ODQ="} +{"index":2232,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"We also note that training-free methods such as ours or prior works [6] [7] [8] , can in-turn help improve the performance of human-feedback based methods by providing a better quality dataset for determining both aesthetics and content alignment of the generated images.","reference_id":"MjkwODM1NzMzNzcwNzE5MzE4MQ=="} +{"index":2233,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"In particular, we show comparisons with 1) Stable Diffusion [1] , 2) Composable Diffusion [6] 3) StructureDiffusion [8] and 4) Attendand-Excite [7]","reference_id":"MTE2NjI2NzUxMDgzMjYzODgxMDk="} +{"index":2234,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"In contrast, Eval-and-Refine is able to describe both presence and relation between We compare the outputs of our iterative refinement approach with prior works [1, [6] [7] [8] on improving quality of generated images with changing number of subjects (underlined) from 2 to 5","reference_id":"MTUzMDIwMDU2Nzc3NDI1MTMyOQ=="} +{"index":2235,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"In this section, we provide further details for the implementation of our approach as well as other baselines [1, [6] [7] [8] used while reporting results in the main paper","reference_id":"MTcyODQ4MzE0MDIyMjM0OTA1MTU="} +{"index":2236,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"StructureDiffusion [8] helps in addressing some missing objects e.g., telescope in example-1, but the generated images tend to be semantically similar to those produced by the original Stable Diffusion model, and thus does not significantly improve text-to-image alignment.","reference_id":"MjI1MjI5ODMwMjQ1NjM0OTAyOA=="} +{"index":2237,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"While the proposed iterative refinement approach shows better performance than previous works [6] [7] [8] , it still has some limitations","reference_id":"MTM0NjE4NDQ5NDg0NTg5MDM0ODk="} +{"index":2238,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"These improvements are also reflected in the pairwise comparisons where human subjects tend to prefer our approach over prior works [6] [7] [8] .","reference_id":"MTUzOTc2MjcwNzU0NDMwODg0NDE="} +{"index":2239,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"However, while remarkable, it has been observed [6] [7] [8] that as the complexity of the input text increases, the generated images do not always accurately align with the semantic meaning of the textual prompt.","reference_id":"MTM1NDI4MjIwODY1MjQ3NDIxNg=="} +{"index":2240,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.05032","reference_contexts":"1 shows comparison for the average inference time (per image) for our approach with prior works [6] [7] [8]","reference_id":"MTIyMjY5ODMzODQzMDk3NDIzMzI="} +{"index":2241,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2104.08718","reference_contexts":"Various protocols for evaluating text-image alignment in a reference-free manner have been proposed [9] [10] [11]","reference_id":"ODMyMTE3NzM0NzA3MjM5MzQyOQ=="} +{"index":2242,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2104.08718","reference_contexts":"To facilitate the reliable use of current text-to-image generation models for practical applications, it is essential to answer two key questions: 1) Can we detect such fine-grain misalignments between the input text and the generated output in a robust manner? and 2) Once detected, can we improve the text-to-image alignment for failure cases? While several metrics for evaluating text-to-image alignment (e.g., CLIP [9] , BLIP [10] , BLIP2 [11] ) exist, it has been observed [7, 12] that a high score with these metrics can be achieved even if the image does not fully correspond with input prompt","reference_id":"MTM3NzExNTc5MTMxNjQ2MjM5OTA="} +{"index":2243,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2104.08718","reference_contexts":"Most prior works [2, 3, 5, 9] typically use the cosine similarity between the text and image embedding from large-scale multi-modal models [9, [36] [37] [38] such as CLIP [9] , BLIP [10] , BLIP-2 [11] for evaluating the alignment scores","reference_id":"MTEyMzgyOTg3MDU5MDM3NjQwNDA="} +{"index":2244,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2104.08718","reference_contexts":"Attend-and-Excite Ours Stable-Diffusion Attend-and-Excite Ours Traditional methods for evaluating text-to-image alignment e.g., CLIP [9] , BLIP-2 [10] and BLIP2-ITM (which provides a binary image-text matching score between 0 and 1) often fail to distinguish between good (right) and bad (left) image outputs and can give high scores even if the generated image is not an accurate match for input prompt (missing yellow car)","reference_id":"MTYwODE3NTkzOTIzNzgzODg2ODY="} +{"index":2245,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2104.08718","reference_contexts":"In particular, we show comparisons with CLIP [9] , BLIP [10] and BLIP2 [11] scores where the text-to-image alignment score is computed using the cosine similarity between the corresponding image and text embeddings","reference_id":"NTkxODg5ODkyNTI0OTI0ODY0OA=="} +{"index":2246,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.12597","reference_contexts":"Various protocols for evaluating text-image alignment in a reference-free manner have been proposed [9] [10] [11]","reference_id":"ODMyMTE3NzM0NzA3MjM5MzQyOQ=="} +{"index":2247,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.12597","reference_contexts":"To facilitate the reliable use of current text-to-image generation models for practical applications, it is essential to answer two key questions: 1) Can we detect such fine-grain misalignments between the input text and the generated output in a robust manner? and 2) Once detected, can we improve the text-to-image alignment for failure cases? While several metrics for evaluating text-to-image alignment (e.g., CLIP [9] , BLIP [10] , BLIP2 [11] ) exist, it has been observed [7, 12] that a high score with these metrics can be achieved even if the image does not fully correspond with input prompt","reference_id":"MTM3NzExNTc5MTMxNjQ2MjM5OTA="} +{"index":2248,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2301.12597","reference_contexts":"In particular, we show comparisons with CLIP [9] , BLIP [10] and BLIP2 [11] scores where the text-to-image alignment score is computed using the cosine similarity between the corresponding image and text embeddings","reference_id":"NTkxODg5ODkyNTI0OTI0ODY0OA=="} +{"index":2249,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2107.02423","reference_contexts":"For instance, [14] [15] [16] [17] [18] use GANs to perform text guided image generation.","reference_id":"NTMyODc3NDQxNjAzMjg0NTkzOA=="} +{"index":2250,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2211.01324","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2251,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2303.07909","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2252,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2212.10562","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2253,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2207.12598","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2254,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2108.02938","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2255,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2211.17084","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2256,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2211.14305","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2257,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2302.08113","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2258,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2211.11742","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2259,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2211.15518","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2260,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2302.05543","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2261,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2302.08453","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2262,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2208.12242","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2263,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2208.01618","reference_contexts":"Similarly, use of an additional guidance input to improve controllability of text-to-image generation have recently been extensively explored [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35]","reference_id":"MTc1NTI4MzQ0NDQ3Mzk0NjkxMDM="} +{"index":2264,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2205.01917","reference_contexts":"Most prior works [2, 3, 5, 9] typically use the cosine similarity between the text and image embedding from large-scale multi-modal models [9, [36] [37] [38] such as CLIP [9] , BLIP [10] , BLIP-2 [11] for evaluating the alignment scores","reference_id":"MTEyMzgyOTg3MDU5MDM3NjQwNDA="} +{"index":2265,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2303.11897","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2266,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2302.12192","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2267,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2302.13971","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2268,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2304.01852","reference_contexts":"This example-set is then used as an in-context dataset which is then fed to the 'gpt-4' [44] model to generate the desired prompt decompositions on prompts across the Decomposable-Captions-4k dataset","reference_id":"NTg3Njg3NDY1OTQ5NTA4NDM3OQ=="} +{"index":2269,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2304.01852","reference_contexts":"In particular, given a desired number of subjects N (between 2 and 5), we first use the GPT-4 model [44] API with 8K context length to come up with an initial random subject e.g., a dog","reference_id":"ODMyNDY0NjU1MjE4MzM1OTAyNw=="} +{"index":2270,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2304.01852","reference_contexts":"To this end, we leverage the diverse language modelling capabilities of large-scale large language models (LLM) [43, 44] in order to generate novel prompts of varying complexity and realism","reference_id":"MzI1NTQzMjY0NDExMTI5OTI1NA=="} +{"index":2271,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2304.01852","reference_contexts":"Instead of relying of human inputs as in prior works, we leverage the in-context learning capability of large-language models (LLMs) [43, 44] for obtaining such decompositions in an autonomous","reference_id":"MTY2MjM3OTI5NDgxMzIxNjA1NTc="} +{"index":2272,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2106.09685","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2273,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2111.02114","reference_contexts":"Moreover unlike prompts found in typical large-scale image-text datasets [48, 49] , the generated prompts should be imaginative and be able to describe novel and often non-realistic combinations of different concepts (e.g","reference_id":"OTA1NTA0NDU3NjMyMjk5MzEyNg=="} +{"index":2274,"citer_arxiv_id":"2307.04749","cited_arxiv_id":"2210.08402","reference_contexts":"Moreover unlike prompts found in typical large-scale image-text datasets [48, 49] , the generated prompts should be imaginative and be able to describe novel and often non-realistic combinations of different concepts (e.g","reference_id":"OTA1NTA0NDU3NjMyMjk5MzEyNg=="} +{"index":2275,"citer_arxiv_id":"1605.09304","cited_arxiv_id":"1507.02379","reference_contexts":"Many handdesigned natural image priors have been experimentally shown to improve image quality such as: Gaussian blur [7] , α-norm [5, 7, 8] , total variation [6, 9] , jitter [10, 6, 9] , data-driven patch priors [8] , center-bias regularization [9] , and initializing from mean images [9]","reference_id":"MTA3OTU0MTg4NjU3MDU4NTcx"} +{"index":2276,"citer_arxiv_id":"1605.09304","cited_arxiv_id":"1507.02379","reference_contexts":"For a fair comparison, the categories were not cherry-picked to showcase our best images, but instead were selected based on the images available in previous papers [5, [7] [8] [9] 6]","reference_id":"MTUyMDI3MDMzNjA1MTUyODE0ODE="} +{"index":2277,"citer_arxiv_id":"1605.09304","cited_arxiv_id":"1408.5093","reference_contexts":"Figure S14 : Visualization of example neuron feature detectors from all eight layers of a CaffeNet DNN [20]","reference_id":"NjQ1NDY4OTY5MjU1MzEwODM3NA=="} +{"index":2278,"citer_arxiv_id":"1605.09304","cited_arxiv_id":"1408.5093","reference_contexts":"For G we use networks made publicly available by [11] that have been trained with the principles of GANs [17] to reconstruct images from hidden-layer feature representations within CaffeNet [20]","reference_id":"MTY5MzgzNjU4OTQ1MTE1NzQxMDA="} +{"index":2279,"citer_arxiv_id":"1605.09304","cited_arxiv_id":"1408.5093","reference_contexts":"Specifically, we target the output neurons of two DNNs downloaded from Caffe Model Zoo [20] ): (1) An AlexNet DNN that was trained on the 2.5-million-image MIT Places dataset to classify 205 types of places with 50.1% accuracy [26]","reference_id":"NDM4NjgzNDE0MzYyNDQ1MjE1OQ=="} +{"index":2280,"citer_arxiv_id":"1605.09304","cited_arxiv_id":"1408.5093","reference_contexts":"For reproducibility, we use pretrained models freely available in Caffe or the Caffe Model Zoo [20] : CaffeNet [20] , GoogleNet [21] , and ResNet [22]","reference_id":"OTkzMzY4NDU5NTQyODY4NTA1MA=="} +{"index":2281,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1711.02132","reference_contexts":"As such, it is only natural that a wealth of research has been dedicated to making fundamental improvements to the model over the past few years (Dehghani et al., 2018; So et al., 2019; Ahmed et al., 2017)","reference_id":"MTI0Mjk3Njk1ODE1MTkwNDAxMzQ="} +{"index":2282,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1607.06450","reference_contexts":"Transformer blocks are characterized by a multi-head self-attention mechanism, a positionwise feed-forward network, layer normalization (Ba et al., 2016) modules and residual connectors","reference_id":"MTAzNzA4ODc3MDY2NjUxMTI0NDE="} +{"index":2283,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2005.14165","reference_contexts":"Transformers are pervasive and have made tremendous impact in many fields such as language understanding (Devlin et al., 2018; Brown et al., 2020; Raffel et al., 2019) and image processing (Parmar et al., 2018; Carion et al., 2020)","reference_id":"MjE3NDg3MzQ5MjcwMDMyODI4MA=="} +{"index":2284,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2005.14165","reference_contexts":"Due to the nature of the loss, these models are often superior for open ended generation (Brown et al., 2020)","reference_id":"MjI5NjcxOTI2NDMwMDA3MTMzNA=="} +{"index":2285,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2005.12872","reference_contexts":"Transformers are pervasive and have made tremendous impact in many fields such as language understanding (Devlin et al., 2018; Brown et al., 2020; Raffel et al., 2019) and image processing (Parmar et al., 2018; Carion et al., 2020)","reference_id":"MjE3NDg3MzQ5MjcwMDMyODI4MA=="} +{"index":2286,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"To this end, most early work in this area is based on block/local patterns such as Image Transformer (Parmar et al., 2018) , Compressed Attention (Liu et al., 2018) , Blockwise Transformer (Qiu et al., 2019) or the local windows in Sparse Transformer (Child et al., 2019) .","reference_id":"MTAxOTU4ODQ4NzQ3NDI5NTY2OTM="} +{"index":2287,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2288,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"The Sparse Transformer (Child et al., 2019) presents a simple initial attempt to reduce the quadratic complexity of the standard self-attention mechanism","reference_id":"MTU3NDc4NTczNDQ4NDgxODI2MTM="} +{"index":2289,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"Models such as Sparse Transformer (Child et al., 2019) and/or Longformer (Beltagy et al., 2020) employ strided or \"dilated\" windows.","reference_id":"MTIxMDg1NzYwNTAyMTcwODM0MzA="} +{"index":2290,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"Meanwhile, local attention models based on fixed and/or learned patterns such as Sparse Transformers (Child et al., 2019) , Longformer (Beltagy et al., 2020) , ETC (Ainslie et al., 2020) or BigBird (Zaheer et al., 2020) have seen more reasonable usage, especially within the areas of long context question answering","reference_id":"MTE1OTYxMDc5MDA5NzEwNzE2NTA="} +{"index":2291,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2292,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"To this end, Sparse Transformers (Child et al., 2019 ), Adaptive Transformers (Correia et al., 2019) , Routing Transformers (Roy et al., 2020) and Reformers (Kitaev et al., 2020) are mainly focused on generative modeling tasks","reference_id":"NzYzMjMzODM0NDA2NDYzMjMyMA=="} +{"index":2293,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"For example, the Sparse Transformer (Child et al., 2019) combines strided and local attention by assigning half of its heads to each pattern","reference_id":"NzYyODgxMjI3Mzg5NDQyMzkwNQ=="} +{"index":2294,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.10509","reference_contexts":"Following which, we move on to models that utilize combinations of patterns such as Sparse Transformers (Child et al., 2019) , CCNet (Huang et al., 2019) , and Axial Transformers (Ho et al., 2019)","reference_id":"NDU3MzU4MzQyNzA5MzI3MDE3NQ=="} +{"index":2295,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2110.12894","reference_contexts":"Efficiency and computational costs is generally a complicated affair and we would suggest readers peruse (Dehghani et al., 2021) for more details on trade-offs, intricacies etc.","reference_id":"MzYxNDMzOTQ2NjExMTA1NDE0"} +{"index":2296,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2110.12894","reference_contexts":"More notes and discussions about this efficiency misnomer can be found in this paper (Dehghani et al., 2021) which we encourage readers to also peruse.","reference_id":"MTIwMTMxMzE4Mjc5MDE4NjQ2NDM="} +{"index":2297,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2004.07320","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2298,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2101.03961","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2299,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2101.03961","reference_contexts":"Hence, a large portion of recent work have explored sparsity (Lepikhin et al., 2020; Fedus et al., 2021) as a means to scale up the FFN without incurring compute costs","reference_id":"MTU1NzYzMjYwMzQzNTc5NjU1NjI="} +{"index":2300,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2101.03961","reference_contexts":"Examples of this class of model includes Switch Transformers (Fedus et al., 2021) , ST-MoE (Zoph et al., 2022) , GShard (Lepikhin et al., 2020) , Product-Key Memory Layers (Lample et al., 2019)","reference_id":"MTI2MjU0Mjk0ODU3OTU4OTg4ODM="} +{"index":2301,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2101.03961","reference_contexts":"We believe that sparse models were a necessarily addition to the scope to this paper given its recent signs of promise (Fedus et al., 2021; Du et al., 2021; Zoph et al., 2022)","reference_id":"NTQ1MzQ5ODEzMjcxNTkzMzY5Nw=="} +{"index":2302,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2101.03961","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2303,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2101.03961","reference_contexts":"We also observe that Sparse models (Lepikhin et al., 2020; Fedus et al., 2021) that are not only applicable to attention modules, are also recently emerging and becoming more popular and have demonstrated considerable success in the recent months (Du et al., 2021) .","reference_id":"NzEyODAwNTkyNTc1MTg3ODQ5NA=="} +{"index":2304,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2101.03961","reference_contexts":"It is good to note that while most of the works within the scope of this survey deals with efficient attention, the scope of sparse models goes beyond the attention module and is generally applied more frequently to the feed forward layers (Lepikhin et al., 2020; Fedus et al., 2021)","reference_id":"MTU2MDA1NTA4MjQwNzU4NjU4Nzg="} +{"index":2305,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1503.02531","reference_contexts":"• Knowledge Distillation -Knowledge distillation (KD) (Hinton et al., 2015) has been a useful technique for transfering the knowledge learned from a larger teacher model to a smaller student model","reference_id":"ODc2Mzc1MjA4MzUwMDU4MjcwOA=="} +{"index":2306,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1912.12180","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2307,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1912.12180","reference_contexts":"Similarly, Axial Transformer (Ho et al., 2019) applies a sequence of self-attention computations given a high dimensional tensor as input, each along a single axis of the input tensor","reference_id":"MTgyOTM3NjIyNjI4MDY0ODQyMzU="} +{"index":2308,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1912.12180","reference_contexts":"Axial Transformer (Ho et al., 2019; Weissenborn et al., 2019) input, Axial Transformer simply applies multiple attentions, each along a single axis of the input tensor","reference_id":"MTQ1MjkzODMzNDU0MzA2MDAzNzA="} +{"index":2309,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1912.12180","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2310,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1912.12180","reference_contexts":"Following which, we move on to models that utilize combinations of patterns such as Sparse Transformers (Child et al., 2019) , CCNet (Huang et al., 2019) , and Axial Transformers (Ho et al., 2019)","reference_id":"NDU3MzU4MzQyNzA5MzI3MDE3NQ=="} +{"index":2311,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2107.14795","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2312,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2107.14795","reference_contexts":"Examples of this class of models include Perceiver (Jaegle et al., 2021) , Funnel Transformers (Dai et al., 2020) , Swin Transformer (Liu et al., 2021b) , and Charformer (Tay et al., 2021c) models","reference_id":"MTExMjMwMDA0NTQ1MzEzMzczNDE="} +{"index":2313,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2107.14795","reference_contexts":"Even though many existing 'memory' models utilize a form of downsampling to achieve a speed and efficiency gain, we erected a new categoriation of 'downsampling' to better reflect this new emerging trend (Dai et al., 2020; Jaegle et al., 2021; Tay et al., 2021c; Ryoo et al., 2021) .","reference_id":"NTc5NTY4OTQ1Njc1NjQzMDg4MQ=="} +{"index":2314,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2107.14795","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2315,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1909.10351","reference_contexts":"For example, DistilBERT (Sanh et al., 2019) , task-specific distillation (Tang et al., 2019) and TinyBERT (Jiao et al., 2019) .","reference_id":"MTQyOTI0Nzk0Nzk4Nzg1MzY5NjY="} +{"index":2316,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2317,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"The Linear Transformer (Katharopoulos et al., 2020 ) also evaluates on speech recognition, which is a rare benchmark amongst this group of papers.","reference_id":"NTYxMjk1MTc1MzIyNjIwNDEyMQ=="} +{"index":2318,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"The usage of kernels (Katharopoulos et al., 2020; Choromanski et al., 2020a) enable clever mathematical re-writing of the self-attention mechanism to avoid explicitly computing the N × N matrix","reference_id":"OTMyNDc3MTU5NzU1Mjg1MjE3Nw=="} +{"index":2319,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"This immense interest has also spurred research into more efficient variants of the model (Kitaev et al., 2020; Roy et al., 2020; Beltagy et al., 2020; Katharopoulos et al., 2020; Tay et al., 2020b; Wang et al., 2020c; Rae et al., 2020; Choromanski et al., 2020b; Dai et al., 2020; Correia et al., 2019; Sukhbaatar et al., 2019a; Vyas et al., 2020) .","reference_id":"MTUxOTE4NzQ1MDMzMzc1MTU3OTE="} +{"index":2320,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"We then discuss models based on kernel approaches such as Performer (Choromanski et al., 2020a) and Linear Transformers (Katharopoulos et al., 2020)","reference_id":"MTczODI3NDgyMTYyMTkzODY2ODQ="} +{"index":2321,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2322,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"The most recent wave of models we've been seeing is models that are based on low-rank approximation or kernel methods, e.g., models such as Low-Rank Transformer (Winata et al., 2020) , Linformer (Wang et al., 2020c) , Performer (Choromanski et al., 2020a) and/or Linear Transformers (Katharopoulos et al., 2020)","reference_id":"MTM0Njc2MTM4MjEwNTkwMDIwMjI="} +{"index":2323,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16236","reference_contexts":"The Linear Transformer (Katharopoulos et al., 2020) improves the complexity of selfattention from quadratic to linear by using a kernel-based formulation of self-attention and the associative property of matrix products","reference_id":"ODcyMzI4MzIxMTE5MjY1NzMwMg=="} +{"index":2324,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1907.05242","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2325,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1907.05242","reference_contexts":"Examples of this class of model includes Switch Transformers (Fedus et al., 2021) , ST-MoE (Zoph et al., 2022) , GShard (Lepikhin et al., 2020) , Product-Key Memory Layers (Lample et al., 2019)","reference_id":"MTI2MjU0Mjk0ODU3OTU4OTg4ODM="} +{"index":2326,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1907.05242","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2327,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2105.03824","reference_contexts":"Different mixing operations have been proposed, such as the G-MLP (Liu et al., 2021a) , FNet (Lee-Thorp et al., 2021)","reference_id":"MTgyOTQzNzYyNDkzNTI5NjU5Njc="} +{"index":2328,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16668","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2329,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16668","reference_contexts":"However, we believe that the emergence of an entire research direction (Roller et al., 2021; Lewis et al., 2021; Lepikhin et al., 2020; Du et al., 2021) based on sparse efficient should warrant a new category of efficient Transformers.","reference_id":"NjEzMzc1MDE2MTAwNDU4NzM3Mw=="} +{"index":2330,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16668","reference_contexts":"Hence, a large portion of recent work have explored sparsity (Lepikhin et al., 2020; Fedus et al., 2021) as a means to scale up the FFN without incurring compute costs","reference_id":"MTU1NzYzMjYwMzQzNTc5NjU1NjI="} +{"index":2331,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16668","reference_contexts":"Examples of this class of model includes Switch Transformers (Fedus et al., 2021) , ST-MoE (Zoph et al., 2022) , GShard (Lepikhin et al., 2020) , Product-Key Memory Layers (Lample et al., 2019)","reference_id":"MTI2MjU0Mjk0ODU3OTU4OTg4ODM="} +{"index":2332,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16668","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2333,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16668","reference_contexts":"We also observe that Sparse models (Lepikhin et al., 2020; Fedus et al., 2021) that are not only applicable to attention modules, are also recently emerging and becoming more popular and have demonstrated considerable success in the recent months (Du et al., 2021) .","reference_id":"NzEyODAwNTkyNTc1MTg3ODQ5NA=="} +{"index":2334,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.16668","reference_contexts":"It is good to note that while most of the works within the scope of this survey deals with efficient attention, the scope of sparse models goes beyond the attention module and is generally applied more frequently to the feed forward layers (Lepikhin et al., 2020; Fedus et al., 2021)","reference_id":"MTU2MDA1NTA4MjQwNzU4NjU4Nzg="} +{"index":2335,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2103.16716","reference_contexts":"However, we believe that the emergence of an entire research direction (Roller et al., 2021; Lewis et al., 2021; Lepikhin et al., 2020; Du et al., 2021) based on sparse efficient should warrant a new category of efficient Transformers.","reference_id":"NjEzMzc1MDE2MTAwNDU4NzM3Mw=="} +{"index":2336,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2103.14030","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2337,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1904.01038","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2338,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1911.02972","reference_contexts":"To this end, most early work in this area is based on block/local patterns such as Image Transformer (Parmar et al., 2018) , Compressed Attention (Liu et al., 2018) , Blockwise Transformer (Qiu et al., 2019) or the local windows in Sparse Transformer (Child et al., 2019) .","reference_id":"MTAxOTU4ODQ4NzQ3NDI5NTY2OTM="} +{"index":2339,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1911.02972","reference_contexts":"Performer (Choromanski et al., 2020) Linformer (Wang et al., 2020b) Linear Transformer (Katharopoulos et al., 2020) Set Transformer (Lee et al., 2019) Transformer-XL (Dai et al., 2019) Memory Compressed (Liu et al., 2018) ETC (Ainslie et al., 2020) Sparse Transformer (Child et al., 2019) Image Transformer (Parmar et al., 2018) Synthesizer (Tay et al., 2020a) Longformer (Beltagy et al., 2020) Big Bird (Zaheer et al., 2020) Axial Transformer (Ho et al., 2019) Blockwise Transformer (Qiu et al., 2019) Sinkhorn Transformer (Tay et al., 2020b) Reformer (Kitaev et al., 2020) Compressive Transformer (Rae et al., 2018) Routing Transformer (Roy et al., 2020) Funnel Transformer (Dai et al., 2020) Random Feature Attention (Peng et al., 2021) Long Short Transformer (Zhu et al., 2021) Poolingformer (Zhang et al., 2021) Nystromformer (Xiong et al., 2019) Adaptive Sparse Transformer (Correia et al., 2019) Product Key Memory (Lample et al., 2019) GShard (Lepikhin et al., 2020) Switch Transformer (Fedus et al., 2021) TokenLearner (Ryoo et al., 2021) Perceiver (Jaegle et al., 2021) Clusterformer (Wang et al., 2020) Scaling Transformer (Jaszczur et al., 2021) Low-Rank Transformer (Winata et al., 2020) Clustered Attention (Vyas et al., 2020) CC-Net (Huang et al., 2018) GLaM (Du et al., 2021) Swin Transformer (Liu et al., 2020) Charformer (Tay et al., 2021) Figure 2 : Taxonomy of Efficient Transformer Architectures","reference_id":"NzkxMTM4MzQyODA3OTk2MTgzOQ=="} +{"index":2340,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1911.02972","reference_contexts":"Examples of models that do this include Blockwise (Qiu et al., 2019) and/or Local Attention (Parmar et al., 2018)","reference_id":"MTg0MzU1NTE0ODEyNDM5MDE0OTQ="} +{"index":2341,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2112.05682","reference_contexts":"As noted by (Rabe and Staats, 2021) , for applications that require to flex on sequence length and memory needs time to time, it might be suffice to 'just sequentially process it' even if that might not be inherently as satisfying as finding a theoretical approximate","reference_id":"MTgyMjE1OTUyNzQ5MzcwMTM4NDE="} +{"index":2342,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2106.04426","reference_contexts":"However, we believe that the emergence of an entire research direction (Roller et al., 2021; Lewis et al., 2021; Lepikhin et al., 2020; Du et al., 2021) based on sparse efficient should warrant a new category of efficient Transformers.","reference_id":"NjEzMzc1MDE2MTAwNDU4NzM3Mw=="} +{"index":2343,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1910.01108","reference_contexts":"For example, DistilBERT (Sanh et al., 2019) , task-specific distillation (Tang et al., 2019) and TinyBERT (Jiao et al., 2019) .","reference_id":"MTQyOTI0Nzk0Nzk4Nzg1MzY5NjY="} +{"index":2344,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1905.07799","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2345,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1907.01470","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2346,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"1903.12136","reference_contexts":"For example, DistilBERT (Sanh et al., 2019) , task-specific distillation (Tang et al., 2019) and TinyBERT (Jiao et al., 2019) .","reference_id":"MTQyOTI0Nzk0Nzk4Nzg1MzY5NjY="} +{"index":2347,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2105.03322","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2348,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2106.12672","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2349,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2006.04768","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2350,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2202.08906","reference_contexts":"We believe that sparse models were a necessarily addition to the scope to this paper given its recent signs of promise (Fedus et al., 2021; Du et al., 2021; Zoph et al., 2022)","reference_id":"NTQ1MzQ5ODEzMjcxNTkzMzY5Nw=="} +{"index":2351,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2202.08906","reference_contexts":"Examples of this class of model includes Switch Transformers (Fedus et al., 2021) , ST-MoE (Zoph et al., 2022) , GShard (Lepikhin et al., 2020) , Product-Key Memory Layers (Lample et al., 2019)","reference_id":"MTI2MjU0Mjk0ODU3OTU4OTg4ODM="} +{"index":2352,"citer_arxiv_id":"2009.06732","cited_arxiv_id":"2202.08906","reference_contexts":"Complexity Decode Class Memory Compressed (Liu et al., 2018) O(N 2 c ) FP+M Image Transformer (Parmar et al., 2018) O(N.m) FP Set Transformer (Lee et al., 2019) O(kN ) M Transformer-XL (Dai et al., 2019) O(N 2 ) RC Sparse Transformer (Child et al., 2019) O(N √ N ) FP Reformer (Kitaev et al., 2020) O(N log N ) LP Routing Transformer (Roy et al., 2020) O(N N ) LP Axial Transformer (Ho et al., 2019) O(N √ N ) FP Compressive Transformer (Rae et al., 2020) O(N 2 ) RC Sinkhorn Transformer (Tay et al., 2020b) O(B 2 ) LP Longformer (Beltagy et al., 2020) O(n(k + m)) FP+M ETC (Ainslie et al., 2020) O(N 2 g + N N g ) FP+M Synthesizer (Tay et al., 2020a) O(N 2 ) LR+LP Performer (Choromanski et al., 2020a) O(N ) KR Funnel Transformer (Dai et al., 2020) O(N 2 ) FP+DS Linformer (Wang et al., 2020c) O(N ) LR Linear Transformers (Katharopoulos et al., 2020) O(N ) KR Big Bird (Zaheer et al., 2020) O(N ) FP+M Random Feature Attention (Peng et al., 2021) O(N ) KR Long Short Transformers (Zhu et al., 2021) O(kN ) FP + LR Poolingformer (Zhang et al., 2021) O(N ) FP+M Nyströmformer (Xiong et al., 2021b) O(kN ) M+DS Perceiver (Jaegle et al., 2021) O(kN ) M+DS Clusterformer (Wang et al., 2020b) O(N log N ) LP Luna (Ma et al., 2021) O(kN ) M TokenLearner (Ryoo et al., 2021) O(k 2 ) DS Adaptive Sparse Transformer (Correia et al., 2019) O(N 2 ) Sparse Product Key Memory (Lample et al., 2019) O(N 2 ) Sparse Switch Transformer (Fedus et al., 2021) O(N 2 ) Sparse ST-MoE (Zoph et al., 2022) O(N 2 ) Sparse GShard (Lepikhin et al., 2020) O(N 2 ) Sparse Scaling Transformers (Jaszczur et al., 2021) O(N 2 ) Sparse GLaM (Du et al., 2021) O(N 2 ) Sparse","reference_id":"MTgxMTQ4MDg2MzgzNDY0Nzk1OTM="} +{"index":2353,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1512.03012","reference_contexts":"It is a basic component in most state-ofthe-art architectures [9] , [10] , [11] , [12] , [13] , [14] , [15] , [16] , and has successfully proliferated throughout various areas of deep learning [17] , [18] , [19]","reference_id":"NjQzODExODg5ODQxODk2MDg5NQ=="} +{"index":2354,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1607.06450","reference_contexts":"This is mainly due to its scale-invariant property [20] , [183] , [184] , [185] , i.e., it does not change the prediction when rescaling parameters and works by adaptively adjusting the learning rate in a layer-wise manner [92] , [94] , [160]","reference_id":"MTM1NzY2MTg4MTgxNTI1NjkxNzQ="} +{"index":2355,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1607.06450","reference_contexts":"LN [20] proposes to standardize the layer input within the neurons for each training sample, to avoid the drawbacks of normalization along batch dimensions","reference_id":"MzYxNzIyMjg1MTA3ODY4NDMxMA=="} +{"index":2356,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1607.06450","reference_contexts":"The scale-invariant property also applies to other methods that normalize the activations [20] , [22] , [34] , [67] or weights [21] , [51] , [52]","reference_id":"MTU4Mjg0NzM2NjI5NjQ3MDc5MDA="} +{"index":2357,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1607.06450","reference_contexts":"Further, a significant number of other normalization techniques have been proposed to address the training issues in particular contexts, further evolving the DNN architectures and their applications [20] , [21] , [22] , [23] , [24]","reference_id":"Njg5NzczNjc2OTc0MjY1NjQ5OQ=="} +{"index":2358,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1607.06450","reference_contexts":"The other strategy is to normalize the activations as a function transformation, which requires backpropagation through this transformation [8] , [20] , [34] .","reference_id":"Njk4MDM3OTQ1OTMzMzQwNzAyOA=="} +{"index":2359,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1607.06450","reference_contexts":"[97] hypothesized that the re-centering invariance produced by centering in LN [20] is dispensable and proposed to perform scaling only for LN, which is referred to as root mean square layer normalization (RMSLN)","reference_id":"MjA2NTY0NTI2MTA0NjExMTc5MQ=="} +{"index":2360,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1607.06450","reference_contexts":"recurrent neural networks (RNNs) [20] , [60] , [61] ) and tasks [29] , [62] , [63] ; 2) BN suffers from the small-batch-size problem -its error increases rapidly as the batch size becomes smaller [22]","reference_id":"ODc2NDE3MzkyMzIxODM3MDE1Mw=="} +{"index":2361,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1412.1193","reference_contexts":"Note that [38] , [42] , [43] have provided empirical evidence to support their effectiveness in approximating the full FIM with block diagonal sub-FIMs","reference_id":"ODU5OTI4MjA1MjA4MzU5OTU2Mw=="} +{"index":2362,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1707.04822","reference_contexts":"[53] were the first to propose block-wise (layer-wise) gradient normalization for training DNNs to front the gradient explosion or vanishing problem","reference_id":"NTM5MzY2MzY2NDIzOTc1OTUzOQ=="} +{"index":2363,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1707.04822","reference_contexts":"(3) Normalizing gradients to exploit the curvature information for GD/SGD, even though the optimization landscape is illconditioned [53] , [54]","reference_id":"NzA2NDMxNzEzNzU5MDM4MDk1Nw=="} +{"index":2364,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1708.03888","reference_contexts":"[54] , for large-batch training","reference_id":"OTg5Mzk0MTU0MjkyMjk4ODc3Nw=="} +{"index":2365,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1708.03888","reference_contexts":"(3) Normalizing gradients to exploit the curvature information for GD/SGD, even though the optimization landscape is illconditioned [53] , [54]","reference_id":"NzA2NDMxNzEzNzU5MDM4MDk1Nw=="} +{"index":2366,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1902.05605","reference_contexts":"Another interesting observation is that normalization is not very common in deep reinforcement learning (DRL) [63]","reference_id":"OTkxMDg2MjAxNDQxNjE3NzczNw=="} +{"index":2367,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1902.05605","reference_contexts":"recurrent neural networks (RNNs) [20] , [60] , [61] ) and tasks [29] , [62] , [63] ; 2) BN suffers from the small-batch-size problem -its error increases rapidly as the batch size becomes smaller [22]","reference_id":"ODc2NDE3MzkyMzIxODM3MDE1Mw=="} +{"index":2368,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1610.06160","reference_contexts":"As an alternative to the L 2 normalization used to control the activation scale in a BN layer [8] , the L 1 normalization was proposed in [90] , [91] , [92] for standardization","reference_id":"MTA2NDk4MzQzODQ2NzU3ODc4MDI="} +{"index":2369,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1610.06160","reference_contexts":"One challenge for this type of method is how to calculate the gradients during backpropagation, since the population statistics are computed by all the previous minibatches, and it is impossible to obtain their exact gradients [90]","reference_id":"NDk4NTg1NzkyNDM3MDQ1MDMzNA=="} +{"index":2370,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1610.06160","reference_contexts":"The more generalized L p was investigated in [90] and [92] , where the standardization deviation is:","reference_id":"MTM0NTgzNjk5MTczMzk3NzE5MDc="} +{"index":2371,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1803.05407","reference_contexts":"However, while this can benefit a model trained with a small batch size, where estimation is the main issue [8] , [113] , [135] , it may lead to degenerated generalization when the batch size is moderate.","reference_id":"MTY4Mjk2NTE1NDk5MjAwOTYzNjM="} +{"index":2372,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1903.10520","reference_contexts":"Weight centering is also advocated for in [95] , [141]","reference_id":"OTExMTkzOTAyNDg2MDExNDEzOA=="} +{"index":2373,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1903.10520","reference_contexts":"This technique was then further used in other methods for learning with different constraints, e.g., orthogonal weight normalization [52] , weight standardization [141] , spectral normalization [23] and weight centralization [95]","reference_id":"MTM3MzExMDE5NTE1NzgyOTIwNzQ="} +{"index":2374,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1903.10520","reference_contexts":"Therefore, WS usually needs to be combined with activation normalization methods (e.g., BN/GN) to relieve this issue [141] .","reference_id":"MTQwMjQxNDQxMjU2OTc2NjA1MTU="} +{"index":2375,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1903.10520","reference_contexts":"The idea of combining normalizing weights and activations to improve performance has been widely studied [24] , [52] , [141] , [171]","reference_id":"MTM2MjQ4NTI1NjM2MjMyMTg5NTk="} +{"index":2376,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1710.02338","reference_contexts":"Another research direction is to analyze the effect of weight decay [189] when combined with scale-invariant normalization methods [92] , [142] , [161] , [185] , [190] , [191] , [192]","reference_id":"MTY4OTQ4MTE3ODI0NzY0ODEyODg="} +{"index":2377,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1710.02338","reference_contexts":"Riemannian Optimization A weight matrix W with constraints can be viewed as an embedded submanifold [52] , [160] , [161] , [162]","reference_id":"NTUzNTU2MjkwNTI0NjU2NTk1Nw=="} +{"index":2378,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1709.08145","reference_contexts":"However, normalizing weights has some drawbacks when used in large-scale networks in practice: 1) It may not effectively improve the optimization efficiency when residual connections are introduced or when the nonlinearity does not satisfy the assumptions required for preserving distributions between different layers, since the Criteria 1 in Section 3 is not readily satisfied in these situations; 2) Normalizing the weight usually has significantly lower test accuracy than BN on large-scale image classification [170]","reference_id":"MTI0NTIxMDU3NjU4MDY5NTU4NzM="} +{"index":2379,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1706.05350","reference_contexts":"Another research direction is to analyze the effect of weight decay [189] when combined with scale-invariant normalization methods [92] , [142] , [161] , [185] , [190] , [191] , [192]","reference_id":"MTY4OTQ4MTE3ODI0NzY0ODEyODg="} +{"index":2380,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1803.02999","reference_contexts":"However, we note that there also exist works exploring how to apply normalization to meta learning [59] , [214] , [215] , reinforcement learning [216] , [217] , [218] , unsupervised representation learning [219] , [220] , permutation-equivariant networks [221] , [222] , graph neural networks [223] , ordinary differential equation (ODE) based networks [224] , symmetric positive definite (SPD) neural networks [225] , and guarding against adversarial attacks [226] , [227] , [228] .","reference_id":"Njc1ODM3OTc5NzY3MDQ2NDY2NA=="} +{"index":2381,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1902.05605","reference_contexts":"However, we note that there also exist works exploring how to apply normalization to meta learning [59] , [214] , [215] , reinforcement learning [216] , [217] , [218] , unsupervised representation learning [219] , [220] , permutation-equivariant networks [221] , [222] , graph neural networks [223] , ordinary differential equation (ODE) based networks [224] , symmetric positive definite (SPD) neural networks [225] , and guarding against adversarial attacks [226] , [227] , [228] .","reference_id":"Njc1ODM3OTc5NzY3MDQ2NDY2NA=="} +{"index":2382,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1511.06434","reference_contexts":"For example, combining samples form different domains in a batch for BN may harm the generalization in domain adaptation, and this also applies for the training of GANs [62] , [252] .","reference_id":"NTg2NDEzMzEzMjIxNzU0NTI1Mg=="} +{"index":2383,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1701.07875","reference_contexts":"One way to remedy this issue is to impose constraints on the discriminator [253]","reference_id":"MTI3MjM4MTU4ODI5NzgzMDQyMTU="} +{"index":2384,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1704.03971","reference_contexts":"[254] leveraged weight normalization to effectively improve the training performance of GANs","reference_id":"NTk0MTEwMTE4MDIyODk4MTU0"} +{"index":2385,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1411.1784","reference_contexts":"As discussed in Section 9.2 for style transfer, the NRR operation of activation normalization can also be used as the side information for GANs, under the scenario of conditional GANs (cGANs) [258]","reference_id":"MTIzMjk3ODE2NTc3ODY0MTMxNTc="} +{"index":2386,"citer_arxiv_id":"2009.12836","cited_arxiv_id":"1702.08231","reference_contexts":"RBN avoids the sum of squares, square-root and reciprocal operations and is more friendly for low-precise training [273] .","reference_id":"MTQ4NzI3OTM0NDg3MjcxOTIzNTI="} +{"index":2387,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1711.04340","reference_contexts":"Generative adversarial networks have also been used for the purpose of generating additional data (e.g., [45, 41, 70, 2, 56] )","reference_id":"MTA1NzI1NTg4NjI0MTI3MTQ2NzA="} +{"index":2388,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1708.04552","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2389,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1711.04528","reference_contexts":"Our approach is inspired by recent advances in architecture search, where reinforcement learning and evolution have been used to discover model architectures from data [71, 4, 72, 7, 35, 13, 34, 46, 49, 63, 48, 9]","reference_id":"MTU2NDAxODg0NDMxMTk4NDg2MDg="} +{"index":2390,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1705.07485","reference_contexts":"All the results of the baseline models, and baseline models with Cutout are replicated in our experiments and match the previously reported results [67, 17, 65, 12]","reference_id":"MTM5OTg5NjgzNzM1MDA0NTA1NTk="} +{"index":2391,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1705.07485","reference_contexts":"1 PyramidNet+ShakeDrop [65] 2.7 2.3 1.5 ± 0.1 Reduced CIFAR-10 Wide-ResNet-28-10 [67] 18.8 16.5 14.1±0.3 Shake-Shake (26 2x96d) [17] 17.1 13.4 10.0 ± 0.2 CIFAR-100 Wide-ResNet-28-10 [67] 18.8 18.4 17.1±0.3 Shake-Shake (26 2x96d) [17] 17.1 16.0 14.3±0.2 PyramidNet+ShakeDrop [65] 14.0 12.2 10.7 ± 0.2 SVHN Wide-ResNet-28-10 [67] 1.5 1.3 1.1 Shake-Shake (26 2x96d) [17] 1.4 1.2 1.0 Reduced SVHN Wide-ResNet-28-10 [67] 13.2 32.5 8.2 Shake-Shake (26 2x96d) [17] 12.3 24.2 5.9 Table 2","reference_id":"ODgzNDc5MzQ5NjY1MzczOTE5NQ=="} +{"index":2392,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1705.07485","reference_contexts":"Wide-ResNet-28-10 [67] 3.9 3.1 2.6±0.1 Shake-Shake (26 2x32d) [17] 3.6 3.0 2.5±0.1 Shake-Shake (26 2x96d) [17] 2.9 2.6 2.0±0.1 Shake-Shake (26 2x112d) [17] 2.8 2.6 1.9±0.1 AmoebaNet-B (6,128) [48] 3.0 2.1 1.8±0.","reference_id":"NjkzNTQ0MDkwMDE1NTg1NjE2Mw=="} +{"index":2393,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1705.07485","reference_contexts":"Other hyperparameters are the same as reported in the papers introducing the models [67, 17, 65] , with the exception of using a cosine learning decay for the Wide-ResNet-28-10","reference_id":"NDg4NzY1NDY5NDI4Mzc1Mjk3NA=="} +{"index":2394,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1705.07485","reference_contexts":"We implement the Wide-ResNet-28-10 [67], Shake-Shake [17] and ShakeDrop [65] models in TensorFlow [1] , and find the weight decay and learning rate hyperparameters that give the best validation set accuracy for regular training with baseline augmentation","reference_id":"MTUyNjkwODA1MjI4MTA4Mzg5MTQ="} +{"index":2395,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1705.07485","reference_contexts":"The baseline pre-processing follows the convention for state-of-the-art CIFAR-10 models: standardizing the data, using horizontal flips with 50% probability, zero-padding and random crops, and finally Cutout with 16x16 pixels [17, 65, 48, 72]","reference_id":"MzYxMzI3NTMxMDI2MjMwNzk0Ng=="} +{"index":2396,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1705.07485","reference_contexts":"Other hyperparameters and training details are identical to the those in the papers introducing the models [67, 17]","reference_id":"MTc1MzAyMDIzODYxNDk2MjM3MzA="} +{"index":2397,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1312.5402","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2398,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1709.01507","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2399,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1805.08974","reference_contexts":"Even on datasets for which fine-tuning weights pre-trained on ImageNet does not help significantly [26] , e.g","reference_id":"OTExMzcyMzU5MzIxNzQ4MzAwMw=="} +{"index":2400,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1610.02242","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2401,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1712.00559","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2402,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1608.03983","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2403,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1805.00932","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2404,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1306.5151","reference_contexts":"Stanford Cars [27] and FGVC Aircraft [38] , training with the ImageNet policy reduces test set error by 1.2% and 1.8%, respectively","reference_id":"MTUyNTMwMDg2NzE1NDkwNjI5OTQ="} +{"index":2405,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1712.04621","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2406,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"On CIFAR-10, we achieve an error rate of 1.5%, which is 0.6% better than the previous state-of-the-art [48]","reference_id":"MTQwMTk0MzQ1OTkxMjQwMjIwODM="} +{"index":2407,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"Yet a large focus of the machine learning and computer vision community has been to engineer better network architectures (e.g., [55, 59, 20, 58, 64, 19, 72, 23, 48] )","reference_id":"MTQ0NjczNTE4MzI0NjYyNDYyODM="} +{"index":2408,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"For AmoebaNets, we use the same hyperparameters that were used in [48] for both baseline augmentation and AutoAugment","reference_id":"MTAzNjk4Nzk4NzM0ODA2ODI1NTc="} +{"index":2409,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"Wide-ResNet-28-10 [67] 3.9 3.1 2.6±0.1 Shake-Shake (26 2x32d) [17] 3.6 3.0 2.5±0.1 Shake-Shake (26 2x96d) [17] 2.9 2.6 2.0±0.1 Shake-Shake (26 2x112d) [17] 2.8 2.6 1.9±0.1 AmoebaNet-B (6,128) [48] 3.0 2.1 1.8±0.","reference_id":"NjkzNTQ0MDkwMDE1NTg1NjE2Mw=="} +{"index":2410,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"The baseline pre-processing follows the convention for state-of-the-art CIFAR-10 models: standardizing the data, using horizontal flips with 50% probability, zero-padding and random crops, and finally Cutout with 16x16 pixels [17, 65, 48, 72]","reference_id":"MzYxMzI3NTMxMDI2MjMwNzk0Ng=="} +{"index":2411,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"AmoebaNet-B,C results with Inception-style preprocessing are replicated in our experiments and match the previously reported result by [48]","reference_id":"MTE0MjUyNDEzMzY5OTY0ODQ0NzA="} +{"index":2412,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"As can be seen from the results, AutoAugment improves over the widely-used Inception Pre-processing [59] across a wide range of models, from ResNet-50 to the state-of-art AmoebaNets [48]","reference_id":"MzY5MDg0NzAzNTE0MDk3MzYwMQ=="} +{"index":2413,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"Our approach is inspired by recent advances in architecture search, where reinforcement learning and evolution have been used to discover model architectures from data [71, 4, 72, 7, 35, 13, 34, 46, 49, 63, 48, 9]","reference_id":"MTU2NDAxODg0NDMxMTk4NDg2MDg="} +{"index":2414,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"It might be possible to use a different discrete search algorithm such as genetic programming [48] or even random search [6] to improve the results in this paper.","reference_id":"NjIyNTQwMzQwMDk1NDY0OTE3Nw=="} +{"index":2415,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.01548","reference_contexts":"In our experiments, we use Reinforcement Learning [71] as the search algorithm, but we believe the results can be further improved if better algorithms are used [48, 39] .","reference_id":"MTMxMDYzMjU1MjMyNTcyOTc5MzE="} +{"index":2416,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1806.00451","reference_contexts":"We also evaluate our best model trained with AutoAugment on a recently proposed CIFAR-10 test set [50]","reference_id":"MTI3ODk5Njk1ODA0NjY5MDY1OTM="} +{"index":2417,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1505.03229","reference_contexts":"For example, on MNIST, most top-ranked models use elastic distortions, scale, translation, and rotation [54, 8, 62, 52]","reference_id":"MjIxOTM3NDE4NjYwMTc3NTYzNg=="} +{"index":2418,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1707.06347","reference_contexts":"Similar to [72] , we employ Proximal Policy Optimization (PPO) [53] with learning rate 0.00035","reference_id":"MTM2MzcwMzQ5ODgwNTU5ODE1NTA="} +{"index":2419,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1707.06347","reference_contexts":"The search algorithm has two components: a controller, which is a recurrent neural network, and the training algorithm, which is the Proximal Policy Optimization algorithm [53]","reference_id":"MTU4MzE3ODYzNTY1MTEyMzYwNDY="} +{"index":2420,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.02375","reference_contexts":"The baseline pre-processing follows the convention for state-of-the-art CIFAR-10 models: standardizing the data, using horizontal flips with 50% probability, zero-padding and random crops, and finally Cutout with 16x16 pixels [17, 65, 48, 72]","reference_id":"MzYxMzI3NTMxMDI2MjMwNzk0Ng=="} +{"index":2421,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.02375","reference_contexts":"All the results of the baseline models, and baseline models with Cutout are replicated in our experiments and match the previously reported results [67, 17, 65, 12]","reference_id":"MTM5OTg5NjgzNzM1MDA0NTA1NTk="} +{"index":2422,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.02375","reference_contexts":"Again, we achieve the state-of-art result on this dataset, beating the previous record of 12.19% error rate by ShakeDrop regularization [65] .","reference_id":"MzI5NjExNDk1OTI5NjYyMzE2"} +{"index":2423,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1802.02375","reference_contexts":"We implement the Wide-ResNet-28-10 [67], Shake-Shake [17] and ShakeDrop [65] models in TensorFlow [1] , and find the weight decay and learning rate hyperparameters that give the best validation set accuracy for regular training with baseline augmentation","reference_id":"MTUyNjkwODA1MjI4MTA4Mzg5MTQ="} +{"index":2424,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1710.09412","reference_contexts":"Sample Pairing [24, 68] Linearly add the image with another image (selected at random from the same mini-batch) with weight magnitude, without changing the label.","reference_id":"MTUxNzkwNDQxMzY5ODUyMjUwMzA="} +{"index":2425,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1710.09412","reference_contexts":"[68] reports an improvement of 1.1% for a Wide-ResNet-28-10 model trained on CIFAR-10.","reference_id":"MTM5NzQ5NDkzNjA1ODg3NTU1MTE="} +{"index":2426,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1710.09412","reference_contexts":"[68] reports an improvement of 1.5% for a ResNet-50 model.","reference_id":"MTQ3NDk1NjA2NDg3MzU1NDE0"} +{"index":2427,"citer_arxiv_id":"1805.09501","cited_arxiv_id":"1708.04896","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2428,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1412.6544","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2429,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1503.02531","reference_contexts":"Several recent publications focus on reducing the test time cost of ensembles, by transferring the \"knowledge\" of cumbersome ensembles into a single model (Bucilu et al., 2006; Hinton et al., 2015)","reference_id":"NjkzODEzODQ0NTA1MDQyNzI1"} +{"index":2430,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.06993","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2431,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1412.2007","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2432,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1605.07110","reference_contexts":"It is well established (Kawaguchi, 2016) that the number of possible local minima grows exponentially with the number of parameters-of which modern neural networks can have millions","reference_id":"MTc5NTk5ODM3NjgyODAzMjMzNTY="} +{"index":2433,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1609.04836","reference_contexts":"SGD tends to avoid sharper local minima because gradients are computed from small mini-batches and are therefore inexact (Keskar et al., 2016)","reference_id":"NzMxMjUzNzgxMzU4ODg3NDc0"} +{"index":2434,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1412.6980","reference_contexts":"Stochastic Gradient Descent (SGD) (Bottou, 2010) and its accelerated variants (Kingma & Ba, 2014; Duchi et al., 2011) have become the de-facto approaches for optimizing deep neural networks","reference_id":"Mzc3MzM5NzgyMTExMzIzODg4Mw=="} +{"index":2435,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1606.01305","reference_contexts":"As an alternative to traditional ensembles, so-called \"implicit\" ensembles have high efficiency during both training and testing (Srivastava et al., 2014; Wan et al., 2013; Huang et al., 2016b; Singh et al., 2016; Krueger et al., 2016)","reference_id":"MTAyNDAwMDAzNTYzMTgyNDQ1NTk="} +{"index":2436,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1610.02242","reference_contexts":"Laine & Aila (2016) propose a temporal ensembling method for semi-supervised learning, which achieves consensus among models trained with different regularization and augmentation conditions for better generalization performance","reference_id":"MTA5NzY3MDI1ODMzOTA0Mjg4MDU="} +{"index":2437,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1605.07648","reference_contexts":"We use a standard data augmentation scheme (Lin et al., 2013; Romero et al., 2014; Lee et al., 2015; Springenberg et al., 2014; Srivastava et al., 2015; Huang et al., 2016b; Larsson et al., 2016) , in which the images are zero-padded with 4 pixels on each side, randomly cropped to produce 32×32 images, and horizontally mirrored with probability 0.5","reference_id":"MTUzMDE1ODkxOTAyMjU1MTU3MDg="} +{"index":2438,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1312.4400","reference_contexts":"We use a standard data augmentation scheme (Lin et al., 2013; Romero et al., 2014; Lee et al., 2015; Springenberg et al., 2014; Srivastava et al., 2015; Huang et al., 2016b; Larsson et al., 2016) , in which the images are zero-padded with 4 pixels on each side, randomly cropped to produce 32×32 images, and horizontally mirrored with probability 0.5","reference_id":"MTUzMDE1ODkxOTAyMjU1MTU3MDg="} +{"index":2439,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.03983","reference_contexts":"Following (Loshchilov & Hutter, 2016) , we update the learning rate at each iteration rather than at every epoch","reference_id":"MTQxOTAzNzE2NTcwMTY5MDg1NzU="} +{"index":2440,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.03983","reference_contexts":"We harness this power with the cyclical learning rate schedule proposed by Loshchilov & Hutter (2016) , saving model snapshots at each point of convergence","reference_id":"MTA4MzA4ODU0NjM0OTAzNTk2ODU="} +{"index":2441,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.03983","reference_contexts":"Our approach is inspired by the observation that training neural networks for fewer epochs and dropping the learning rate earlier has minor impact on the final test error (Loshchilov & Hutter, 2016)","reference_id":"MTE3MDUwMDM4NjkyMjUzNTEyNzQ="} +{"index":2442,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.03983","reference_contexts":"To converge to multiple local minima, we follow a cyclic annealing schedule as proposed by Loshchilov & Hutter (2016)","reference_id":"MTI1MjQxNDQ2MTg0MDk2MzI4Nw=="} +{"index":2443,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.03983","reference_contexts":"Our work is inspired by the recent findings of Loshchilov & Hutter (2016) and Smith (2016) , who show that cyclic learning rates can be effective for training convolutional neural networks","reference_id":"MTcyMjA1MjA0MDU5MzMxMDkwNDU="} +{"index":2444,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.03983","reference_contexts":"More specifically, we adopt the cycling procedure suggested by Loshchilov & Hutter (2016) , in which the learning rate is abruptly raised and then quickly lowered to follow a cosine function","reference_id":"MTg1MjgyODcyNTMyODY5NjUyMw=="} +{"index":2445,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1608.03983","reference_contexts":"In our experiments, we set f to be the shifted cosine function proposed by Loshchilov & Hutter (2016) :","reference_id":"MTM3Njc1ODM0MDY4NTAwOTE2ODc="} +{"index":2446,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1412.6550","reference_contexts":"We use a standard data augmentation scheme (Lin et al., 2013; Romero et al., 2014; Lee et al., 2015; Springenberg et al., 2014; Srivastava et al., 2015; Huang et al., 2016b; Larsson et al., 2016) , in which the images are zero-padded with 4 pixels on each side, randomly cropped to produce 32×32 images, and horizontally mirrored with probability 0.5","reference_id":"MTUzMDE1ODkxOTAyMjU1MTU3MDg="} +{"index":2447,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1606.02891","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2448,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1605.06465","reference_contexts":"As an alternative to traditional ensembles, so-called \"implicit\" ensembles have high efficiency during both training and testing (Srivastava et al., 2014; Wan et al., 2013; Huang et al., 2016b; Singh et al., 2016; Krueger et al., 2016)","reference_id":"MTAyNDAwMDAzNTYzMTgyNDQ1NTk="} +{"index":2449,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1605.06465","reference_contexts":"Finally, Swapout (Singh et al., 2016) is a stochastic training method that generalizes Dropout and Stochastic Depth","reference_id":"MTMwMDA5NDM1MDkzOTEyNzE0NjA="} +{"index":2450,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1506.01186","reference_contexts":"Our work is inspired by the recent findings of Loshchilov & Hutter (2016) and Smith (2016) , who show that cyclic learning rates can be effective for training convolutional neural networks","reference_id":"MTcyMjA1MjA0MDU5MzMxMDkwNDU="} +{"index":2451,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1412.6806","reference_contexts":"We use a standard data augmentation scheme (Lin et al., 2013; Romero et al., 2014; Lee et al., 2015; Springenberg et al., 2014; Srivastava et al., 2015; Huang et al., 2016b; Larsson et al., 2016) , in which the images are zero-padded with 4 pixels on each side, randomly cropped to produce 32×32 images, and horizontally mirrored with probability 0.5","reference_id":"MTUzMDE1ODkxOTAyMjU1MTU3MDg="} +{"index":2452,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1505.00387","reference_contexts":"We use a standard data augmentation scheme (Lin et al., 2013; Romero et al., 2014; Lee et al., 2015; Springenberg et al., 2014; Srivastava et al., 2015; Huang et al., 2016b; Larsson et al., 2016) , in which the images are zero-padded with 4 pixels on each side, randomly cropped to produce 32×32 images, and horizontally mirrored with probability 0.5","reference_id":"MTUzMDE1ODkxOTAyMjU1MTU3MDg="} +{"index":2453,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1605.07146","reference_contexts":"We test several state-of-the-art architectures, including residual networks (ResNet) (He et al., 2016a) , Wide ResNet (Zagoruyko & Komodakis, 2016) and DenseNet (Huang et al., 2016a)","reference_id":"MTAxMjk1NDE0MTg2ODgyNDczNjc="} +{"index":2454,"citer_arxiv_id":"1704.00109","cited_arxiv_id":"1605.07146","reference_contexts":"We evaluate the efficacy of Snapshot Ensembles on three state-of-the-art deep learning architectures for object recognition: ResNet (He et al., 2016b) , Wide-ResNet (Zagoruyko & Komodakis, 2016) , and DenseNet (Huang et al., 2016a)","reference_id":"Njg5MjA5NzU0NDY3ODQ0NDAxMA=="} +{"index":2455,"citer_arxiv_id":"1707.00683","cited_arxiv_id":"1605.02697","reference_contexts":"For instance, it was shown that we can improve upon classic concatenation by an element-wise product [1, 13, 14] , Tucker decomposition [2] , bilinear pooling [9] or more exotic approaches [18]","reference_id":"ODE0MzczMjc4MTY0MjIwNDA0OQ=="} +{"index":2456,"citer_arxiv_id":"1903.02271","cited_arxiv_id":"1801.01973","reference_contexts":"Although it has some flaws (Barratt & Sharma, 2018) , we report it to enable comparison with existing methods","reference_id":"MTE0Mzc2Mzc0NDk4NzIyODU0OQ=="} +{"index":2457,"citer_arxiv_id":"1903.02271","cited_arxiv_id":"1706.02677","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2458,"citer_arxiv_id":"1903.02271","cited_arxiv_id":"1606.01583","reference_contexts":"(2016) and Odena (2016) train a discriminator that classifies its input into K + 1 classes: K image classes for real images, and one class for generated images","reference_id":"MTQ0NzgyNjAwNzc2MzU2Njk3NDA="} +{"index":2459,"citer_arxiv_id":"1903.02271","cited_arxiv_id":"1708.05789","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2460,"citer_arxiv_id":"1903.02271","cited_arxiv_id":"1609.03499","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2461,"citer_arxiv_id":"1903.02271","cited_arxiv_id":"1905.03670","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2462,"citer_arxiv_id":"1603.00748","cited_arxiv_id":"1412.6980","reference_contexts":"We found the most sensitive hyperparameters to be presence or absence of batch normalization, base learning rate for ADAM (Kingma & Ba, 2014 ) ∈ {1e -4 , 1e -3 , 1e -2 }, and exploration noise scale ∈ {0.1, 0.3, 1.0}","reference_id":"MTM0MzY5OTg0OTUxMDY3MTI3OTA="} +{"index":2463,"citer_arxiv_id":"1603.00748","cited_arxiv_id":"1504.00702","reference_contexts":"In such domains, the methods of choice have been efficient model-free algorithms that use more suitable, task-specific representations (Peters et al., 2010; Deisenroth et al., 2013) , as well as model-based algorithms that learn a model of the system with supervised learning and optimize a policy under this model (Deisenroth & Rasmussen, 2011; Levine et al., 2015)","reference_id":"OTUzMDA0NjU3MzI3MjcyNTAx"} +{"index":2464,"citer_arxiv_id":"1603.00748","cited_arxiv_id":"1504.00702","reference_contexts":"Extending the benefits of time-varying linear models to less restrictive settings is a promising direction and build on prior work (Levine et al., 2015; Fu et al., 2015)","reference_id":"MjEzNzU5MTc5ODYwODk0NTU2Nw=="} +{"index":2465,"citer_arxiv_id":"1603.00748","cited_arxiv_id":"1511.05952","reference_contexts":"Deep reinforcement learning and related methods have been applied to learn policies to play Atari games (Mnih et al., 2015; Schaul et al., 2015) and perform a wide variety of simulated and real-world robotic control tasks (Hafner & Riedmiller, 2011; Lillicrap et al., 2016; Levine & Koltun, 2013; de Bruin et al., 2015; Hafner & Riedmiller, 2011)","reference_id":"MTIxMjkxODIyNDQzNzI3NTcwNg=="} +{"index":2466,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1611.01838","reference_contexts":"While this connection has held the promise of enabling new approaches to model training that yield better generalization, practical efficient algorithms that specifically seek out flatter minima and furthermore effectively improve generalization on a range of state-of-the-art models have thus far been elusive (e.g., see (Chaudhari et al., 2016; Izmailov et al., 2018) ; we include a more detailed discussion of prior work in Section 5).","reference_id":"MzY2NDExMzgzNzA0NDc5MDg3MQ=="} +{"index":2467,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1611.01838","reference_contexts":"Other ideas which also penalize sharp minima include operating on diffused loss landscape (Mobahi, 2016) and regularizing local entropy (Chaudhari et al., 2016)","reference_id":"NTA5MjYxMDkwMTUyNDkwNjM1Nw=="} +{"index":2468,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1805.09501","reference_contexts":"Relatedly, a panoply of methods for modifying the training process have been proposed, including dropout (Srivastava et al., 2014) , batch normalization (Ioffe & Szegedy, 2015) , stochastic depth (Huang et al., 2016) , data augmentation (Cubuk et al., 2018) , and mixed sample augmentations (Zhang et al., 2017; Harris et al., 2020) .","reference_id":"Njc1OTgyNDgwMzgzMDA1ODU5Mg=="} +{"index":2469,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1805.09501","reference_contexts":"We have ensured that our implementations' generalization performance in the absence of SAM matches or exceeds that reported in prior work (Cubuk et al., 2018; Lim et al., 2019) All results use basic data augmentations (horizontal flip, padding by four pixels, and random crop)","reference_id":"MzA4NTUzNDgxNzc5MDgwNjk0OA=="} +{"index":2470,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1805.09501","reference_contexts":"For auto-augment, we use the best policy found on this dataset as described in (Cubuk et al., 2018) plus cutout (Devries & Taylor, 2017)","reference_id":"ODg0NjA0MjA3NDkyOTAwMDkyMQ=="} +{"index":2471,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1708.04552","reference_contexts":"For auto-augment, we use the best policy found on this dataset as described in (Cubuk et al., 2018) plus cutout (Devries & Taylor, 2017)","reference_id":"ODg0NjA0MjA3NDkyOTAwMDkyMQ=="} +{"index":2472,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1708.04552","reference_contexts":"We also evaluate in the setting of more advanced data augmentation methods such as cutout regularization (Devries & Taylor, 2017) and AutoAugment (Cubuk et al., 2018) , which are utilized by prior work to achieve state-of-the-art results.","reference_id":"MTM5NjM2MDYwODg2NzA3MTY0NzY="} +{"index":2473,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1703.04933","reference_contexts":"The idea of searching for \"flat\" minima can be traced back to Hochreiter & Schmidhuber (1995) , and its connection to generalization has seen significant study (Shirish Keskar et al., 2016; Dziugaite & Roy, 2017; Neyshabur et al., 2017; Dinh et al., 2017)","reference_id":"MTc3Mzg0ODY1NzY5ODQ1MjIzNDk="} +{"index":2474,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"2010.11929","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2475,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1703.11008","reference_contexts":"The idea of searching for \"flat\" minima can be traced back to Hochreiter & Schmidhuber (1995) , and its connection to generalization has seen significant study (Shirish Keskar et al., 2016; Dziugaite & Roy, 2017; Neyshabur et al., 2017; Dinh et al., 2017)","reference_id":"MTc3Mzg0ODY1NzY5ODQ1MjIzNDk="} +{"index":2476,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1703.11008","reference_contexts":"The connection between the geometry of the loss landscape-in particular, the flatness of minimaand generalization has been studied extensively from both theoretical and empirical perspectives (Shirish Keskar et al., 2016; Dziugaite & Roy, 2017; Jiang et al., 2019)","reference_id":"MTE0OTU2ODg1MzA4NTk1OTAyNTE="} +{"index":2477,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1705.07485","reference_contexts":"We first evaluate SAM's impact on generalization for today's state-of-the-art models on CIFAR-10 and CIFAR-100 (without pretraining): WideResNets with ShakeShake regularization (Zagoruyko & Komodakis, 2016; Gastaldi, 2017) and PyramidNet with ShakeDrop regularization (Han et al., 2016; Yamada et al., 2018)","reference_id":"MTAyODY4MjMwMTg4NzgyNzk3OTk="} +{"index":2478,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1901.10159","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2479,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1610.02915","reference_contexts":"We first evaluate SAM's impact on generalization for today's state-of-the-art models on CIFAR-10 and CIFAR-100 (without pretraining): WideResNets with ShakeShake regularization (Zagoruyko & Komodakis, 2016; Gastaldi, 2017) and PyramidNet with ShakeDrop regularization (Han et al., 2016; Yamada et al., 2018)","reference_id":"MTAyODY4MjMwMTg4NzgyNzk3OTk="} +{"index":2480,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1512.03385","reference_contexts":"To assess SAM's performance at larger scale, we apply it to ResNets (He et al., 2015) of different depths (50, 101, 152) trained on ImageNet (Deng et al., 2009)","reference_id":"MTU1MTM0NTc5NzcwMjA0OTU1OTk="} +{"index":2481,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1512.03385","reference_contexts":"In this setting, following prior work (He et al., 2015; Szegedy et al., 2015) , we resize and crop images to 224-pixel resolution, normalize them, and use batch size 4096, initial learning rate 1.0, cosine learning rate schedule, SGD optimizer with momentum 0.9, label smoothing of 0.1, and weight decay 0.0001","reference_id":"MjUwMzc4NTY1Mzk2MzA2MTY2NQ=="} +{"index":2482,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1603.09382","reference_contexts":"Relatedly, a panoply of methods for modifying the training process have been proposed, including dropout (Srivastava et al., 2014) , batch normalization (Ioffe & Szegedy, 2015) , stochastic depth (Huang et al., 2016) , data augmentation (Cubuk et al., 2018) , and mixed sample augmentations (Zhang et al., 2017; Harris et al., 2020) .","reference_id":"Njc1OTgyNDgwMzgzMDA1ODU5Mg=="} +{"index":2483,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1811.06965","reference_contexts":"In particular, we apply SAM to finetuning EfficentNet-b7 (pretrained on ImageNet) and EfficientNet-L2 (pretrained on ImageNet plus unlabeled JFT; input resolution 475) (Tan & Le, 2019; Kornblith et al., 2018; Huang et al., 2018)","reference_id":"MTE0ODE1NDk5NDk5NTczMDc4MDg="} +{"index":2484,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1811.06965","reference_contexts":"Such heavy overparameterization is currently required to achieve stateof-the-art results in a variety of domains (Tan & Le, 2019; Kolesnikov et al., 2020; Huang et al., 2018)","reference_id":"MjAzOTA1MjcyOTU4NzMyNDc0Mw=="} +{"index":2485,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1502.03167","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2486,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1803.05407","reference_contexts":"While this connection has held the promise of enabling new approaches to model training that yield better generalization, practical efficient algorithms that specifically seek out flatter minima and furthermore effectively improve generalization on a range of state-of-the-art models have thus far been elusive (e.g., see (Chaudhari et al., 2016; Izmailov et al., 2018) ; we include a more detailed discussion of prior work in Section 5).","reference_id":"MzY2NDExMzgzNzA0NDc5MDg3MQ=="} +{"index":2487,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1806.07572","reference_contexts":"As a result, the choice of optimizer (and associated optimizer settings) from among the many available (e.g., stochastic gradient descent (Nesterov, 1983) , Adam (Kingma & Ba, 2014) , RMSProp (Hinton et al.) , and others (Duchi et al., 2011; Dozat, 2016; Martens & Grosse, 2015) ) has become an important design choice, though understanding of its relationship to model generalization remains nascent (Shirish Keskar et al., 2016; Wilson et al., 2017; Shirish Keskar & Socher, 2017; Agarwal et al., 2020; Jacot et al., 2018)","reference_id":"NTQ2OTE2NzMyOTU4NjU0MzE5Nw=="} +{"index":2488,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1412.6980","reference_contexts":"As a result, the choice of optimizer (and associated optimizer settings) from among the many available (e.g., stochastic gradient descent (Nesterov, 1983) , Adam (Kingma & Ba, 2014) , RMSProp (Hinton et al.) , and others (Duchi et al., 2011; Dozat, 2016; Martens & Grosse, 2015) ) has become an important design choice, though understanding of its relationship to model generalization remains nascent (Shirish Keskar et al., 2016; Wilson et al., 2017; Shirish Keskar & Socher, 2017; Agarwal et al., 2020; Jacot et al., 2018)","reference_id":"NTQ2OTE2NzMyOTU4NjU0MzE5Nw=="} +{"index":2489,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1805.08974","reference_contexts":"In particular, we apply SAM to finetuning EfficentNet-b7 (pretrained on ImageNet) and EfficientNet-L2 (pretrained on ImageNet plus unlabeled JFT; input resolution 475) (Tan & Le, 2019; Kornblith et al., 2018; Huang et al., 2018)","reference_id":"MTE0ODE1NDk5NDk5NTczMDc4MDg="} +{"index":2490,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1712.09913","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2491,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1905.00397","reference_contexts":"We have ensured that our implementations' generalization performance in the absence of SAM matches or exceeds that reported in prior work (Cubuk et al., 2018; Lim et al., 2019) All results use basic data augmentations (horizontal flip, padding by four pixels, and random crop)","reference_id":"MzA4NTUzNDgxNzc5MDgwNjk0OA=="} +{"index":2492,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1503.05671","reference_contexts":"As a result, the choice of optimizer (and associated optimizer settings) from among the many available (e.g., stochastic gradient descent (Nesterov, 1983) , Adam (Kingma & Ba, 2014) , RMSProp (Hinton et al.) , and others (Duchi et al., 2011; Dozat, 2016; Martens & Grosse, 2015) ) has become an important design choice, though understanding of its relationship to model generalization remains nascent (Shirish Keskar et al., 2016; Wilson et al., 2017; Shirish Keskar & Socher, 2017; Agarwal et al., 2020; Jacot et al., 2018)","reference_id":"NTQ2OTE2NzMyOTU4NjU0MzE5Nw=="} +{"index":2493,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1601.04114","reference_contexts":"Other ideas which also penalize sharp minima include operating on diffused loss landscape (Mobahi, 2016) and regularizing local entropy (Chaudhari et al., 2016)","reference_id":"NTA5MjYxMDkwMTUyNDkwNjM1Nw=="} +{"index":2494,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1811.07056","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2495,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1609.04836","reference_contexts":"As a result, the choice of optimizer (and associated optimizer settings) from among the many available (e.g., stochastic gradient descent (Nesterov, 1983) , Adam (Kingma & Ba, 2014) , RMSProp (Hinton et al.) , and others (Duchi et al., 2011; Dozat, 2016; Martens & Grosse, 2015) ) has become an important design choice, though understanding of its relationship to model generalization remains nascent (Shirish Keskar et al., 2016; Wilson et al., 2017; Shirish Keskar & Socher, 2017; Agarwal et al., 2020; Jacot et al., 2018)","reference_id":"NTQ2OTE2NzMyOTU4NjU0MzE5Nw=="} +{"index":2496,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1609.04836","reference_contexts":"The training loss landscapes of today's models are commonly complex and non-convex, with a multiplicity of local and global minima, and with different global minima yielding models with different generalization abilities (Shirish Keskar et al., 2016)","reference_id":"MTQ3MzM0NjM3NjUyMjQ4MjQzMDM="} +{"index":2497,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1609.04836","reference_contexts":"The connection between the geometry of the loss landscape-in particular, the flatness of minimaand generalization has been studied extensively from both theoretical and empirical perspectives (Shirish Keskar et al., 2016; Dziugaite & Roy, 2017; Jiang et al., 2019)","reference_id":"MTE0OTU2ODg1MzA4NTk1OTAyNTE="} +{"index":2498,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1609.04836","reference_contexts":"The idea of searching for \"flat\" minima can be traced back to Hochreiter & Schmidhuber (1995) , and its connection to generalization has seen significant study (Shirish Keskar et al., 2016; Dziugaite & Roy, 2017; Neyshabur et al., 2017; Dinh et al., 2017)","reference_id":"MTc3Mzg0ODY1NzY5ODQ1MjIzNDk="} +{"index":2499,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1905.11946","reference_contexts":"We train all models with weight decay 1e -5 as suggested in (Tan & Le, 2019 ), but we reduce the learning rate to 0.016 as the models tend to diverge for higher values","reference_id":"MTQ4NDcwMzA5NDQwMjgzODQ0OQ=="} +{"index":2500,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1905.11946","reference_contexts":"In particular, we apply SAM to finetuning EfficentNet-b7 (pretrained on ImageNet) and EfficientNet-L2 (pretrained on ImageNet plus unlabeled JFT; input resolution 475) (Tan & Le, 2019; Kornblith et al., 2018; Huang et al., 2018)","reference_id":"MTE0ODE1NDk5NDk5NTczMDc4MDg="} +{"index":2501,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1905.11946","reference_contexts":"Such heavy overparameterization is currently required to achieve stateof-the-art results in a variety of domains (Tan & Le, 2019; Kolesnikov et al., 2020; Huang et al., 2018)","reference_id":"MjAzOTA1MjcyOTU4NzMyNDc0Mw=="} +{"index":2502,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1708.07747","reference_contexts":"Beyond CIFAR-{10, 100}, we have also evaluated SAM on the SVHN (Netzer et al., 2011) and Fashion-MNIST datasets (Xiao et al., 2017)","reference_id":"OTQ4NzMyODIxNjg4NTk4MjM4Ng=="} +{"index":2503,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1802.02375","reference_contexts":"We first evaluate SAM's impact on generalization for today's state-of-the-art models on CIFAR-10 and CIFAR-100 (without pretraining): WideResNets with ShakeShake regularization (Zagoruyko & Komodakis, 2016; Gastaldi, 2017) and PyramidNet with ShakeDrop regularization (Han et al., 2016; Yamada et al., 2018)","reference_id":"MTAyODY4MjMwMTg4NzgyNzk3OTk="} +{"index":2504,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1605.07146","reference_contexts":"We first evaluate SAM's impact on generalization for today's state-of-the-art models on CIFAR-10 and CIFAR-100 (without pretraining): WideResNets with ShakeShake regularization (Zagoruyko & Komodakis, 2016; Gastaldi, 2017) and PyramidNet with ShakeDrop regularization (Han et al., 2016; Yamada et al., 2018)","reference_id":"MTAyODY4MjMwMTg4NzgyNzk3OTk="} +{"index":2505,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1611.03530","reference_contexts":"Indeed, many modern neural networks can easily memorize the training data and have the capacity to readily overfit (Zhang et al., 2016)","reference_id":"MTY3NjMyMjM0MDE0ODQxNjIzOTk="} +{"index":2506,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1710.09412","reference_contexts":"Relatedly, a panoply of methods for modifying the training process have been proposed, including dropout (Srivastava et al., 2014) , batch normalization (Ioffe & Szegedy, 2015) , stochastic depth (Huang et al., 2016) , data augmentation (Cubuk et al., 2018) , and mixed sample augmentations (Zhang et al., 2017; Harris et al., 2020) .","reference_id":"Njc1OTgyNDgwMzgzMDA1ODU5Mg=="} +{"index":2507,"citer_arxiv_id":"2010.01412","cited_arxiv_id":"1710.09412","reference_contexts":"We evaluate five variants of model training: standard SGD, SGD with Mixup (Zhang et al., 2017) , SAM, and \"bootstrapped\" variants of SGD with Mixup and SAM (wherein the model is first trained as usual and then retrained from scratch on the labels predicted by the initially trained model)","reference_id":"MTI3NjE3OTQ0NzE5NjIyMjEyNDU="} +{"index":2508,"citer_arxiv_id":"1609.03552","cited_arxiv_id":"1506.03365","reference_contexts":"Datasets: We experiment with multiple photo collections from various sources as follows: \"shoes\" dataset [37] , which has 50K shoes collected from Zappos.com (the shoes are roughly centered, but not well aligned, and roughly facing left, with frontal to side view); \"church outdoor\" dataset (126K images) from the LSUN challenge [38] ; \"outdoor natural\" images (150K) from the MIT Places dataset [39] ; and two query-based product collections downloaded from Amazon, including \"handbags\" (138K) and \"shirts\" (137K)","reference_id":"MjY0ODUxNDk4ODIzNzAzNTI4Mw=="} +{"index":2509,"citer_arxiv_id":"2103.10693","cited_arxiv_id":"1612.00410","reference_contexts":"The deep learning information bottleneck theory reveals that the essence of deep learning is to eliminate useless information and leave real effective information [1]","reference_id":"Nzk5NjkyNDQzMDEwOTIxODY5NQ=="} +{"index":2510,"citer_arxiv_id":"2103.10693","cited_arxiv_id":"1808.06670","reference_contexts":"However, this goal turns to be even more difficult [8, 17] when a whole sequence needs to be accurately reconstructed in SVAE","reference_id":"MTEwODU4MzY5NDExMjQ0NzEzMA=="} +{"index":2511,"citer_arxiv_id":"2103.10693","cited_arxiv_id":"1702.08235","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2512,"citer_arxiv_id":"2103.10693","cited_arxiv_id":"1807.03748","reference_contexts":"However, this goal turns to be even more difficult [8, 17] when a whole sequence needs to be accurately reconstructed in SVAE","reference_id":"MTEwODU4MzY5NDExMjQ0NzEzMA=="} +{"index":2513,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1809.11096","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2514,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"2002.05709","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2515,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1910.04302","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2516,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1807.00734","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2517,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1711.01558","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2518,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1807.03748","reference_contexts":"Note that unlike in contrastive self-supervised learning methods [62, 22, 6] GAN Objective (Set level fidelity) Contrastive Objective (Instance level fidelity) Contrastive Head Discriminator Head z μ σ Encoder Decoder Dual Contradistinctive Training VAE where two views (independent augmentations) of an instance constitutes a positive pair, an input instance x i and its reconstruction f θ (z) comprises a positive pair in DC-VAE","reference_id":"MTcyNTYxMDI0NzU2OTg5MzIxOQ=="} +{"index":2519,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1807.03748","reference_contexts":"Inspired by the literature on instance-level classification [50] , approximating likelihood by classification [60] , and contrastive learning [20, 65, 22] , we propose to model instance-level fidelity by contrastive loss (commonly referred to as InfoNCE loss) [62]","reference_id":"NjYwMTMwNjg3NTkxNzEzNDM3Mg=="} +{"index":2520,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1807.03748","reference_contexts":"Unlike in self-supervised representation learning methods [62, 22, 6] , where self-supervision requires generating a transformed input (via data augmentation operations), the reconstruction naturally fits into the contrastive term that encourages the matching between the reconstruction and the input image instance, while pushing the reconstruction away from the rest of the images in the entire training set","reference_id":"OTU2NTY0NTIyMDQwNzUwMDM2OQ=="} +{"index":2521,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1807.03748","reference_contexts":"Motivated by the above observations, we develop a new generative autoencoder model with dual contradistinctive losses by adopting a discriminative loss performing instance-level classification (enforcing the instance-level fidelity), which is rooted in metric learning [40] and contrastive learning [20, 65, 62]","reference_id":"MTQ4NTgyMDYzMjM5NDM0NDQyNzg="} +{"index":2522,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1807.03748","reference_contexts":"In this paper, we aim to improve VAE [37] by introducing a contrastive loss [62] to address instance-level fidelity between the input and the reconstruction in the induced feature space","reference_id":"MTU1MDgzMDgxOTUxMDI3OTc1OTc="} +{"index":2523,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1506.03365","reference_contexts":"In Figures 8 and 14 we show a large collection of additional recontruction images on the CelebA-HQ [34] and LSUN Bedroom [67] datasets.","reference_id":"NTgyMzAyMTY5MDQyODAwODg2Nw=="} +{"index":2524,"citer_arxiv_id":"2011.10063","cited_arxiv_id":"1506.03365","reference_contexts":"Datasets To validate our method, we train our method on several different datasets -CIFAR-10 [38], STL-10 [10], CelebA [48], CelebA-HQ [34] , and LSUN bedroom [67] .","reference_id":"MTg3MTc1MzIyODI5NTY1ODgwNA=="} +{"index":2525,"citer_arxiv_id":"1710.11381","cited_arxiv_id":"1707.05776","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2526,"citer_arxiv_id":"1710.11381","cited_arxiv_id":"1512.09300","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2527,"citer_arxiv_id":"1710.11381","cited_arxiv_id":"1301.3781","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2528,"citer_arxiv_id":"1710.11381","cited_arxiv_id":"1609.04468","reference_contexts":"The proposal by White (2016) is to use spherical linear interpolation","reference_id":"MTcyNzM2OTc0OTEwNDQxODQ0NzQ="} +{"index":2529,"citer_arxiv_id":"1710.11381","cited_arxiv_id":"1609.04468","reference_contexts":"(2015) ; White (2016) )","reference_id":"MTIzOTk2NTA1NDQ4NjA2NTU3OTA="} +{"index":2530,"citer_arxiv_id":"1710.11381","cited_arxiv_id":"1609.04468","reference_contexts":"We have already eluded to the proposal of White (2016) , which uses the interpolation with θ := ∠(z 0 , z 1 ),","reference_id":"Mzk1OTA4NDgzNTY0Mjc5MDk2NA=="} +{"index":2531,"citer_arxiv_id":"1710.11381","cited_arxiv_id":"1609.04468","reference_contexts":"This observation was also made in White (2016) who suggested replacing linear interpolation with spherical linear interpolation which prevents diverging from the model's prior distribution.","reference_id":"MzY1NTM4MjMzNzI0MzU4NDQ5"} +{"index":2532,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"1810.02419","reference_contexts":"Similar to image synthesis, two kinds of approaches are proposed including GAN-based (Vondrick et al., 2016; Saito et al., 2017; Acharya et al., 2018; Tulyakov et al., 2018) and transformer-based (Ge et al., 2022; Yan et al., 2021; Srivastava et al., 2015; Le Moing et al., 2021; Hong et al., 2023)","reference_id":"NTAzMjExNzM3MTg3MDk0NDMxMw=="} +{"index":2533,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"1905.10985","reference_contexts":"Automatic image synthesis is seen as a major milestone towards general artificial intelligence (Goertzel & Pennachin, 2007; Clune, 2019; Fjelland, 2020; Zhang et al., 2023)","reference_id":"NTg0OTk5OTk5NjM0MzY4MzE2MQ=="} +{"index":2534,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2302.03011","reference_contexts":"Text-to-video synthesis (Esser et al., 2023; Blattmann et al., 2023; Ge et al., 2023) recently has become an increasingly popular research topic in the field of Artificial Intelligence Generated Content (AIGC) following the success of Diffusion Models for image synthesis (Rombach et al., 2022)","reference_id":"MTQ5NDU0NzYzODgyMzQ0NDU3MTA="} +{"index":2535,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2302.03011","reference_contexts":"Recently, motivated by DM-based image synthesis, several works (Esser et al., 2023; Ge et al., 2023; Blattmann et al., 2023; Khachatryan et al., 2023; He et al., 2022; Dabral et al., 2023; Luo et al., 2023; Brooks et al., 2022) propose to explore DMs for conditional video synthesis","reference_id":"MTE4NTI3ODc1MjA0ODY1NjM3MDE="} +{"index":2536,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Since models represented by Stable Diffusion (Rombach et al., 2022) show amazing results in image synthesis, loading its pre-trained LDM is preferable for efficiently training an LDM for high-quality video synthesis (Blattmann et al., 2023; Ge et al., 2023)","reference_id":"MTM4NDk1MzM4NzE0MjkyNDQwMA=="} +{"index":2537,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Also, existing works (Ho et al., 2022; Wu et al., 2022; Ge et al., 2023; Blattmann et al., 2023) have revealed that it is beneficial to train a video generation model with image-text data for better appearances of video content","reference_id":"MTk0NzI0OTQwNzAwNDU0NjA5OA=="} +{"index":2538,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Almost at the same time, PYoCo (Ge et al., 2023) is proposed as an improved method for extending LDM from image synthesis to video synthesis","reference_id":"MTQxNzYwNzQ3MzQzNzgyMDYzMjQ="} +{"index":2539,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"As previous works have revealed, for generating smooth videos, the initial noise of LDMs for video synthesis is essential (Ge et al., 2023) , and sharing a base noise across video frames also helps (Luo et al., 2023)","reference_id":"MzU1NTYwNTQwMzM0OTc0NTA1OQ=="} +{"index":2540,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Similar to existing works like Video LDM (Blattmann et al., 2023) and PYoCo (Ge et al., 2023) , only part of our network layers are trainable for efficient training","reference_id":"NTM5NTgxODE1MTU0MjY5NDU2Mw=="} +{"index":2541,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Similar to existing works like Video LDM (Blattmann et al., 2023) and PYoCo (Ge et al., 2023) , we adapt the original U-Net for image diffusion to video synthesis by injecting temporal layers marked with dashed boxes in the figure","reference_id":"MjE0OTE4MTc4MTA1MzExMzc1"} +{"index":2542,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Text-to-video synthesis (Esser et al., 2023; Blattmann et al., 2023; Ge et al., 2023) recently has become an increasingly popular research topic in the field of Artificial Intelligence Generated Content (AIGC) following the success of Diffusion Models for image synthesis (Rombach et al., 2022)","reference_id":"MTQ5NDU0NzYzODgyMzQ0NDU3MTA="} +{"index":2543,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Recently, motivated by DM-based image synthesis, several works (Esser et al., 2023; Ge et al., 2023; Blattmann et al., 2023; Khachatryan et al., 2023; He et al., 2022; Dabral et al., 2023; Luo et al., 2023; Brooks et al., 2022) propose to explore DMs for conditional video synthesis","reference_id":"MTE4NTI3ODc1MjA0ODY1NjM3MDE="} +{"index":2544,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10474","reference_contexts":"Following previous works like Make-A-Video (Singer et al., 2023) , PY-oCo (Ge et al., 2023) and Video LDM (Blattmann et al., 2023) , the following metrics for quantitative evaluation are used: (i) Fréchet Video Distance (FVD) (Unterthiner et al., 2019) : Following Make-A-Video (Singer et al., 2023) , we use a trained I3D model (Carreira & Zisserman, 2017) for calculating FVD","reference_id":"MjI0NzUxMDQ5NzMzODM2MTkzNQ=="} +{"index":2545,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2211.13221","reference_contexts":"CogVideo (Hong et al., 2023) 5.4M 25.27 701.59 MagicVideo (Zhou et al., 2022) 27.0M -699.00 LVDM (He et al., 2022) 2.0M -641.80 ModelScope (Wang et al., 2023a) --639.90 Video LDM (Blattmann et al., 2023) 10.7M 33.45 550.61 Make-A-Video (Singer et al., 2023) 20.0M 33.00 367.23 VideoFactory (Wang et al., 2023b) 140 Quantitative Evaluation","reference_id":"MTA2Njg1NzQzMjU0MTYzMzU2NzA="} +{"index":2546,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2211.13221","reference_contexts":"Recently, motivated by DM-based image synthesis, several works (Esser et al., 2023; Ge et al., 2023; Blattmann et al., 2023; Khachatryan et al., 2023; He et al., 2022; Dabral et al., 2023; Luo et al., 2023; Brooks et al., 2022) propose to explore DMs for conditional video synthesis","reference_id":"MTE4NTI3ODc1MjA0ODY1NjM3MDE="} +{"index":2547,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2210.02303","reference_contexts":"Also, existing works (Ho et al., 2022; Wu et al., 2022; Ge et al., 2023; Blattmann et al., 2023) have revealed that it is beneficial to train a video generation model with image-text data for better appearances of video content","reference_id":"MTk0NzI0OTQwNzAwNDU0NjA5OA=="} +{"index":2548,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2210.02303","reference_contexts":"To this end, example videos are generated with the same text prompts by VidRD and the other models including Make-A-Video (Singer et al., 2023) , Imagen Video (Ho et al., 2022)","reference_id":"MTczNDk2NTA5NDIyNzg1NTEwODg="} +{"index":2549,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2302.09778","reference_contexts":"Due to the surprising results of DMs, massive works (Kawar et al., 2023; Bhunia et al., 2023; Liu et al., 2023b; Fan et al., 2023; Dabral et al., 2023; Huang et al., 2023; Ramesh et al., 2022; Saharia et al., 2022) focus on this technology","reference_id":"NzAyNzUwMjc1NzEyMTAyNDE0Ng=="} +{"index":2550,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2303.13439","reference_contexts":"Moreover, the training of LDMs usually relies on a massive amount of data to ensure the quality of the generative content (Khachatryan et al., 2023)","reference_id":"MTc1NTQxNTkwMzMxMTI2NDk3NDk="} +{"index":2551,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2303.13439","reference_contexts":"Recently, motivated by DM-based image synthesis, several works (Esser et al., 2023; Ge et al., 2023; Blattmann et al., 2023; Khachatryan et al., 2023; He et al., 2022; Dabral et al., 2023; Luo et al., 2023; Brooks et al., 2022) propose to explore DMs for conditional video synthesis","reference_id":"MTE4NTI3ODc1MjA0ODY1NjM3MDE="} +{"index":2552,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2301.12597","reference_contexts":"For this problem, we employ an LLM with visual comprehension capabilities: BLIP-2 (Li et al., 2023)","reference_id":"MjE1NTEwMTk1MDA2MDU4NTYzMw=="} +{"index":2553,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2303.04761","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2554,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2302.01329","reference_contexts":"With the satisfactory results of LDM-based video synthesis, there are also some works (Molad et al., 2023; Qi et al., 2023; Liu et al., 2023a) on controllable video editing.","reference_id":"NjcwODk0MDU5OTQwMDgzMTY5"} +{"index":2555,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2112.10741","reference_contexts":"GLIDE (Nichol et al., 2021) and Stable Diffusion (Rombach et al., 2022) are two representative DM-based works that employ Vision-Language Models (VLMs) such as CLIP (Radford et al., 2021) for text-guided image synthesis","reference_id":"Mzk0NjM1MzM3MzE2NjAyNTAxMg=="} +{"index":2556,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2303.09535","reference_contexts":"With the satisfactory results of LDM-based video synthesis, there are also some works (Molad et al., 2023; Qi et al., 2023; Liu et al., 2023a) on controllable video editing.","reference_id":"NjcwODk0MDU5OTQwMDgzMTY5"} +{"index":2557,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2204.06125","reference_contexts":"Due to the surprising results of DMs, massive works (Kawar et al., 2023; Bhunia et al., 2023; Liu et al., 2023b; Fan et al., 2023; Dabral et al., 2023; Huang et al., 2023; Ramesh et al., 2022; Saharia et al., 2022) focus on this technology","reference_id":"NzAyNzUwMjc1NzEyMTAyNDE0Ng=="} +{"index":2558,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2010.10864","reference_contexts":"For those video datasets mostly containing short videos originally used for action recognition such as Kinetics-700 (Smaira et al., 2020) , the problem is how to give a proper text caption based on its classification label to each video","reference_id":"MTI5MzY1NjU5ODcyMTQ4MjIyMjI="} +{"index":2559,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2010.10864","reference_contexts":"2) Short video classification datasets: Moments-In-Time (Monfort et al., 2021) and Kinetics-700 (Smaira et al., 2020)","reference_id":"MTE1ODk2ODA2MTkwMzA4NjYxNTA="} +{"index":2560,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"1212.0402","reference_contexts":"Following previous works (Singer et al., 2023; Hong et al., 2023; Blattmann et al., 2023) , we use UCF-101 (Soomro et al., 2012) , a dataset for video recognition, for evaluating FVD and IS","reference_id":"MzMxODEzMDU3Nzg4OTIyMDk0OA=="} +{"index":2561,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"1212.0402","reference_contexts":"On the benchmark of video generation based on UCF-101 (Soomro et al., 2012) , we achieve Fréchet Video Distance (FVD) of 363.19 and Inception Score (IS) of 39.37.","reference_id":"MTU1MDAzOTYxMTgyODA2MDM3MjM="} +{"index":2562,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2308.06571","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2563,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2305.10874","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2564,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2104.10157","reference_contexts":"In VideoGPT (Yan et al., 2021) , VQ-VAE (Van Den Oord et al., 2016) is used to learn subsampled discrete latent representations of raw videos","reference_id":"MTc3NTIzNjM0OTI2NTc1NjUxMTk="} +{"index":2565,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2104.10157","reference_contexts":"Similar to image synthesis, two kinds of approaches are proposed including GAN-based (Vondrick et al., 2016; Saito et al., 2017; Acharya et al., 2018; Tulyakov et al., 2018) and transformer-based (Ge et al., 2022; Yan et al., 2021; Srivastava et al., 2015; Le Moing et al., 2021; Hong et al., 2023)","reference_id":"NTAzMjExNzM3MTg3MDk0NDMxMw=="} +{"index":2566,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2303.07909","reference_contexts":"However, the autoregressive nature of the above methods results in high computation costs and sequential error accumulation (Zhang et al., 2023)","reference_id":"MTY4ODE5NzMyMTEwMzc2NzQ2OTI="} +{"index":2567,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2303.07909","reference_contexts":"Automatic image synthesis is seen as a major milestone towards general artificial intelligence (Goertzel & Pennachin, 2007; Clune, 2019; Fjelland, 2020; Zhang et al., 2023)","reference_id":"NTg0OTk5OTk5NjM0MzY4MzE2MQ=="} +{"index":2568,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2211.11018","reference_contexts":"CogVideo (Hong et al., 2023) 5.4M 25.27 701.59 MagicVideo (Zhou et al., 2022) 27.0M -699.00 LVDM (He et al., 2022) 2.0M -641.80 ModelScope (Wang et al., 2023a) --639.90 Video LDM (Blattmann et al., 2023) 10.7M 33.45 550.61 Make-A-Video (Singer et al., 2023) 20.0M 33.00 367.23 VideoFactory (Wang et al., 2023b) 140 Quantitative Evaluation","reference_id":"MTA2Njg1NzQzMjU0MTYzMzU2NzA="} +{"index":2569,"citer_arxiv_id":"2309.03549","cited_arxiv_id":"2304.10592","reference_contexts":"For each video containing multiple frames, we employ CLIP (Radford et al., 2021) with vision-language alignment ability and MiniGPT-4 (Zhu et al., 2023) , an LLM with vision-language understanding to mark those frames irrelevant to the classification label","reference_id":"MTE4MTA1MzU0MjU0MzUxMTE1Mjg="} +{"index":2570,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2211.09707","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2571,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2004.05150","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2572,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2212.04495","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2573,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2207.12598","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2574,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2209.00349","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2575,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2204.06125","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2576,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2209.14916","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2577,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2212.02500","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2578,"citer_arxiv_id":"2305.04919","cited_arxiv_id":"2208.15001","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2579,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2310.00426","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2580,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2307.09481","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2581,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2210.11416","reference_contexts":"DALLE-3 (Betker et al., 2023) recaptions the training data pairs and utilizes T5-XXL (Chung et al., 2022) as the text encoder to strengthen the prompt-following ability.","reference_id":"ODcwMjIzMjcyNDQ0NDA3NjE0MA=="} +{"index":2582,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2303.05499","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2583,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2310.02992","reference_contexts":"Emu2 (Sun et al., 2023) , Kosmos-G (Pan et al., 2023) , and λ-ECLIPSE (Patel et al., 2024) are all MLLM-based methods, while λ-ECLIPSE is reported to have better performance","reference_id":"MTQ5MjcwOTY2NTc3NjkzMDYwNDE="} +{"index":2584,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2310.02992","reference_contexts":"To narrow the gap between image and text prompts, Kosmos-G (Pan et al., 2023) and λ-ECLIPSE (Patel et al., 2024) conduct multi-modal training to unite the inputs by text-image interleaving","reference_id":"MTEwMzU4MDgyMjEzMTYyNzc2MzM="} +{"index":2585,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2310.02992","reference_contexts":"Furthermore, recent works have employed multi-subject driven customization methodologies to concatenate visual concepts with textual prompts, offering a glimpse of the potential in techniques like SSR-Encoder (Zhang et al., 2024) , λ-ECLIPSE (Patel et al., 2024) , Emu2 (Sun et al., 2023) , and KOSMOS-G (Pan et al., 2023)","reference_id":"ODE3NTg0OTgyMTA5NDMxNTI2"} +{"index":2586,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2402.05195","reference_contexts":"• λ-ECLIPSE (Patel et al., 2024) trains an independent multi-modal encoder and employs Kandinsky as the generative backbone","reference_id":"OTg5ODg4NDc3NTUxMTE2MTU1NQ=="} +{"index":2587,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2402.05195","reference_contexts":"Emu2 (Sun et al., 2023) , Kosmos-G (Pan et al., 2023) , and λ-ECLIPSE (Patel et al., 2024) are all MLLM-based methods, while λ-ECLIPSE is reported to have better performance","reference_id":"MTQ5MjcwOTY2NTc3NjkzMDYwNDE="} +{"index":2588,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2402.05195","reference_contexts":"Furthermore, recent works have employed multi-subject driven customization methodologies to concatenate visual concepts with textual prompts, offering a glimpse of the potential in techniques like SSR-Encoder (Zhang et al., 2024) , λ-ECLIPSE (Patel et al., 2024) , Emu2 (Sun et al., 2023) , and KOSMOS-G (Pan et al., 2023)","reference_id":"ODE3NTg0OTgyMTA5NDMxNTI2"} +{"index":2589,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2402.05195","reference_contexts":"For single-subject, the results for IP-Adapter, Emu2, and Kosmos-G were obtained from λ-ECLIPSE (Patel et al., 2024)","reference_id":"MTQ5MjEwMjk0NjczNjE1ODQ3NTQ="} +{"index":2590,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2402.05195","reference_contexts":"To narrow the gap between image and text prompts, Kosmos-G (Pan et al., 2023) and λ-ECLIPSE (Patel et al., 2024) conduct multi-modal training to unite the inputs by text-image interleaving","reference_id":"MTEwMzU4MDgyMjEzMTYyNzc2MzM="} +{"index":2591,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2307.01952","reference_contexts":"The pre-trained model employed in MS-Diffusion is Stable Diffusion XL (SDXL) (Podell et al., 2023)","reference_id":"ODg5MjY1NzM0Nzg1NDk5MTk3Mg=="} +{"index":2592,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2307.01952","reference_contexts":"Text-to-image generative models (Saharia et al., 2022; Bao et al., 2023; Esser et al., 2024; Podell et al., 2023) are capable of producing high-quality images using user-provided text prompts","reference_id":"ODAzMjI1Nzk3Mzc0NDA4ODAzNw=="} +{"index":2593,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2307.01952","reference_contexts":"In this work, we employ L IP with SDXL (Podell et al., 2023) as the pre-trained model, which contains two CLIP text encoders and additional condition inputs besides c and t, omitted for brevity.","reference_id":"MTcxMjQxNDc0OTEwNzY5MDM1MTA="} +{"index":2594,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2312.13286","reference_contexts":"Emu2 (Sun et al., 2023) , Kosmos-G (Pan et al., 2023) , and λ-ECLIPSE (Patel et al., 2024) are all MLLM-based methods, while λ-ECLIPSE is reported to have better performance","reference_id":"MTQ5MjcwOTY2NTc3NjkzMDYwNDE="} +{"index":2595,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2312.13286","reference_contexts":"Furthermore, recent works have employed multi-subject driven customization methodologies to concatenate visual concepts with textual prompts, offering a glimpse of the potential in techniques like SSR-Encoder (Zhang et al., 2024) , λ-ECLIPSE (Patel et al., 2024) , Emu2 (Sun et al., 2023) , and KOSMOS-G (Pan et al., 2023)","reference_id":"ODE3NTg0OTgyMTA5NDMxNTI2"} +{"index":2596,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2401.07519","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2597,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2312.03626","reference_contexts":"While some approaches (Wang et al., 2023; 2024b; Zhou et al., 2024; Kim et al., 2023) in non-personalized text-to-image tasks address multi-object generation conflicts through text cross-attention, this inevitably requires tuning the diffusion model's parameters, thereby affecting the plug-and-play nature, which is not preferred by us","reference_id":"MzI3MTU4NjQwMjg3NjQ3MTQyNg=="} +{"index":2598,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2312.03626","reference_contexts":"Recent works (Wang et al., 2023; Wei et al., 2024) have studied utilizing an objective on the cross-attention maps in multi-subject generation","reference_id":"MTQ4Nzg0MzY1MDMzNTMyNTYxOTA="} +{"index":2599,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2308.06721","reference_contexts":"Single-subject driven personalization methods, IP-Adapter (Ye et al., 2023) and ELITE (Wei et al., 2023) for instance, introduce a specialized cross-attention mechanism that distinctly processes text and image features, thereby affording the possibility to employ reference images directly as visual prompts within the model","reference_id":"NTYyMDQ0NjEyODU2ODc3OTk0"} +{"index":2600,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2308.06721","reference_contexts":"IP-Adapter (Ye et al., 2023) proposes a zero-shot personalized model by projecting the image embedding to the crossattention layers","reference_id":"NzQ3NjM0NjU3NzY5MDczNDcyOA=="} +{"index":2601,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2308.06721","reference_contexts":"In zero-shot image personalization architectures like IP-Adapter (Ye et al., 2023) , images can also be considered a condition of the diffusion model","reference_id":"MTA1MjExMjcyNjczNzI2NTkzNjM="} +{"index":2602,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2308.06721","reference_contexts":"For GRS, we replace it with a linear projection layer and a normal resampler (Alayrac et al., 2022; Ye et al., 2023)","reference_id":"MTAxNzM1NjcyODgxMTI4MjU1NTc="} +{"index":2603,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2308.06721","reference_contexts":"Following the training of IP-adapter (Ye et al., 2023) , we set γ = 1.0 in cross-attention layers and dropped the text and image condition using the same probability","reference_id":"MTYwNzQ3NDg5NjU3OTIxMDQ3MDI="} +{"index":2604,"citer_arxiv_id":"2406.07209","cited_arxiv_id":"2308.06721","reference_contexts":"Different from Flamingo (Alayrac et al., 2022) and IP-Adapter (Ye et al., 2023) , we propose the integration of a grounding resampler that functions as an alternative form of image projector","reference_id":"NjEwODY3MTE4MDY5OTkyMzY2Nw=="} +{"index":2605,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2309.15564","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2606,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2004.05150","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2607,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2204.02311","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2608,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2309.11499","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2609,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2010.11929","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2610,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2307.08041","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2611,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2305.18247","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2612,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2106.09685","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2613,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2403.18814","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2614,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2311.07575","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2615,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2306.00973","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2616,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2310.03744","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2617,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2310.03744","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2618,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2304.08485","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2619,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2110.10834","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2620,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2105.10026","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2621,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2306.14824","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2622,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2306.14824","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2623,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2307.01952","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2624,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2307.05222","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2625,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2401.10208","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2626,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2302.13971","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2627,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2309.05519","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2628,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2309.17453","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2629,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2309.02591","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2630,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2309.15112","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2631,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2304.10592","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2632,"citer_arxiv_id":"2407.08683","cited_arxiv_id":"2312.09251","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2633,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2209.15571","reference_contexts":"Nowadays, Diffusion Models (DMs) [133] , learning to invert diffusion paths from real data towards random noise, have gradually become the de-facto approach of generative AI across multiple domains, with numerous practical applications [106, 8, 49, 107, 1, 114, 44, 2]","reference_id":"MTM0NTEzOTAxNjE2MjU3MzM4ODQ="} +{"index":2634,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2209.15571","reference_contexts":"Specifically, the foundational text-to-image model, Lumina-T2I, utilizes the flow matching framework [92, 86, 4] and is trained on a meticulously curated dataset of highresolution photorealistic image-text pairs, achieving remarkably realistic results with merely a small proportion of computational resources","reference_id":"MTExODYwMTY5NTU0MzAyMDczMzg="} +{"index":2635,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2303.08797","reference_contexts":"Apart from generating high-quality images following text instruction, various applications, including high-resolution generation [55, 43, 69, 170, 33, 25] , compositional generation [74, 12, 162] , style-consistent generation [58, 143] , image editing [57, 18, 78, 102] , and controllable generation [164, 103, 168, 101] , have been proposed to further extend the applicability of pretrained T2I models","reference_id":"MTM2MDYyNDQ3MzA5OTU2MzY1NzA="} +{"index":2636,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2303.08797","reference_contexts":"In contrast, flow matching [86, 5] emerges as a simple alternative that linearly interpolates between noise and data in a straight line","reference_id":"MTYwNjMyMDgyODQzMTIwOTU1MTQ="} +{"index":2637,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2305.10403","reference_contexts":"However, the largest parameter size of DiT is only limited at 600M which is far less than LLMs (e.g., PaLM-540B [35, 7] , Grok-1-300B, LLaMa3-400B [145, 146] )","reference_id":"MTM5NjI0NzE0MTk0ODIwNTY0MTE="} +{"index":2638,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1607.06450","reference_contexts":"Specifically, Flag-DiT substitutes all LayerNorm [9] with RMSNorm [163] to improve training stability","reference_id":"MTI1MjExNzk4MTU2MjE5NTcxNzY="} +{"index":2639,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.15127","reference_contexts":"To bridge these gaps, we present Lumina-T2X, a family of Flow-based Large Diffusion Transformers (Flag-DiT) designed to transform noise into images [114, 123] , videos [14, 108] , multi-views of 3D objects [131, 130] , and audio clips [138] based on textual instructions","reference_id":"ODQ3MzQ1MDM2NzgzMDY5ODMwOQ=="} +{"index":2640,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.15127","reference_contexts":"It operates over latent patches extracted from a pretrained VAE [79, 14] , then utilizes a transformer [150, 111] as denoising backbone to predict the mean and variance according to DDPM formulation [134, 135, 61, 105] from different levels of noised latent patches conditioned on time steps and class labels","reference_id":"MTg4NDcyNDk5NDE5NDA3NTE0Ng=="} +{"index":2641,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.15127","reference_contexts":"Analysis of Gate Distribution in Zero-Initialized Attention Cross-attention [139, 14] is the defacto standard for text conditioning","reference_id":"MTY2NzU2ODkwODUyODA5Njk1OTE="} +{"index":2642,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.15127","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2643,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.15127","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2644,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1809.11096","reference_contexts":"In the era of deep learning, Generative Adversarial Networks (GANs) [48, 173, 70, 155, 17, 76] stand as a pioneering method in this field due to their efficient sampling capabilities, yet they face issues of","reference_id":"MTI3MzM4NzY2MjQzMzIxMjk5NTM="} +{"index":2645,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.19512","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2646,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.19512","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2647,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.09047","reference_contexts":"The success of diffusion models over the past four years can be attributed to the progress in several areas, including reformulating diffusion models to predict noise instead of pixels [61] , improvements in sampling methods for better efficiency [134, 94, 95, 77] , the introduction of classifier-free guidance that enables direct conversion of text to images [59] , and cascaded/latent space models that reduce the computational cost of high-resolution generation [63, 123, 142]","reference_id":"MTc5NzU4MDY3NTU5ODA5MTc1MDA="} +{"index":2648,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.09047","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2649,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.00426","reference_contexts":"[13, 24, 25] , our Lumina-T2X model is able to generate high-resolution images and coherent videos with minimal computational demands","reference_id":"ODEzMDgyMDIzOTk1ODg4MzA5Mw=="} +{"index":2650,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.00426","reference_contexts":"Influence of ImageNet Initialization PixArt-α [24, 25] utilizes ImageNet-pretrained DiT, which learns pixel dependency, as an initialization for the subsequent T2I model","reference_id":"NDM2MjUxNDYxOTQwOTM4NzA0MQ=="} +{"index":2651,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.00426","reference_contexts":"Conversely, Pixart-α [24] , which uses standard positional embeddings instead of RoPE [136] , does not show comparable generalization capabilities at test resolutions","reference_id":"MTYwMzI5MTk4NzMzMDQ2ODg3NjA="} +{"index":2652,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.00426","reference_contexts":"Comparison with Pixart-α Compared to PixArt-α [24] , Lumina-T2I can generate images at resolutions ranging from 512 2 pixels to 1792 2 pixels","reference_id":"MTczNDI1OTk3ODgzMzgwMjY4NjY="} +{"index":2653,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.00426","reference_contexts":"Motivated by previous approaches [62, 24] , Lumina-T2I, Lumina-T2V, and Lumina-T2MV employ a multi-stage training approach, starting from low-resolution, short-duration data while ending with high-resolution, long-duration data","reference_id":"MTE2MDMxMzU5ODkwMDk4NDM1MTc="} +{"index":2654,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.00426","reference_contexts":"Recent advancements in foundational diffusion models, such as Sora [108], Stable Diffusion 3 [44] , PixArt-α [24] , and PixArt-Σ [25] , have yielded remarkable success in generating photorealistic images and videos","reference_id":"ODQ2MjI2MTExMTM1MzM4Mjk4OQ=="} +{"index":2655,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.00426","reference_contexts":"Built on DiT, PixArt-α [24] unleashes this potential for generating images based on arbitrary textual instructions","reference_id":"MTQ1Mzg5Mzc0NzAyOTMyODU3MjA="} +{"index":2656,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.04692","reference_contexts":"[13, 24, 25] , our Lumina-T2X model is able to generate high-resolution images and coherent videos with minimal computational demands","reference_id":"ODEzMDgyMDIzOTk1ODg4MzA5Mw=="} +{"index":2657,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.04692","reference_contexts":"Influence of ImageNet Initialization PixArt-α [24, 25] utilizes ImageNet-pretrained DiT, which learns pixel dependency, as an initialization for the subsequent T2I model","reference_id":"NDM2MjUxNDYxOTQwOTM4NzA0MQ=="} +{"index":2658,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.04692","reference_contexts":"Therefore, high-resolution fine-tuning/adaptation [170, 33, 50, 25] and tuning-free high-resolution generation [55, 43, 53, 69, 66] are the mainstream choices today","reference_id":"MTc1MTE3NzQ5MjQ4NTc5OTQwNDk="} +{"index":2659,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.04692","reference_contexts":"Recent advancements in foundational diffusion models, such as Sora [108], Stable Diffusion 3 [44] , PixArt-α [24] , and PixArt-Σ [25] , have yielded remarkable success in generating photorealistic images and videos","reference_id":"ODQ2MjI2MTExMTM1MzM4Mjk4OQ=="} +{"index":2660,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.04692","reference_contexts":"Apart from generating high-quality images following text instruction, various applications, including high-resolution generation [55, 43, 69, 170, 33, 25] , compositional generation [74, 12, 162] , style-consistent generation [58, 143] , image editing [57, 18, 78, 102] , and controllable generation [164, 103, 168, 101] , have been proposed to further extend the applicability of pretrained T2I models","reference_id":"MTM2MDYyNDQ3MzA5OTU2MzY1NzA="} +{"index":2661,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.04557","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2662,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.04557","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2663,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2306.15595","reference_contexts":"This echoes the findings on LLMs with RoPE -the long-range decay of RoPE is insufficient to suppress the anomalies brought about by unseen context lengths [30]","reference_id":"NzU1NjA5OTU4NTQ1Mjc4NzY2OA=="} +{"index":2664,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2306.15595","reference_contexts":"Many subsequent techniques further explore and unlock this potential, e.g., position interpolation [30] , NTK-aware scaled RoPE [3] , Yarn [112] , etc","reference_id":"Njk4NDgxMTQwMzE4MzU2ODQ1OQ=="} +{"index":2665,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2306.15595","reference_contexts":"Flow-based Large Diffusion Transformers (Flag-DiT): Lumina-T2X utilizes the Flag-DiT architecture inspired by the core design principles from Large Language Models (LLMs) [145, 146, 19, 117, 122, 141, 166] , such as scalable architecture [19, 150, 56, 163, 136, 36] and context window extension [112, 136, 30, 3] for increasing parameter size and sequence length","reference_id":"MTE3NDg4ODc4NjczMTc5NTY1MDk="} +{"index":2666,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2306.15595","reference_contexts":"Motivated by recent LLMs exhibiting strong context extrapolation capabilities [112, 136, 30, 3] , we replace APE with RoPE [136] which injects relative position information in a layerwise manner, following Equations 1 and 2.","reference_id":"MTEwMDQxMDU2MTA1NDI1MDQ2"} +{"index":2667,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2402.19479","reference_contexts":"Lumina-T2V is independently trained on a subset of the Panda-70M dataset [31] and the collected Pexel dataset, comprising of 15 million and 40,000 videos, respectively","reference_id":"NjIzMDA1MDg1NTE3ODc1NDY3OA=="} +{"index":2668,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1904.10509","reference_contexts":"Following this, autoregressive models (ARMs) [149, 148, 34, 26] have shown exceptional performance but come with higher computational demands, and the sequential sampling mechanism is more suited to 1-D data.","reference_id":"MTY3MDU2MjQ1MjAyOTYxMzEyMDg="} +{"index":2669,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1410.8516","reference_contexts":"Meanwhile, Variational Autoencoders (VAEs) [79, 82, 6, 147, 128] and flow-based models [40, 41] demonstrate better training stability and interpretability but lag behind GANs in terms of image quality","reference_id":"MTc5OTc4OTAwODYzODUyMTg0MzQ="} +{"index":2670,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1605.08803","reference_contexts":"Meanwhile, Variational Autoencoders (VAEs) [79, 82, 6, 147, 128] and flow-based models [40, 41] demonstrate better training stability and interpretability but lag behind GANs in terms of image quality","reference_id":"MTc5OTc4OTAwODYzODUyMTg0MzQ="} +{"index":2671,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2010.11929","reference_contexts":"We find that this stems from the design choice that DiT leverages absolute positional embedding (APE) [42, 144] and adds it to latent tokens in the first layer following vision transformers","reference_id":"MTc2MzY1MzI4NzU2MTIyODk0NDg="} +{"index":2672,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.03206","reference_contexts":"Previous works [77, 44] have pointed out that the learning of score function in diffusion models or velocity field in flow matching is more challenging in the middle of the schedule","reference_id":"NTg0MDQxMjYyMDkzMjkzNTQw"} +{"index":2673,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.03206","reference_contexts":"Inspired by the observation that intermediate timesteps are critical for both diffusion models [77] and flow-based models [44] , we adopt the time resampling strategy to sample timestep from a log-norm distribution during training","reference_id":"MTExODYzMTU4OTc4Njk2MDQwMTM="} +{"index":2674,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.03206","reference_contexts":"Nowadays, Diffusion Models (DMs) [133] , learning to invert diffusion paths from real data towards random noise, have gradually become the de-facto approach of generative AI across multiple domains, with numerous practical applications [106, 8, 49, 107, 1, 114, 44, 2]","reference_id":"MTM0NTEzOTAxNjE2MjU3MzM4ODQ="} +{"index":2675,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.03206","reference_contexts":"Recent advancements in foundational diffusion models, such as Sora [108], Stable Diffusion 3 [44] , PixArt-α [24] , and PixArt-Σ [25] , have yielded remarkable success in generating photorealistic images and videos","reference_id":"ODQ2MjI2MTExMTM1MzM4Mjk4OQ=="} +{"index":2676,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.03206","reference_contexts":"(1) NTK-aware scaled RoPE [3] that rescales the rotary base of RoPE to achieve a gradual position interpolation of the low-frequency components, (2) Time Shifting [44] that reschedules the timesteps to ensure consistent SNR across denoising processes of different resolutions, and (3) Proportional Attention [75] that rescales the attention score to ensure stable attention entropy across various sequence lengths","reference_id":"MTE2MTM2NDA5MzcwMTE1MzIzNDU="} +{"index":2677,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.03206","reference_contexts":"Afterwards, SiT [98] and SD3 [44] further demonstrate that an interpolation or flow-matching framework [92, 86, 4, 5] can better enhance the stability and scalability of DiT -pointing the way for diffusion models to scale up to the next level.","reference_id":"MjAwMTA1ODgwMTk3MzQxNjc0MQ=="} +{"index":2678,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.03206","reference_contexts":"This coincides with the Time Shifting schedule in [44] and other counterparts in diffusion literature [64, 69]","reference_id":"MTU3Mjc0ODkwNjA1NTEwMTE3MTU="} +{"index":2679,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2304.15010","reference_contexts":"Besides simple label conditioning for class-conditioned generation, Flag-DiT can flexibly support arbitrary text instruction with zero-initialized attention [165, 45, 164, 10]","reference_id":"MTQ3MTY3Nzg2MTAzNzQ0Nzk3Ng=="} +{"index":2680,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2304.15010","reference_contexts":"Unlike previous methods, Lumina-T2I employs zero-initialized attention mechanism [45, 165] , which incorporates a zero-initialized gating mechanism to adaptively control the influence of text-conditioning across various heads and layers","reference_id":"NDQxNTY1OTQ3NDkwNzM2Mzc2"} +{"index":2681,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2402.05935","reference_contexts":"The largest model within the Lumina-T2X family comprises a Flag-DiT with 7 billion parameters and a multi-modal large language model, SPHINX [46, 85] , as the text encoder, with 13 billion parameters, capable of handling 128K tokens","reference_id":"MTExNjU1MzEyMTE0NzAzNjg3NQ=="} +{"index":2682,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2304.13731","reference_contexts":"The similar framework, with suitable adjustments, has also been applied to audio generation [67, 90, 47, 161]","reference_id":"NzA4MTMzNjg1NjU5MzYyMzY0OQ=="} +{"index":2683,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.04725","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2684,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.04725","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2685,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.06662","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2686,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.06662","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2687,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2208.01626","reference_contexts":"Apart from generating high-quality images following text instruction, various applications, including high-resolution generation [55, 43, 69, 170, 33, 25] , compositional generation [74, 12, 162] , style-consistent generation [58, 143] , image editing [57, 18, 78, 102] , and controllable generation [164, 103, 168, 101] , have been proposed to further extend the applicability of pretrained T2I models","reference_id":"MTM2MDYyNDQ3MzA5OTU2MzY1NzA="} +{"index":2688,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2208.01626","reference_contexts":"High-Resolution Editing Following the methods outlined in Section 2.5, we perform style and subject editing on high-resolution images [57, 18, 78, 129]","reference_id":"MTMwMDQzNjA2MjM3MDg5ODMwNTY="} +{"index":2689,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2208.01626","reference_contexts":"As shown in Figure 1 , Lumina-T2I can generate highquality images at arbitrary resolutions and aspect ratios, and further enables advanced functionalities including resolution extrapolation [43, 55] , high-resolution editing [57, 18, 78, 129] , compositional generation [12, 162] , and style-consistent generation [58, 143] , all of which are seamlessly integrated into the framework in a training-free manner","reference_id":"MjMxMjY5MjMzOTE1NjYwNzU3Mg=="} +{"index":2690,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2208.01626","reference_contexts":"High-Resolution Editing Beyond high-resolution generation, our Lumina-T2I can also perform image editing [57, 18] , especially for high-resolution images","reference_id":"MTA5MTUwODQ3ODU5MzU5NzMwNTQ="} +{"index":2691,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.02133","reference_contexts":"As shown in Figure 1 , Lumina-T2I can generate highquality images at arbitrary resolutions and aspect ratios, and further enables advanced functionalities including resolution extrapolation [43, 55] , high-resolution editing [57, 18, 78, 129] , compositional generation [12, 162] , and style-consistent generation [58, 143] , all of which are seamlessly integrated into the framework in a training-free manner","reference_id":"MjMxMjY5MjMzOTE1NjYwNzU3Mg=="} +{"index":2692,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.02133","reference_contexts":"Style-Consistent Generation Batch generation of style-consistent content holds immense value for practical application scenarios [58, 143]","reference_id":"NDM1NzA5MzM4NDkxODM1MDM3NA=="} +{"index":2693,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.02133","reference_contexts":"A representative approach is shared attention [58] , which enables generating style-aligned batches without specific tuning of the model","reference_id":"MTgxNzA4OTkwMjY2Njg1NTM4OQ=="} +{"index":2694,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.02133","reference_contexts":"Apart from generating high-quality images following text instruction, various applications, including high-resolution generation [55, 43, 69, 170, 33, 25] , compositional generation [74, 12, 162] , style-consistent generation [58, 143] , image editing [57, 18, 78, 102] , and controllable generation [164, 103, 168, 101] , have been proposed to further extend the applicability of pretrained T2I models","reference_id":"MTM2MDYyNDQ3MzA5OTU2MzY1NzA="} +{"index":2695,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.02133","reference_contexts":"Thanks to the full-attention model architecture, we can obtain results comparable to those in [58] without using any tricks such as Adaptive Instance Normalization (AdaIN) [68]","reference_id":"NTIzOTE1ODg4NDQ0MDc2MjQ3Mw=="} +{"index":2696,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2207.12598","reference_contexts":"Comparison with SOTA Approaches As shown in Table 2 , Large-DiT-7B significantly surpasses all approaches on FID and IS score without using Classifier-free Guidance (CFG) [60] , reducing the FID score from 8.60 to 6.09","reference_id":"MTA5NDU5MzU4MTgzNTI1NzgxODI="} +{"index":2697,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02303","reference_contexts":"Motivated by previous approaches [62, 24] , Lumina-T2I, Lumina-T2V, and Lumina-T2MV employ a multi-stage training approach, starting from low-resolution, short-duration data while ending with high-resolution, long-duration data","reference_id":"MTE2MDMxMzU5ODkwMDk4NDM1MTc="} +{"index":2698,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02303","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2699,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02303","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2700,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.17117","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2701,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.17117","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2702,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.14509","reference_contexts":"As Flag-DiT is a pure transformer-based architecture, it can borrow the well-validated parallel strategies [132, 121, 169, 88, 87, 89, 72] designed for LLMs, including FSDP [169] and sequence parallel [88, 87, 89, 72] to support large parameter scales and longer sequences","reference_id":"MTM1ODg1Mzc0ODQxMTAyODQzNzM="} +{"index":2703,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2302.02412","reference_contexts":"Compositional Generation Compositional, or multi-concepts text-to-image generation [74, 12, 162] , which requires the model to generate multiple subjects at different regions of a single image, is seamlessly supported by our transformer-based framework","reference_id":"MTUxNDM2Mzg5MjE5ODE0NDg2NzA="} +{"index":2704,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2302.02412","reference_contexts":"Apart from generating high-quality images following text instruction, various applications, including high-resolution generation [55, 43, 69, 170, 33, 25] , compositional generation [74, 12, 162] , style-consistent generation [58, 143] , image editing [57, 18, 78, 102] , and controllable generation [164, 103, 168, 101] , have been proposed to further extend the applicability of pretrained T2I models","reference_id":"MTM2MDYyNDQ3MzA5OTU2MzY1NzA="} +{"index":2705,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1312.6114","reference_contexts":"Meanwhile, Variational Autoencoders (VAEs) [79, 82, 6, 147, 128] and flow-based models [40, 41] demonstrate better training stability and interpretability but lag behind GANs in terms of image quality","reference_id":"MTc5OTc4OTAwODYzODUyMTg0MzQ="} +{"index":2706,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1312.6114","reference_contexts":"It operates over latent patches extracted from a pretrained VAE [79, 14] , then utilizes a transformer [150, 111] as denoising backbone to predict the mean and variance according to DDPM formulation [134, 135, 61, 105] from different levels of noised latent patches conditioned on time steps and class labels","reference_id":"MTg4NDcyNDk5NDE5NDA3NTE0Ng=="} +{"index":2707,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.06214","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2708,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.07575","reference_contexts":"The largest model within the Lumina-T2X family comprises a Flag-DiT with 7 billion parameters and a multi-modal large language model, SPHINX [46, 85] , as the text encoder, with 13 billion parameters, capable of handling 128K tokens","reference_id":"MTExNjU1MzEyMTE0NzAzNjg3NQ=="} +{"index":2709,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02747","reference_contexts":"Specifically, the foundational text-to-image model, Lumina-T2I, utilizes the flow matching framework [92, 86, 4] and is trained on a meticulously curated dataset of highresolution photorealistic image-text pairs, achieving remarkably realistic results with merely a small proportion of computational resources","reference_id":"MTExODYwMTY5NTU0MzAyMDczMzg="} +{"index":2710,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02747","reference_contexts":"Next, we switch from the DDPM setting in DiT to the flow matching formulation [98, 92, 86] , offering another flexibility to Flag-DiT","reference_id":"MTcyNDA3MjI3NzczMzk0MjM2MzU="} +{"index":2711,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02747","reference_contexts":"Afterwards, SiT [98] and SD3 [44] further demonstrate that an interpolation or flow-matching framework [92, 86, 4, 5] can better enhance the stability and scalability of DiT -pointing the way for diffusion models to scale up to the next level.","reference_id":"MjAwMTA1ODgwMTk3MzQxNjc0MQ=="} +{"index":2712,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02747","reference_contexts":"which is named Conditional Flow Matching loss [86] , sharing similarity with the noise prediction or score prediction losses in diffusion models.","reference_id":"MTAyMjIxNzY4MDM5ODk3OTEyMzU="} +{"index":2713,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02747","reference_contexts":"Moreover, Flag-DiT improves upon the original DiT by adopting the flow matching formulation [98, 86] , which builds continuous-time diffusion paths via linear interpolation between noise and data","reference_id":"NDIxMzc2MjA5MTk2NzU2NTQ0Ng=="} +{"index":2714,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.02747","reference_contexts":"In contrast, flow matching [86, 5] emerges as a simple alternative that linearly interpolates between noise and data in a straight line","reference_id":"MTYwNjMyMDgyODQzMTIwOTU1MTQ="} +{"index":2715,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.01889","reference_contexts":"As Flag-DiT is a pure transformer-based architecture, it can borrow the well-validated parallel strategies [132, 121, 169, 88, 87, 89, 72] designed for LLMs, including FSDP [169] and sequence parallel [88, 87, 89, 72] to support large parameter scales and longer sequences","reference_id":"MTM1ODg1Mzc0ODQxMTAyODQzNzM="} +{"index":2716,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2402.08268","reference_contexts":"As Flag-DiT is a pure transformer-based architecture, it can borrow the well-validated parallel strategies [132, 121, 169, 88, 87, 89, 72] designed for LLMs, including FSDP [169] and sequence parallel [88, 87, 89, 72] to support large parameter scales and longer sequences","reference_id":"MTM1ODg1Mzc0ODQxMTAyODQzNzM="} +{"index":2717,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2301.12503","reference_contexts":"The similar framework, with suitable adjustments, has also been applied to audio generation [67, 90, 47, 161]","reference_id":"NzA4MTMzNjg1NjU5MzYyMzY0OQ=="} +{"index":2718,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2209.03003","reference_contexts":"Next, we switch from the DDPM setting in DiT to the flow matching formulation [98, 92, 86] , offering another flexibility to Flag-DiT","reference_id":"MTcyNDA3MjI3NzczMzk0MjM2MzU="} +{"index":2719,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2209.03003","reference_contexts":"Afterwards, SiT [98] and SD3 [44] further demonstrate that an interpolation or flow-matching framework [92, 86, 4, 5] can better enhance the stability and scalability of DiT -pointing the way for diffusion models to scale up to the next level.","reference_id":"MjAwMTA1ODgwMTk3MzQxNjc0MQ=="} +{"index":2720,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2209.03003","reference_contexts":"Specifically, the foundational text-to-image model, Lumina-T2I, utilizes the flow matching framework [92, 86, 4] and is trained on a meticulously curated dataset of highresolution photorealistic image-text pairs, achieving remarkably realistic results with merely a small proportion of computational resources","reference_id":"MTExODYwMTY5NTU0MzAyMDczMzg="} +{"index":2721,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2211.01095","reference_contexts":"The success of diffusion models over the past four years can be attributed to the progress in several areas, including reformulating diffusion models to predict noise instead of pixels [61] , improvements in sampling methods for better efficiency [134, 94, 95, 77] , the introduction of classifier-free guidance that enables direct conversion of text to images [59] , and cascaded/latent space models that reduce the computational cost of high-resolution generation [63, 123, 142]","reference_id":"MTc5NzU4MDY3NTU5ODA5MTc1MDA="} +{"index":2722,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.17172","reference_contexts":"Moreover, it incorporates keyquery normalization (KQ-Norm) [36, 56, 96] before key-query dot product attention computation","reference_id":"MTAyNzczMTU5MDgyMjQyNjMwMDE="} +{"index":2723,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.08740","reference_contexts":"Next, we switch from the DDPM setting in DiT to the flow matching formulation [98, 92, 86] , offering another flexibility to Flag-DiT","reference_id":"MTcyNDA3MjI3NzczMzk0MjM2MzU="} +{"index":2724,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.08740","reference_contexts":"Furthermore, SiT [98] utilizes the stochastic interpolant framework (or flow matching) to connect different distributions in a more flexible manner than DDPM","reference_id":"MTUwMDU0NTk5NzIxNTk3NDEyODg="} +{"index":2725,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.08740","reference_contexts":"Afterwards, SiT [98] and SD3 [44] further demonstrate that an interpolation or flow-matching framework [92, 86, 4, 5] can better enhance the stability and scalability of DiT -pointing the way for diffusion models to scale up to the next level.","reference_id":"MjAwMTA1ODgwMTk3MzQxNjc0MQ=="} +{"index":2726,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.08740","reference_contexts":"Moreover, Flag-DiT improves upon the original DiT by adopting the flow matching formulation [98, 86] , which builds continuous-time diffusion paths via linear interpolation between noise and data","reference_id":"NDIxMzc2MjA5MTk2NzU2NTQ0Ng=="} +{"index":2727,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1710.03740","reference_contexts":"Besides, DiT requires full precision training which doubles the GPU memory costs and training speed compared with mixed precision training [99]","reference_id":"MTczMDAxNzUzNTA1MTcwNzcyNDI="} +{"index":2728,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2103.03841","reference_contexts":"We additionally report sFID [124] , Inception Score [104] , and Precision/Recall [83] for an extensive evaluation.","reference_id":"MTM1NzI5ODA4MTEzMDk2NTY3MA=="} +{"index":2729,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.00071","reference_contexts":"Many subsequent techniques further explore and unlock this potential, e.g., position interpolation [30] , NTK-aware scaled RoPE [3] , Yarn [112] , etc","reference_id":"Njk4NDgxMTQwMzE4MzU2ODQ1OQ=="} +{"index":2730,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.00071","reference_contexts":"Flow-based Large Diffusion Transformers (Flag-DiT): Lumina-T2X utilizes the Flag-DiT architecture inspired by the core design principles from Large Language Models (LLMs) [145, 146, 19, 117, 122, 141, 166] , such as scalable architecture [19, 150, 56, 163, 136, 36] and context window extension [112, 136, 30, 3] for increasing parameter size and sequence length","reference_id":"MTE3NDg4ODc4NjczMTc5NTY1MDk="} +{"index":2731,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.00071","reference_contexts":"Motivated by recent LLMs exhibiting strong context extrapolation capabilities [112, 136, 30, 3] , we replace APE with RoPE [136] which injects relative position information in a layerwise manner, following Equations 1 and 2.","reference_id":"MTEwMDQxMDU2MTA1NDI1MDQ2"} +{"index":2732,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.01952","reference_contexts":"To bridge these gaps, we present Lumina-T2X, a family of Flow-based Large Diffusion Transformers (Flag-DiT) designed to transform noise into images [114, 123] , videos [14, 108] , multi-views of 3D objects [131, 130] , and audio clips [138] based on textual instructions","reference_id":"ODQ3MzQ1MDM2NzgzMDY5ODMwOQ=="} +{"index":2733,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.01952","reference_contexts":"Although this paradigm has achieved notable success at the current model scale [114, 113, 171] , subsequent works have proven the better potential of diffusion models based on vision transformers (so-called Diffusion Transformer, DiT) [111]","reference_id":"MzM4MzE4OTAyNTI1NzU1NzU1OA=="} +{"index":2734,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.01952","reference_contexts":"Nowadays, Diffusion Models (DMs) [133] , learning to invert diffusion paths from real data towards random noise, have gradually become the de-facto approach of generative AI across multiple domains, with numerous practical applications [106, 8, 49, 107, 1, 114, 44, 2]","reference_id":"MTM0NTEzOTAxNjE2MjU3MzM4ODQ="} +{"index":2735,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.01952","reference_contexts":"PixArt-α significantly reduces training costs compared with SDXL [114] and Raphael [159] , while maintaining high sample quality","reference_id":"MTIyODk3ODU2MDMxMzc1MTI2MDY="} +{"index":2736,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.05530","reference_contexts":"Flow-based Large Diffusion Transformers (Flag-DiT): Lumina-T2X utilizes the Flag-DiT architecture inspired by the core design principles from Large Language Models (LLMs) [145, 146, 19, 117, 122, 141, 166] , such as scalable architecture [19, 150, 56, 163, 136, 36] and context window extension [112, 136, 30, 3] for increasing parameter size and sequence length","reference_id":"MTE3NDg4ODc4NjczMTc5NTY1MDk="} +{"index":2737,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2111.02114","reference_contexts":"This suggests that synthetic images from JourneyDB and natural images collected online (e.g., LAION [126, 127] , COYO [20] , SAM [80] , and ImageNet [38] ) belong to distinct distributions","reference_id":"MTY3ODgxNzEzNjU0MjM1NTk1Mzc="} +{"index":2738,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.10089","reference_contexts":"High-Resolution Editing Following the methods outlined in Section 2.5, we perform style and subject editing on high-resolution images [57, 18, 78, 129]","reference_id":"MTMwMDQzNjA2MjM3MDg5ODMwNTY="} +{"index":2739,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.10089","reference_contexts":"As shown in Figure 1 , Lumina-T2I can generate highquality images at arbitrary resolutions and aspect ratios, and further enables advanced functionalities including resolution extrapolation [43, 55] , high-resolution editing [57, 18, 78, 129] , compositional generation [12, 162] , and style-consistent generation [58, 143] , all of which are seamlessly integrated into the framework in a training-free manner","reference_id":"MjMxMjY5MjMzOTE1NjYwNzU3Mg=="} +{"index":2740,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.15110","reference_contexts":"To bridge these gaps, we present Lumina-T2X, a family of Flow-based Large Diffusion Transformers (Flag-DiT) designed to transform noise into images [114, 123] , videos [14, 108] , multi-views of 3D objects [131, 130] , and audio clips [138] based on textual instructions","reference_id":"ODQ3MzQ1MDM2NzgzMDY5ODMwOQ=="} +{"index":2741,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.15110","reference_contexts":"Following Zero123++ [130] , we put the 12 rendered images into a single large image in the form of a 3 × 4 grid","reference_id":"NzkxODE0NTY0NzQ0ODk1MzgyNg=="} +{"index":2742,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.15110","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2743,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2308.16512","reference_contexts":"To bridge these gaps, we present Lumina-T2X, a family of Flow-based Large Diffusion Transformers (Flag-DiT) designed to transform noise into images [114, 123] , videos [14, 108] , multi-views of 3D objects [131, 130] , and audio clips [138] based on textual instructions","reference_id":"ODQ3MzQ1MDM2NzgzMDY5ODMwOQ=="} +{"index":2744,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2308.16512","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2745,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"1909.08053","reference_contexts":"As Flag-DiT is a pure transformer-based architecture, it can borrow the well-validated parallel strategies [132, 121, 169, 88, 87, 89, 72] designed for LLMs, including FSDP [169] and sequence parallel [88, 87, 89, 72] to support large parameter scales and longer sequences","reference_id":"MTM1ODg1Mzc0ODQxMTAyODQzNzM="} +{"index":2746,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2010.02502","reference_contexts":"The success of diffusion models over the past four years can be attributed to the progress in several areas, including reformulating diffusion models to predict noise instead of pixels [61] , improvements in sampling methods for better efficiency [134, 94, 95, 77] , the introduction of classifier-free guidance that enables direct conversion of text to images [59] , and cascaded/latent space models that reduce the computational cost of high-resolution generation [63, 123, 142]","reference_id":"MTc5NzU4MDY3NTU5ODA5MTc1MDA="} +{"index":2747,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2010.02502","reference_contexts":"It operates over latent patches extracted from a pretrained VAE [79, 14] , then utilizes a transformer [150, 111] as denoising backbone to predict the mean and variance according to DDPM formulation [134, 135, 61, 105] from different levels of noised latent patches conditioned on time steps and class labels","reference_id":"MTg4NDcyNDk5NDE5NDA3NTE0Ng=="} +{"index":2748,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2011.13456","reference_contexts":"Thus plenty of diffusion schedules are carefully designed and used, including VE [135] , VP [61] , and EDM [77] .","reference_id":"MjAyNjMwOTM1MDkyOTEwNDQx"} +{"index":2749,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2011.13456","reference_contexts":"It operates over latent patches extracted from a pretrained VAE [79, 14] , then utilizes a transformer [150, 111] as denoising backbone to predict the mean and variance according to DDPM formulation [134, 135, 61, 105] from different levels of noised latent patches conditioned on time steps and class labels","reference_id":"MTg4NDcyNDk5NDE5NDA3NTE0Ng=="} +{"index":2750,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2210.04885","reference_contexts":"Analysis of Gate Distribution in Zero-Initialized Attention Cross-attention [139, 14] is the defacto standard for text conditioning","reference_id":"MTY2NzU2ODkwODUyODA5Njk1OTE="} +{"index":2751,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.11805","reference_contexts":"Flow-based Large Diffusion Transformers (Flag-DiT): Lumina-T2X utilizes the Flag-DiT architecture inspired by the core design principles from Large Language Models (LLMs) [145, 146, 19, 117, 122, 141, 166] , such as scalable architecture [19, 150, 56, 163, 136, 36] and context window extension [112, 136, 30, 3] for increasing parameter size and sequence length","reference_id":"MTE3NDg4ODc4NjczMTc5NTY1MDk="} +{"index":2752,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.03350","reference_contexts":"The success of diffusion models over the past four years can be attributed to the progress in several areas, including reformulating diffusion models to predict noise instead of pixels [61] , improvements in sampling methods for better efficiency [134, 94, 95, 77] , the introduction of classifier-free guidance that enables direct conversion of text to images [59] , and cascaded/latent space models that reduce the computational cost of high-resolution generation [63, 123, 142]","reference_id":"MTc5NzU4MDY3NTU5ODA5MTc1MDA="} +{"index":2753,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.03350","reference_contexts":"However, similar to the observation in diffusion schedules [142, 64, 69] , we found that the high-resolution images are less corrupted and retain the global structure for a wider range of time under the linear interpolation schedule in flow matching","reference_id":"MTgyMjg1NDU4ODAxNjA5NDE3NTg="} +{"index":2754,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2402.03286","reference_contexts":"Apart from generating high-quality images following text instruction, various applications, including high-resolution generation [55, 43, 69, 170, 33, 25] , compositional generation [74, 12, 162] , style-consistent generation [58, 143] , image editing [57, 18, 78, 102] , and controllable generation [164, 103, 168, 101] , have been proposed to further extend the applicability of pretrained T2I models","reference_id":"MTM2MDYyNDQ3MzA5OTU2MzY1NzA="} +{"index":2755,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2402.03286","reference_contexts":"As shown in Figure 1 , Lumina-T2I can generate highquality images at arbitrary resolutions and aspect ratios, and further enables advanced functionalities including resolution extrapolation [43, 55] , high-resolution editing [57, 18, 78, 129] , compositional generation [12, 162] , and style-consistent generation [58, 143] , all of which are seamlessly integrated into the framework in a training-free manner","reference_id":"MjMxMjY5MjMzOTE1NjYwNzU3Mg=="} +{"index":2756,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2402.03286","reference_contexts":"Style-Consistent Generation Batch generation of style-consistent content holds immense value for practical application scenarios [58, 143]","reference_id":"NDM1NzA5MzM4NDkxODM1MDM3NA=="} +{"index":2757,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2302.13971","reference_contexts":"Flow-based Large Diffusion Transformers (Flag-DiT): Lumina-T2X utilizes the Flag-DiT architecture inspired by the core design principles from Large Language Models (LLMs) [145, 146, 19, 117, 122, 141, 166] , such as scalable architecture [19, 150, 56, 163, 136, 36] and context window extension [112, 136, 30, 3] for increasing parameter size and sequence length","reference_id":"MTE3NDg4ODc4NjczMTc5NTY1MDk="} +{"index":2758,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2302.13971","reference_contexts":"➀ Stability Flag-DiT builds on top of DiT [111] and incorporates modifications from ViT-22B [36] and LLaMa [145, 146] to improve the training stability","reference_id":"Mzk1Nzc3MzgwNDU1ODYwOTUyMQ=="} +{"index":2759,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2302.13971","reference_contexts":"It is worth noting that the default Lumina-T2I configuration, equipped with a 5 billion Flag-DiT and a 7 billion LLaMA [145, 146] as its text encoder, requires only 35% of the computational resources compared to PixArt-α, which builds upon a 600 million DiT backbone and 3 billion T5 [120] as its text encoder","reference_id":"NzMxMjMxMDc0NjYzNzI4Njk5Ng=="} +{"index":2760,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2302.13971","reference_contexts":"By default, all images in this technical report are generated using a 5B Flag-DiT coupled with a 7B LLaMa text encoder [145, 146]","reference_id":"NDEwNTM3MDY3NjY2NzkyNjMyNg=="} +{"index":2761,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2302.13971","reference_contexts":"However, the largest parameter size of DiT is only limited at 600M which is far less than LLMs (e.g., PaLM-540B [35, 7] , Grok-1-300B, LLaMa3-400B [145, 146] )","reference_id":"MTM5NjI0NzE0MTk0ODIwNTY0MTE="} +{"index":2762,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.09288","reference_contexts":"Flow-based Large Diffusion Transformers (Flag-DiT): Lumina-T2X utilizes the Flag-DiT architecture inspired by the core design principles from Large Language Models (LLMs) [145, 146, 19, 117, 122, 141, 166] , such as scalable architecture [19, 150, 56, 163, 136, 36] and context window extension [112, 136, 30, 3] for increasing parameter size and sequence length","reference_id":"MTE3NDg4ODc4NjczMTc5NTY1MDk="} +{"index":2763,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.09288","reference_contexts":"➀ Stability Flag-DiT builds on top of DiT [111] and incorporates modifications from ViT-22B [36] and LLaMa [145, 146] to improve the training stability","reference_id":"Mzk1Nzc3MzgwNDU1ODYwOTUyMQ=="} +{"index":2764,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.09288","reference_contexts":"It is worth noting that the default Lumina-T2I configuration, equipped with a 5 billion Flag-DiT and a 7 billion LLaMA [145, 146] as its text encoder, requires only 35% of the computational resources compared to PixArt-α, which builds upon a 600 million DiT backbone and 3 billion T5 [120] as its text encoder","reference_id":"NzMxMjMxMDc0NjYzNzI4Njk5Ng=="} +{"index":2765,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.09288","reference_contexts":"By default, all images in this technical report are generated using a 5B Flag-DiT coupled with a 7B LLaMa text encoder [145, 146]","reference_id":"NDEwNTM3MDY3NjY2NzkyNjMyNg=="} +{"index":2766,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2307.09288","reference_contexts":"However, the largest parameter size of DiT is only limited at 600M which is far less than LLMs (e.g., PaLM-540B [35, 7] , Grok-1-300B, LLaMa3-400B [145, 146] )","reference_id":"MTM5NjI0NzE0MTk0ODIwNTY0MTE="} +{"index":2767,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2301.02111","reference_contexts":"Evaluation We report word error rate (WER) to evaluate the intelligibility of speech by transcribing it using a whisper [119] ASR system following [152]","reference_id":"NjgxNTYwMTUzOTEyNzkzMTAzNA=="} +{"index":2768,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2308.06571","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2769,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2308.06571","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2770,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2312.02201","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2771,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.15103","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2772,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2309.15103","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2773,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.12190","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2774,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2310.12190","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2775,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.11708","reference_contexts":"Compositional Generation As illustrated in Figure 9 , we present demos of compositional generation [162, 12] using the method described in Section 2.5","reference_id":"MTA0NDYwNDAyODM2ODU0MDU0ODQ="} +{"index":2776,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.11708","reference_contexts":"As shown in Figure 1 , Lumina-T2I can generate highquality images at arbitrary resolutions and aspect ratios, and further enables advanced functionalities including resolution extrapolation [43, 55] , high-resolution editing [57, 18, 78, 129] , compositional generation [12, 162] , and style-consistent generation [58, 143] , all of which are seamlessly integrated into the framework in a training-free manner","reference_id":"MjMxMjY5MjMzOTE1NjYwNzU3Mg=="} +{"index":2777,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.11708","reference_contexts":"Compositional Generation Compositional, or multi-concepts text-to-image generation [74, 12, 162] , which requires the model to generate multiple subjects at different regions of a single image, is seamlessly supported by our transformer-based framework","reference_id":"MTUxNDM2Mzg5MjE5ODE0NDg2NzA="} +{"index":2778,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2401.11708","reference_contexts":"Apart from generating high-quality images following text instruction, various applications, including high-resolution generation [55, 43, 69, 170, 33, 25] , compositional generation [74, 12, 162] , style-consistent generation [58, 143] , image editing [57, 18, 78, 102] , and controllable generation [164, 103, 168, 101] , have been proposed to further extend the applicability of pretrained T2I models","reference_id":"MTM2MDYyNDQ3MzA5OTU2MzY1NzA="} +{"index":2779,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2303.16199","reference_contexts":"Besides simple label conditioning for class-conditioned generation, Flag-DiT can flexibly support arbitrary text instruction with zero-initialized attention [165, 45, 164, 10]","reference_id":"MTQ3MTY3Nzg2MTAzNzQ0Nzk3Ng=="} +{"index":2780,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2303.16199","reference_contexts":"Unlike previous methods, Lumina-T2I employs zero-initialized attention mechanism [45, 165] , which incorporates a zero-initialized gating mechanism to adaptively control the influence of text-conditioning across various heads and layers","reference_id":"NDQxNTY1OTQ3NDkwNzM2Mzc2"} +{"index":2781,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.14624","reference_contexts":"Flow-based Large Diffusion Transformers (Flag-DiT): Lumina-T2X utilizes the Flag-DiT architecture inspired by the core design principles from Large Language Models (LLMs) [145, 146, 19, 117, 122, 141, 166] , such as scalable architecture [19, 150, 56, 163, 136, 36] and context window extension [112, 136, 30, 3] for increasing parameter size and sequence length","reference_id":"MTE3NDg4ODc4NjczMTc5NTY1MDk="} +{"index":2782,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.04145","reference_contexts":"Departing from previous approaches [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] that rely on T2I checkpoints for T2V initialization and adopt decoupled spatial-temporal attention, Lumina-T2V takes a different route by initializing the Flag-DiT weights randomly and leveraging a full-attention mechanism that allows for interaction among all spatial-temporal tokens","reference_id":"NTkxOTM3OTU3NzI2ODQ2MTU0Mw=="} +{"index":2783,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2311.04145","reference_contexts":"Additionally, pre-trained T2I models are also applied with a decoupled temporal attention to generate videos [51, 156, 73, 22, 23, 16, 172, 62, 15, 65, 29, 153, 167, 158, 52, 157] and multi-views of 3D object [131, 84, 154, 174, 32, 151, 54, 93, 140, 91, 130]","reference_id":"OTQ4NjUxMzEzNzUzNzU5NjQ2MQ=="} +{"index":2784,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2304.11277","reference_contexts":"As Flag-DiT is a pure transformer-based architecture, it can borrow the well-validated parallel strategies [132, 121, 169, 88, 87, 89, 72] designed for LLMs, including FSDP [169] and sequence parallel [88, 87, 89, 72] to support large parameter scales and longer sequences","reference_id":"MTM1ODg1Mzc0ODQxMTAyODQzNzM="} +{"index":2785,"citer_arxiv_id":"2405.05945","cited_arxiv_id":"2403.05121","reference_contexts":"Although this paradigm has achieved notable success at the current model scale [114, 113, 171] , subsequent works have proven the better potential of diffusion models based on vision transformers (so-called Diffusion Transformer, DiT) [111]","reference_id":"MzM4MzE4OTAyNTI1NzU1NzU1OA=="} +{"index":2786,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2107.03374","reference_contexts":"Therefore, we default to pass@𝑘 evaluation (Chen et al., 2021) and report pass@1 using a non-zero temperature","reference_id":"MTQ1MDEyNjk1Mjg1MzA5ODg2OTM="} +{"index":2787,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2407.21783","reference_contexts":"We open-source the distilled Qwen and Llama (Dubey et al., 2024) series","reference_id":"NjkzNDc4MzM0NDA5MTUwOTg0MQ=="} +{"index":2788,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2404.04475","reference_contexts":"Specifically, we adhere to the original configurations of AlpacaEval 2.0 (Dubois et al., 2024) and Arena-Hard (Li et al., 2024) , which leverage GPT-4-Turbo-1106 as judges for pairwise comparisons","reference_id":"MTI5MjcyOTEzMjE2Mzc0MTQ1NDI="} +{"index":2789,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2210.10760","reference_contexts":"Third, once a model-based PRM is introduced, it inevitably leads to reward hacking (Gao et al., 2022) , and retraining the reward model needs additional training resources and it complicates the whole training pipeline","reference_id":"MTI3MzEwMTE0MjA0MDUxMDY3NTM="} +{"index":2790,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2406.04127","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2791,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2009.03300","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2792,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2305.08322","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2793,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2403.07974","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2794,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2409.12917","reference_contexts":"Several prior works have explored various approaches, including process-based reward models (Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023) , reinforcement learning (Kumar et al., 2024) , and search algorithms such as Monte Carlo Tree Search and Beam Search (Feng et al., 2024; Trinh et al., 2024; Xin et al., 2024)","reference_id":"ODI2NjczMDA2Njc0NjQ4NjcwNg=="} +{"index":2795,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2306.09212","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2796,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2406.11939","reference_contexts":"Specifically, we adhere to the original configurations of AlpacaEval 2.0 (Dubois et al., 2024) and Arena-Hard (Li et al., 2024) , which leverage GPT-4-Turbo-1106 as judges for pairwise comparisons","reference_id":"MTI5MjcyOTEzMjE2Mzc0MTQ1NDI="} +{"index":2797,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2305.20050","reference_contexts":"PRM is a reasonable method to guide the model toward better approaches for solving reasoning tasks (Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023)","reference_id":"MTczNjcyNTA3OTYxODAyOTI3NDY="} +{"index":2798,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2305.20050","reference_contexts":"Several prior works have explored various approaches, including process-based reward models (Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023) , reinforcement learning (Kumar et al., 2024) , and search algorithms such as Monte Carlo Tree Search and Beam Search (Feng et al., 2024; Trinh et al., 2024; Xin et al., 2024)","reference_id":"ODI2NjczMDA2Njc0NjQ4NjcwNg=="} +{"index":2799,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2311.12022","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2800,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2402.03300","reference_contexts":"Group Relative Policy Optimization In order to save the training costs of RL, we adopt Group Relative Policy Optimization (GRPO) (Shao et al., 2024) , which foregoes the critic model that is typically the same size as the policy model, and estimates the baseline from group scores instead","reference_id":"OTk4MDYwOTE4OTgwNjMwNTcwMA=="} +{"index":2801,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2402.03300","reference_contexts":"Reinforcement learning has demonstrated significant effectiveness in reasoning tasks, as evidenced by our previous works (Shao et al., 2024; Wang et al., 2023)","reference_id":"NzQyMDA3MTAxMTY4MjYyMDQ4OQ=="} +{"index":2802,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2402.03300","reference_contexts":"Specifically, we use DeepSeek-V3-Base as the base model and employ GRPO (Shao et al., 2024) as the RL framework to improve model performance in reasoning.","reference_id":"MTExNjk5MDU4NTM1OTkzNTYwMDg="} +{"index":2803,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"1712.01815","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2804,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2211.14275","reference_contexts":"PRM is a reasonable method to guide the model toward better approaches for solving reasoning tasks (Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023)","reference_id":"MTczNjcyNTA3OTYxODAyOTI3NDY="} +{"index":2805,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2211.14275","reference_contexts":"Several prior works have explored various approaches, including process-based reward models (Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023) , reinforcement learning (Kumar et al., 2024) , and search algorithms such as Monte Carlo Tree Search and Beam Search (Feng et al., 2024; Trinh et al., 2024; Xin et al., 2024)","reference_id":"ODI2NjczMDA2Njc0NjQ4NjcwNg=="} +{"index":2806,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2312.08935","reference_contexts":"Reinforcement learning has demonstrated significant effectiveness in reasoning tasks, as evidenced by our previous works (Shao et al., 2024; Wang et al., 2023)","reference_id":"NzQyMDA3MTAxMTY4MjYyMDQ4OQ=="} +{"index":2807,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2312.08935","reference_contexts":"PRM is a reasonable method to guide the model toward better approaches for solving reasoning tasks (Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023)","reference_id":"MTczNjcyNTA3OTYxODAyOTI3NDY="} +{"index":2808,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2312.08935","reference_contexts":"Several prior works have explored various approaches, including process-based reward models (Lightman et al., 2023; Uesato et al., 2022; Wang et al., 2023) , reinforcement learning (Kumar et al., 2024) , and search algorithms such as Monte Carlo Tree Search and Beam Search (Feng et al., 2024; Trinh et al., 2024; Xin et al., 2024)","reference_id":"ODI2NjczMDA2Njc0NjQ4NjcwNg=="} +{"index":2809,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2203.11171","reference_contexts":"For AIME 2024, we also report consensus (majority vote) results (Wang et al., 2022) using 64 samples, denoted as cons@64.","reference_id":"MTIyODY1MTY0Njg2NTQ3MTQ2OTg="} +{"index":2810,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2406.01574","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2811,"citer_arxiv_id":"2501.12948","cited_arxiv_id":"2311.07911","reference_contexts":"Benchmarks We evaluate models on MMLU (Hendrycks et al., 2020) , MMLU-Redux (Gema et al., 2024) , MMLU-Pro (Wang et al., 2024) , C-Eval (Huang et al., 2023) , and CMMLU (Li et al., 2023) , IFEval (Zhou et al., 2023) , FRAMES (Krishna et al., 2024) , GPQA Diamond (Rein et al., 2023) , SimpleQA (OpenAI, 2024c), C-SimpleQA (He et al., 2024) , SWE-Bench Verified (OpenAI, 2024d) , Aider foot_0 , LiveCodeBench (Jain et al., 2024 (Jain et al., ) (2024 (Jain et al., -08 -2025-01)-01) , Codeforces foot_1 , Chinese National High School Mathematics Olympiad (CNMO 2024) foot_2 , and American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024)","reference_id":"NDU5MDYzODcyMDI0NTcwOTYyMA=="} +{"index":2812,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2309.16609","reference_contexts":"Here, we adopt the Qwen [4] with 500M parameters as the decoder because it has a relatively small number of parameters while incorporating prior knowledge of multiple languages","reference_id":"NzE5MzExMjAzMTI0NDgxMjMxMw=="} +{"index":2813,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2308.12966","reference_contexts":"In the past year, Large Vision Language models (LVLMs) [5, 9, 24, 27, 36, 46, 49] have developed rapidly and showcased impressive performance","reference_id":"MTAwMjMzMjA0NDUxNDk1MDUzNTk="} +{"index":2814,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2308.12966","reference_contexts":"To lift the OCR accuracy and support other languages, e.g., Chinese, Qwen-VL [5] unfreezes its image encoder (a CLIP-G) and uses lots of OCR data in its stage-two training","reference_id":"Nzc3MjA1Mzg2NTI2Njg0Mzg0MQ=="} +{"index":2815,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2308.12966","reference_contexts":"Large Vision-Language Models (LVLMs) [5, 9, 20, 24, 27, 46, 49] have attracted lots of attention in the AI-community due to their powerful generalization capabilities","reference_id":"Njk5ODg4MDI1MjExMzEzMjk2"} +{"index":2816,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2308.13418","reference_contexts":"Although some OCR-1.0 models, e.g., Nougat [6] can directly process documents at the whole page level, they are often designed and trained for a specific sub-task, leading to unsatisfactory general ability","reference_id":"Mjg5Mzg4MTg1MTkzNTY5NzI0OA=="} +{"index":2817,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2308.13418","reference_contexts":"We hope that with GOT, researchers no longer have to worry about PDF ground truth page breaks (e.g., Nougat [6] ), as they can train on multiple pages directly","reference_id":"MTA4ODQ5MTYxODUwNDA5NjI0MzM="} +{"index":2818,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2308.13418","reference_contexts":"Using the Nougat [6] method, we obtain about 0.5M English markdown PDF-text pairs","reference_id":"NzM2Mzc5MDUwMDgwODUzNzE1NA=="} +{"index":2819,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2404.16821","reference_contexts":"InternVL-1.5 [9] and other models [27, 50] utilize a sliding window manner to crop the whole image into multiple sub-patches for high-resolution OCR","reference_id":"MTcxMTM3MTExNzIwMzA1NTI3NDA="} +{"index":2820,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2404.16821","reference_contexts":"We use the InternVL-1.5 [9] cropping method with tiles max to 12","reference_id":"MTQ3MzMwMDg4MzY3NzkwMTkyODQ="} +{"index":2821,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2404.16821","reference_contexts":"To alleviate this problem, some models [9, 27, 50] adopt a sliding window manner to decompose input images into smaller patches","reference_id":"OTgzMzQ4OTU0NDQ0NDI3OTEyNQ=="} +{"index":2822,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2404.16821","reference_contexts":"In the past year, Large Vision Language models (LVLMs) [5, 9, 24, 27, 36, 46, 49] have developed rapidly and showcased impressive performance","reference_id":"MTAwMjMzMjA0NDUxNDk1MDUzNTk="} +{"index":2823,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2404.16821","reference_contexts":"Imagine how wasteful it is to use thousands of image tokens, e.g., the image-cropping manner [9, 23] , to encode an equal amount of optical characters (e.g., texts within only an A4-PDF page)","reference_id":"NDM3MzA1MTMzMDk3OTY0MjExMA=="} +{"index":2824,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2404.16821","reference_contexts":"Large Vision-Language Models (LVLMs) [5, 9, 20, 24, 27, 46, 49] have attracted lots of attention in the AI-community due to their powerful generalization capabilities","reference_id":"Njk5ODg4MDI1MjExMzEzMjk2"} +{"index":2825,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2403.12895","reference_contexts":"English Chinese region color region color DocOwl1.5 [13] Fox [20] GOT Fox [20] GOT Fox [20] GOT Fox [20] GOT Metric Deplot (1.3B) [22] UniChart (0.26B) [31] ChartVLM (7.3B) [48] GPT-4V (>100B) [36] Qwen-VL (>72B) [ 5] GOT (0.58B) ChartQA-SE AP@strict 0.614 0.423 0.718 0.504 0.586 0.747 AP@slight 0.709 53.18 0.814 0.606 0.685 0.845 AP@high 0.729 0.560 0.842 0.643 0.727 0.867 PlotQA-SE AP@strict 0.031 0.105 0.038 0.073 0.005 0.133 AP@slight 16.49 0.260 0.468 0.194 0.042 0.596 AP@high 26.50 0.269 0.540 0.223 0.120 0.640","reference_id":"ODk5ODg1NTk5OTMxMzMzOTE0OA=="} +{"index":2826,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2301.12597","reference_contexts":"To quickly obtain the QA-gain benefits from LLMs, most LVLMs [15, 24, 49] align image tokens to text ones","reference_id":"ODg3ODY3MTMwMzM2MjM3ODM2Ng=="} +{"index":2827,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2405.14295","reference_contexts":"We use the open-source Fox [20] benchmark to test the performance of GOT on both Chinese and English PDF OCR","reference_id":"MTA5OTM3MTcwNzEwMTM0MDU0MQ=="} +{"index":2828,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2405.14295","reference_contexts":"For the document-level fine-grained OCR, following Fox [20] , we filter out those with the scanned format in the downloaded PDF files and parse the left part using Python packages (Fitz/PDFminer)","reference_id":"MjA2Mzg1MTEwNzA1MjcyOTE1MA=="} +{"index":2829,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2405.14295","reference_contexts":"As shown in Table 4 , the GOT is overall better than Fox [20] on both the bounding box-based and color-based referential OCR tasks, indicating that our model enjoys excellent interactive OCR capabilities.","reference_id":"MTc5Nzk5MDg3MDU2MjI2MTI5NzA="} +{"index":2830,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2405.14295","reference_contexts":"Specifically, fine-grained and muti-crop/page synthetic data are generated and added for GOT to support region prompt OCR [20] , huge image OCR, and batched PDF OCR features.","reference_id":"MTI5MzI0NjQyOTcxMTg1MDY3Nw=="} +{"index":2831,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2405.14295","reference_contexts":"English Chinese region color region color DocOwl1.5 [13] Fox [20] GOT Fox [20] GOT Fox [20] GOT Fox [20] GOT Metric Deplot (1.3B) [22] UniChart (0.26B) [31] ChartVLM (7.3B) [48] GPT-4V (>100B) [36] Qwen-VL (>72B) [ 5] GOT (0.58B) ChartQA-SE AP@strict 0.614 0.423 0.718 0.504 0.586 0.747 AP@slight 0.709 53.18 0.814 0.606 0.685 0.845 AP@high 0.729 0.560 0.842 0.643 0.727 0.867 PlotQA-SE AP@strict 0.031 0.105 0.038 0.073 0.005 0.133 AP@slight 16.49 0.260 0.468 0.194 0.042 0.596 AP@high 26.50 0.269 0.540 0.223 0.120 0.640","reference_id":"ODk5ODg1NTk5OTMxMzMzOTE0OA=="} +{"index":2832,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2405.14295","reference_contexts":"Large Vision-Language Models (LVLMs) [5, 9, 20, 24, 27, 46, 49] have attracted lots of attention in the AI-community due to their powerful generalization capabilities","reference_id":"Njk5ODg4MDI1MjExMzEzMjk2"} +{"index":2833,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2405.14295","reference_contexts":"As a high-interactivity feature, fine-grained OCR [20] is the region-level visual perception controlled by spatial coordinates or colors","reference_id":"MTI2OTEyMDIxMTA0MzE2Mjg2NTA="} +{"index":2834,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2212.10505","reference_contexts":"Output: [21] , and GuidedBackpropagation [22] ) to explain image captioning predictions with respect to the image content and the words of the sentence generated so far","reference_id":"MTgyMjM3Nzg0OTYxMzE4MjAyNw=="} +{"index":2835,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2212.10505","reference_contexts":"English Chinese region color region color DocOwl1.5 [13] Fox [20] GOT Fox [20] GOT Fox [20] GOT Fox [20] GOT Metric Deplot (1.3B) [22] UniChart (0.26B) [31] ChartVLM (7.3B) [48] GPT-4V (>100B) [36] Qwen-VL (>72B) [ 5] GOT (0.58B) ChartQA-SE AP@strict 0.614 0.423 0.718 0.504 0.586 0.747 AP@slight 0.709 53.18 0.814 0.606 0.685 0.845 AP@high 0.729 0.560 0.842 0.643 0.727 0.867 PlotQA-SE AP@strict 0.031 0.105 0.038 0.073 0.005 0.133 AP@slight 16.49 0.260 0.468 0.194 0.042 0.596 AP@high 26.50 0.269 0.540 0.223 0.120 0.640","reference_id":"ODk5ODg1NTk5OTMxMzMzOTE0OA=="} +{"index":2836,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2403.04473","reference_contexts":"The hallucination problem is also discussed in [27] , where the authors state that it is possibly caused by language priors or visual mis-classification, which could be partially due to the biases present in the dataset.","reference_id":"MTE3MzI4MjQ2MTM2NTA1OTU3NTY="} +{"index":2837,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2403.04473","reference_contexts":"In the past year, Large Vision Language models (LVLMs) [5, 9, 24, 27, 36, 46, 49] have developed rapidly and showcased impressive performance","reference_id":"MTAwMjMzMjA0NDUxNDk1MDUzNTk="} +{"index":2838,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2403.04473","reference_contexts":"Large Vision-Language Models (LVLMs) [5, 9, 20, 24, 27, 46, 49] have attracted lots of attention in the AI-community due to their powerful generalization capabilities","reference_id":"Njk5ODg4MDI1MjExMzEzMjk2"} +{"index":2839,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2403.04473","reference_contexts":"InternVL-1.5 [9] and other models [27, 50] utilize a sliding window manner to crop the whole image into multiple sub-patches for high-resolution OCR","reference_id":"MTcxMTM3MTExNzIwMzA1NTI3NDA="} +{"index":2840,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"1608.03983","reference_contexts":"We utilize the AdamW [29] optimizer and a cosine annealing scheduler [28] with a start learning rate of 1e-4","reference_id":"MTAwNzk2Mzg2MzY3MzA0NDMyMzA="} +{"index":2841,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"1608.03983","reference_contexts":"[28] incorporates the inductive bias of natural language with scene graphs to facilitate image captioning","reference_id":"ODQ2ODEzMTIyMjAyMTU4NTA1OA=="} +{"index":2842,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2305.14761","reference_contexts":"However, language bias is not always correct, for example, not only men ride snowboards [29] and bananas are not always yellow [30] , [31]","reference_id":"MTUyNTA1MTE4NDY3MjQxMjY3MA=="} +{"index":2843,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2305.14761","reference_contexts":"English Chinese region color region color DocOwl1.5 [13] Fox [20] GOT Fox [20] GOT Fox [20] GOT Fox [20] GOT Metric Deplot (1.3B) [22] UniChart (0.26B) [31] ChartVLM (7.3B) [48] GPT-4V (>100B) [36] Qwen-VL (>72B) [ 5] GOT (0.58B) ChartQA-SE AP@strict 0.614 0.423 0.718 0.504 0.586 0.747 AP@slight 0.709 53.18 0.814 0.606 0.685 0.845 AP@high 0.729 0.560 0.842 0.643 0.727 0.867 PlotQA-SE AP@strict 0.031 0.105 0.038 0.073 0.005 0.133 AP@slight 16.49 0.260 0.468 0.194 0.042 0.596 AP@high 26.50 0.269 0.540 0.223 0.120 0.640","reference_id":"ODk5ODg1NTk5OTMxMzMzOTE0OA=="} +{"index":2844,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2203.10244","reference_contexts":"For chart OCR, we use structure-extraction version [8] ChartQA [32] and PlotQA [35] as benchmarks","reference_id":"MTUxMjkwMjQxMjczMDE4Nzg3NTQ="} +{"index":2845,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2203.10244","reference_contexts":"We implement the LRP-IFT on top of pre-trained image captioning models trained with Flickr30K [32] and MSCOCO2017 [33] datasets and effectively improve the mean average precision (mAP) of predicted frequent object words evaluated across the test set","reference_id":"MTA5OTA4NDIyMjA0MTE4NDU3ODk="} +{"index":2846,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"1601.07140","reference_contexts":"and ShopSign [51] , and COCO-Text [44] dataset","reference_id":"MTMyMzkwNzQyNTU2NTA2MDA1NDQ="} +{"index":2847,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2312.06109","reference_contexts":"Besides, following Vary [46, 47] , we gather another 0.5M Chinese markdown pairs","reference_id":"MTg0MjkzNzg2MjkyMzI4NDEzOTI="} +{"index":2848,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2312.06109","reference_contexts":"Innovatively, Vary [46] generates a new high-resolution OCR vision vocabulary paralleling the CLIP branch to deal with document-level dense OCR","reference_id":"MTQxMzM3MjA3MDUxNzY2MjE5MDQ="} +{"index":2849,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2312.06109","reference_contexts":"We follow the Vary-tiny setting [46] to design the last two layers of the encoder, which will transfer a 1024×1024×3 input image to 256×1024 image tokens","reference_id":"MTE5NTg5NTg3MTgyNjM5NTg5NzA="} +{"index":2850,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2312.06109","reference_contexts":"In the past year, Large Vision Language models (LVLMs) [5, 9, 24, 27, 36, 46, 49] have developed rapidly and showcased impressive performance","reference_id":"MTAwMjMzMjA0NDUxNDk1MDUzNTk="} +{"index":2851,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2312.06109","reference_contexts":"Large Vision-Language Models (LVLMs) [5, 9, 20, 24, 27, 46, 49] have attracted lots of attention in the AI-community due to their powerful generalization capabilities","reference_id":"Njk5ODg4MDI1MjExMzEzMjk2"} +{"index":2852,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2401.12503","reference_contexts":"Besides, following Vary [46, 47] , we gather another 0.5M Chinese markdown pairs","reference_id":"MTg0MjkzNzg2MjkyMzI4NDEzOTI="} +{"index":2853,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2307.02499","reference_contexts":"In the past year, Large Vision Language models (LVLMs) [5, 9, 24, 27, 36, 46, 49] have developed rapidly and showcased impressive performance","reference_id":"MTAwMjMzMjA0NDUxNDk1MDUzNTk="} +{"index":2854,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2307.02499","reference_contexts":"Large Vision-Language Models (LVLMs) [5, 9, 20, 24, 27, 46, 49] have attracted lots of attention in the AI-community due to their powerful generalization capabilities","reference_id":"Njk5ODg4MDI1MjExMzEzMjk2"} +{"index":2855,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2307.02499","reference_contexts":"To quickly obtain the QA-gain benefits from LLMs, most LVLMs [15, 24, 49] align image tokens to text ones","reference_id":"ODg3ODY3MTMwMzM2MjM3ODM2Ng=="} +{"index":2856,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2310.05126","reference_contexts":"To alleviate this problem, some models [9, 27, 50] adopt a sliding window manner to decompose input images into smaller patches","reference_id":"OTgzMzQ4OTU0NDQ0NDI3OTEyNQ=="} +{"index":2857,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2310.05126","reference_contexts":"InternVL-1.5 [9] and other models [27, 50] utilize a sliding window manner to crop the whole image into multiple sub-patches for high-resolution OCR","reference_id":"MTcxMTM3MTExNzIwMzA1NTI3NDA="} +{"index":2858,"citer_arxiv_id":"2409.01704","cited_arxiv_id":"2205.01068","reference_contexts":"Then, these image tokens are projected into language model (OPT-125M [53] ) dimension via a 1024×768 linear layer","reference_id":"MTM1MzMxNjkzMjA2OTMxNjk4NDU="} +{"index":2859,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2303.08774","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2860,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2211.14305","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2861,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2312.00785","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2862,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2301.00704","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2863,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2209.14491","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2864,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2203.13131","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2865,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2208.01618","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2866,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2307.08041","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2867,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2207.12598","reference_contexts":"∅ denotes leveraging a special empty token to replace the original token to disable the additional conditional information Ho & Salimans (2022) .","reference_id":"ODYyMDU5ODMxODYzNjk4Mzk2OA=="} +{"index":2868,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2207.12598","reference_contexts":"(2021) ; Peebles & Xie (2023) to model the large-scale image distribution and incorporate additional controls with classifier-free guidance Ho & Salimans (2022)","reference_id":"MTYzNzk2MTc3MjgwNDkwNjc0Ng=="} +{"index":2869,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2207.12598","reference_contexts":"Inspired by the classifier-free guidance Ho & Salimans (2022) from diffusion models, we empirically find a similar form of guidance that can be used for autoregressive sample conditional images based on teacher forcing","reference_id":"MTc1ODIyNDcxNjk3ODExNTA1NTk="} +{"index":2870,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2302.09778","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2871,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2302.11797","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2872,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2305.14720","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2873,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2310.04378","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2874,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2108.01073","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2875,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2302.08453","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2876,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2402.09812","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2877,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2112.10741","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2878,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2102.12092","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2879,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2208.12242","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2880,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2010.02502","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2881,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2310.14189","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2882,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2011.13456","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2883,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2303.01469","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2884,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2404.02905","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2885,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2307.09288","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2886,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2312.02116","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2887,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"1908.00463","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2888,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2303.12733","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2889,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2110.04627","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2890,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2310.05737","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2891,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2403.05121","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2892,"citer_arxiv_id":"2406.09750","cited_arxiv_id":"2312.03045","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2893,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2303.08774","reference_contexts":"Unlike existing methods focusing on edited image quality, we leverage diffusion prior and vision-language model (i.e., GPT-4o [1] ) to create better-aligned instructions with original-edited image pairs, providing more accurate supervision.","reference_id":"NTk3NDEwNTIyMDEzNTIyODE0Mg=="} +{"index":2894,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2303.08774","reference_contexts":"Considering that the Real-Edit [16] benchmark is evaluated by GPT-4o [1] , and its evaluation results closely align with human ratings [16] , we choose this benchmark to conduct ablation experiments in Tab","reference_id":"NzIyMjc0MDA3MTE2NzEyOTcwMQ=="} +{"index":2895,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2303.08774","reference_contexts":"Finally, ensuring the accuracy and effectiveness of correction and contrastive instructions requires the use of GPT-4o [1] , which may incur additional costs as the amount of data increases.","reference_id":"MTA5NDQ3NDY3ODA5NDM3ODIyMzk="} +{"index":2896,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2303.08774","reference_contexts":"Specifically, MultiReward [16] uses high-resolution images from the Unsplash community as a test dataset and combines them with GPT-4o [1] to create an automated evaluation method for singleturn editing","reference_id":"MTM4NDYzODAwMjE0Mjk5ODA4MjU="} +{"index":2897,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2303.08774","reference_contexts":"To further facilitate supervision signal effectiveness, we drew on successful alignment experiences from large language models [1, 32, 39 ] and text-to-image diffusion models [7, 52, 54] : constructing positive and negative sample pairs and guiding the model to assign a higher generation probability to positive samples compared to negative ones","reference_id":"MTQxMjQ2ODQ5MzE4MjYzOTQ2MzI="} +{"index":2898,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2303.08774","reference_contexts":"To address this, we first analyze the ability of different VLMs to understand the differences between original and edited images, showing that GPT-4o [1] is the most capable of rectifying editing instructions","reference_id":"MTc2NTU2MDcxMDY1MjQ2ODcwNDY="} +{"index":2899,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2211.01324","reference_contexts":"Previous work has shown that different timesteps play distinct roles in image generation for text-to-image diffusion models, regardless of the text prompt [2, 17, 19, 36, 46, 56, 61]","reference_id":"Mzg1NzU2NDMwNjU3MDc5ODQzNQ=="} +{"index":2900,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2211.01324","reference_contexts":"stages correspond to the generation of different image attributes, independent of the input text prompt [2, 17, 19, 36, 46, 56, 61] or editing instructions","reference_id":"MTQ5MDc3ODY4MDQwOTE1MjA0ODU="} +{"index":2901,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2405.04007","reference_contexts":"3 , existing image editing datasets [4, 13, 58] typically use only Steps 1 and 2: LLMs construct editing prompts and captions, and then text-to-image diffusion models synthesize edited images","reference_id":"NTM5ODYwNDU3MDYxODQ3Mjk2"} +{"index":2902,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2405.04007","reference_contexts":"Specifically, we extracted 10,177, 8,807, and 21,016 editing pairs from InstructPix2Pix [4], MagicBrush [58], and Seed-Data-Edit [13] , respectively, resulting in a total of 40,000 training samples","reference_id":"MTcxNjc5NTUyNDA0MzMwMjgyNDU="} +{"index":2903,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2405.04007","reference_contexts":"Due to the difficulty of scaling instruction-based image editing data through manual collection, existing efforts [4, 13, 25, 58, 62] aim to automatically modify text descriptions of original images and generate edited images with T2I diffusion models","reference_id":"OTk3NTUwNDU3NDQ4MjE4MjE1Mg=="} +{"index":2904,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2405.04007","reference_contexts":"When the same editing instruction is applied during mid (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) or late (20) (21) (22) (23) (24) (25) (26) (27) (28) (29) (30) stages, the model fails to properly modify the global layout, maintaining the original forest background despite the editing instructions","reference_id":"MTcxMjkzODY2ODk0ODQwNDAwMTQ="} +{"index":2905,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2405.04007","reference_contexts":"To address the issue of scarce training data, existing efforts [13, 19, 25, 62] have attempted to develop various automated pipelines to synthesize large datasets","reference_id":"MTU0MDEzNjY5ODUwMTMyOTMxOTg="} +{"index":2906,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2207.12598","reference_contexts":"Following InstructPix2Pix's implementation [4], we enable classifier-free diffusion guidance [20] for both the image condition and the text condition with 5% mask probability during training","reference_id":"MTcxNzk0Nzg3NDMwMjQ1MzIwNjE="} +{"index":2907,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2207.12598","reference_contexts":"When the same editing instruction is applied during mid (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) or late (20) (21) (22) (23) (24) (25) (26) (27) (28) (29) (30) stages, the model fails to properly modify the global layout, maintaining the original forest background despite the editing instructions","reference_id":"MTcxMjkzODY2ODk0ODQwNDAwMTQ="} +{"index":2908,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2403.05135","reference_contexts":"When the same editing instruction is applied during mid (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) or late (20) (21) (22) (23) (24) (25) (26) (27) (28) (29) (30) stages, the model fails to properly modify the global layout, maintaining the original forest background despite the editing instructions","reference_id":"MTcxMjkzODY2ODk0ODQwNDAwMTQ="} +{"index":2909,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2403.05135","reference_contexts":"This indicates that inherent biases in pre-trained text-to-image diffusion models [15, 22] , such as difficulties in understanding quantity, position, and spatial relationships, persist in editing models","reference_id":"MTM4NzMzMzgzMTMzNTMwMDE0NTE="} +{"index":2910,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2403.05135","reference_contexts":"Additionally, inherent issues present in pre-trained text-to-image diffusion models [15, [22] [23] [24] , such as difficulty in understanding quantity, position, or object relationships, persist in the editing models","reference_id":"MTYyNzIyNDMyNjc1ODYwNDcxMjg="} +{"index":2911,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2404.09990","reference_contexts":"Due to the difficulty of scaling instruction-based image editing data through manual collection, existing efforts [4, 13, 25, 58, 62] aim to automatically modify text descriptions of original images and generate edited images with T2I diffusion models","reference_id":"OTk3NTUwNDU3NDQ4MjE4MjE1Mg=="} +{"index":2912,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2404.09990","reference_contexts":"When the same editing instruction is applied during mid (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) or late (20) (21) (22) (23) (24) (25) (26) (27) (28) (29) (30) stages, the model fails to properly modify the global layout, maintaining the original forest background despite the editing instructions","reference_id":"MTcxMjkzODY2ODk0ODQwNDAwMTQ="} +{"index":2913,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2404.09990","reference_contexts":"To address the issue of scarce training data, existing efforts [13, 19, 25, 62] have attempted to develop various automated pipelines to synthesize large datasets","reference_id":"MTU0MDEzNjY5ODUwMTMyOTMxOTg="} +{"index":2914,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2309.00267","reference_contexts":"The success of alignment training in large language models (LLMs) [26, 32, 39] has been applied to diffusion models for better image generation","reference_id":"MTEzODMwNDgwMzIxNTQ1Mjc4NjA="} +{"index":2915,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2309.00267","reference_contexts":"When the same editing instruction is applied during mid (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) or late (20) (21) (22) (23) (24) (25) (26) (27) (28) (29) (30) stages, the model fails to properly modify the global layout, maintaining the original forest background despite the editing instructions","reference_id":"MTcxMjkzODY2ODk0ODQwNDAwMTQ="} +{"index":2916,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2204.06125","reference_contexts":"In recent years, significant progress has been made in text-to-image (T2I) generation [10, 37, [40] [41] [42] due to the development of diffusion models [9, 21, 48, 49]","reference_id":"MTQ4MjQ2NDc1MDA3NDI3OTQ4ODI="} +{"index":2917,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2204.06125","reference_contexts":"Building on advancements in text-to-image (T2I) diffusion models [10, 37, [40] [41] [42] 44] , recent research has explored them for image editing [4, 19]","reference_id":"MTcxMTc2ODkwOTIzMTM4OTc4NDM="} +{"index":2918,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2304.06790","reference_contexts":"Mask-based methods [6, 18, 28, 47, 53, 57] enable fine-grained local edits with user-provided or predicted masks and corresponding text descriptions","reference_id":"MTQ1MDgwMzA3ODI0ODMyODAxODI="} +{"index":2919,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2407.05282","reference_contexts":"Due to the difficulty of scaling instruction-based image editing data through manual collection, existing efforts [4, 13, 25, 58, 62] aim to automatically modify text descriptions of original images and generate edited images with T2I diffusion models","reference_id":"OTk3NTUwNDU3NDQ4MjE4MjE1Mg=="} +{"index":2920,"citer_arxiv_id":"2505.02370","cited_arxiv_id":"2407.05282","reference_contexts":"To address the issue of scarce training data, existing efforts [13, 19, 25, 62] have attempted to develop various automated pipelines to synthesize large datasets","reference_id":"MTU0MDEzNjY5ODUwMTMyOTMxOTg="} +{"index":2921,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2309.05793","reference_contexts":"To improve the ID fidelity, aside from enhancements on the ID encoder [50, 44, 53] , another universal and parallel improvement is the introducing of an ID loss [6, 3, 29] during the training","reference_id":"MTI2Njc3ODg2NzA2NjYzNjQ3NTU="} +{"index":2922,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2309.05793","reference_contexts":"Another line of work [48, 50, 3, 42, 22, 21, 44] forgoes the necessity of fine-tuning for each ID, instead resorting to pre-training an ID adapter [13, 28] on an expansive portrait dataset","reference_id":"NjYyOTczMjA2MjQwMDMzODg1"} +{"index":2923,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2309.05793","reference_contexts":"Consequently, some methods [3] predict x 0 directly from the current timestep and then calculate the ID loss","reference_id":"OTE5Nzc1NTk0MDI2MDg4MzExNw=="} +{"index":2924,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2309.05793","reference_contexts":"(3) Experiments show that our method achieves SOTA performance in terms of both ID fidelity and editability","reference_id":"MTMyMzg2MTM2MjIwMTEyNzg4NTI="} +{"index":2925,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2309.05793","reference_contexts":"To improve ID fidelity, ID loss is employed in previous works [18, 3] , motivated by its effectiveness in prior GAN-based works [35, 45]","reference_id":"MTI2NzUwNDI5NTUzNDUyNjE3MDk="} +{"index":2926,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2304.02642","reference_contexts":"Tuning-free customization [16, 48, 50] methods typically employ an encoder to extract ID features from C id","reference_id":"NTY2NTczNjA0MDk5MDAwMTU5NQ=="} +{"index":2927,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2211.01095","reference_contexts":"During inference, such a lengthy process can be shortened to a few dozen steps with the help of advanced sampling methods [39, 25, 17]","reference_id":"ODI5OTY2Mjg4MTE5NzM3OTc5Ng=="} +{"index":2928,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2310.04378","reference_contexts":"We notice a concurrent work, LCM-Lookahead [8] , which also uses fast sampling technology (i.e., LCM [26] ) to achieve a more precise prediction of x 0","reference_id":"NjU2NTAwNzcxOTQwMjE5NDQ2NQ=="} +{"index":2929,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2310.04378","reference_contexts":"Leveraging recent fast sampling methods [26, 38, 23] , the Lighting T2I branch can generate high-quality images from pure noise with a limited and manageable number of steps","reference_id":"MTAyMzY5ODI1NzAyNDAwODg1ODE="} +{"index":2930,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2310.04378","reference_contexts":"Recent distill-based works [23, 26, 38] further accelerate this generation process within 10 steps","reference_id":"MzQyMDA1NDg5MTkwNDM4MzQ3MA=="} +{"index":2931,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2310.04378","reference_contexts":"Leveraging recent fast sampling methods [26, 38, 23] , the lighting T2I branch can generate high-quality images from pure noise with a limited and manageable number of steps","reference_id":"MTYyMzg5Nzc2MTgxNTY0NDQwNTA="} +{"index":2932,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2204.06125","reference_contexts":"The denoising model ϵ θ in modern T2I diffusion models [37, 33, 30] is predominantly a UNET composed of residual blocks [11] , self-attention layers, and cross-attention [43] layers","reference_id":"MzE0NTI0NTIxNDY2NzAyNzU4OQ=="} +{"index":2933,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2311.17042","reference_contexts":"Recent distill-based works [23, 26, 38] further accelerate this generation process within 10 steps","reference_id":"MzQyMDA1NDg5MTkwNDM4MzQ3MA=="} +{"index":2934,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2311.17042","reference_contexts":"Leveraging recent fast sampling methods [26, 38, 23] , the Lighting T2I branch can generate high-quality images from pure noise with a limited and manageable number of steps","reference_id":"MTAyMzY5ODI1NzAyNDAwODg1ODE="} +{"index":2935,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2311.17042","reference_contexts":"Secondly, to enhance prompt editing capability, LCM-Lookahead capitalized on the mode collapse phenomenon of SDXL-Turbo [38] to synthesis an ID-consistent dataset","reference_id":"Nzc1ODQxNjQ5MzY5MjYxMjE1Mw=="} +{"index":2936,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2311.17042","reference_contexts":"Leveraging recent fast sampling methods [26, 38, 23] , the lighting T2I branch can generate high-quality images from pure noise with a limited and manageable number of steps","reference_id":"MTYyMzg5Nzc2MTgxNTY0NDQwNTA="} +{"index":2937,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2303.15389","reference_contexts":"For the ID encoder, we use antelopev2 [6] as the face recognition model and EVA-CLIP [40] as the CLIP Image encoder","reference_id":"NjY0OTc0MjIyMDcyNTgzMTgwNg=="} +{"index":2938,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"In terms of minimizing the disruption, one plausible approach is to utilize a face recognition model [6] to extract more abstract and relevant facial domain-specific representations, as done by IP-Apdater-FaceID [1] and InstantID [44]","reference_id":"MTU1NjEyNTMxNDk1NTcwMTA0NDQ="} +{"index":2939,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"To improve the ID fidelity, aside from enhancements on the ID encoder [50, 44, 53] , another universal and parallel improvement is the introducing of an ID loss [6, 3, 29] during the training","reference_id":"MTI2Njc3ODg2NzA2NjYzNjQ3NTU="} +{"index":2940,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"We observe that the comparison methods, namely InstantID [44] and IPAdapter (more specifically, IPAdapter-FaceID [1]) are highly compatible with the SDXL-Lightning model","reference_id":"MzcxNjE4Mjc2MTU2MDQ3OTIxMQ=="} +{"index":2941,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"1, the ID condition C id is usually cropped from the target image x 0 [50, 44]","reference_id":"NTQyMzQ3MjU0ODc0NzcwMzk0NA=="} +{"index":2942,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"While some research [50, 44, 22] has shown the ability for stylized ID generation, notable style degradation occurs when compared with images before ID insertion (as depicted in Fig","reference_id":"OTk1NTU3NTQwNTY1MzA2ODQ3NA=="} +{"index":2943,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"Another line of work [48, 50, 3, 42, 22, 21, 44] forgoes the necessity of fine-tuning for each ID, instead resorting to pre-training an ID adapter [13, 28] on an expansive portrait dataset","reference_id":"NjYyOTczMjA2MjQwMDMzODg1"} +{"index":2944,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"InstantID [44] builds on this by including an additional ID&Landmark ControlNet [52] for more effective modulation","reference_id":"NzY1MjU0MjAxNTI1OTM1MDE4Mg=="} +{"index":2945,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2401.07519","reference_contexts":"To ease the resource demand necessitated by tuning-based methods [7, 36, 14, 19, 10, 41] , a series of tuning-free methods [42, 44, 29, 50, 22, 48, 51, 4] have emerged, which directly encode ID information into the generation process","reference_id":"NTU2MTg0NTM5NTg3ODc5MTE3OQ=="} +{"index":2946,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2405.00760","reference_contexts":"Nonetheless, this issue is not uniquely associated to our method, as similar phenomena has been observed in reward tuning methods [49, 31, 5, 47] , with [5] attributing it to the reward-hacking problem","reference_id":"MTUyNjk0Nzg0ODIyOTI3Mzk5MzQ="} +{"index":2947,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"To improve the ID fidelity, aside from enhancements on the ID encoder [50, 44, 53] , another universal and parallel improvement is the introducing of an ID loss [6, 3, 29] during the training","reference_id":"MTI2Njc3ODg2NzA2NjYzNjQ3NTU="} +{"index":2948,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"Tuning-free customization [16, 48, 50] methods typically employ an encoder to extract ID features from C id","reference_id":"NTY2NTczNjA0MDk5MDAwMTU5NQ=="} +{"index":2949,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"1, the ID condition C id is usually cropped from the target image x 0 [50, 44]","reference_id":"NTQyMzQ3MjU0ODc0NzcwMzk0NA=="} +{"index":2950,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"While some research [50, 44, 22] has shown the ability for stylized ID generation, notable style degradation occurs when compared with images before ID insertion (as depicted in Fig","reference_id":"OTk1NTU3NTQwNTY1MzA2ODQ3NA=="} +{"index":2951,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"Another line of work [48, 50, 3, 42, 22, 21, 44] forgoes the necessity of fine-tuning for each ID, instead resorting to pre-training an ID adapter [13, 28] on an expansive portrait dataset","reference_id":"NjYyOTczMjA2MjQwMDMzODg1"} +{"index":2952,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"IPAdapter [50, 1] shifted from early-version CLIP extraction of grid features to utilizing face recognition backbone [6] to extract more abstract and relevant ID information","reference_id":"Njg3MTAzODI0ODE4MDAyNDY1Ng=="} +{"index":2953,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"To ease the resource demand necessitated by tuning-based methods [7, 36, 14, 19, 10, 41] , a series of tuning-free methods [42, 44, 29, 50, 22, 48, 51, 4] have emerged, which directly encode ID information into the generation process","reference_id":"NTU2MTg0NTM5NTg3ODc5MTE3OQ=="} +{"index":2954,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"This technique, proposed by IP-Adapter [50] , has been widely used, we also adopt it for embedding ID features in this study.","reference_id":"MTE0NTc2MTcxMjYyNTAzMzM2MDM="} +{"index":2955,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2308.06721","reference_contexts":"As a special category of customized text-to-image (T2I) generation [7, 36, 14, 19, 46, 50] , identity (ID) customization allow users to adapt pre-trained T2I diffusion models to align with their personalized ID","reference_id":"ODA1MjE5MjE4MDI2MDUzMTA0OA=="} +{"index":2956,"citer_arxiv_id":"2404.16022","cited_arxiv_id":"2403.17008","reference_contexts":"To improve the ID fidelity, aside from enhancements on the ID encoder [50, 44, 53] , another universal and parallel improvement is the introducing of an ID loss [6, 3, 29] during the training","reference_id":"MTI2Njc3ODg2NzA2NjYzNjQ3NTU="} +{"index":2957,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2401.12945","reference_contexts":"Iterative denoising methods such as diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching (Lipman et al., 2022; Liu et al., 2022) have seen remarkable success in modelling diverse images (Rombach et al., 2022; Esser et al., 2024) , video (Ho et al., 2022; Bar-Tal et al., 2024) , audio (Kong et al., 2020) , and proteins (Abramson et al., 2024 )","reference_id":"MTI0MTM4Njg3NTI3NTA5MzI1NzY="} +{"index":2958,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2303.04248","reference_contexts":"Our work most closely relates to techniques using binary time-distillation (Salimans & Ho, 2022; Meng et al., 2023; Berthelot et al., 2023) , which divides distillation into log 2 (T ) stages of increasing step-size, shortening the required bootstrap paths","reference_id":"MTQ5OTYyNzc2Mzc4ODU3MTA4NzQ="} +{"index":2959,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"1809.11096","reference_contexts":"A APPENDIX FID Sampling Steps Param Count Epochs Trained Adversarial approaches BigGAN-deep (Brock, 2018) 6.96 1 --StyleGAN-XL (Sauer et al., 2022) 2.3 1 --Denoising approaches DiT-XL (Peebles & Xie, 2023) 2.27 500 675M 640 ADM-G (Dhariwal & Nichol, 2021) Table 3: Hyperparameters used during training","reference_id":"MzMwNjc0OTE3ODM1MjUyNzA3MA=="} +{"index":2960,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2303.04137","reference_contexts":"We compare to iterative denoising methods IBC (Florence et al., 2022) and diffusion policy (Chi et al., 2023) , as well as one-step methods LSTM-GMM (Mandlekar et al., 2021) and BET (Shafiullah et al., 2022)","reference_id":"NjczMTU4MzE2MjY3ODQ1ODMwOQ=="} +{"index":2961,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2207.12598","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2962,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"1312.6114","reference_contexts":"In other generative model flavors (e.g., GANs (Goodfellow et al., 2020) , VAEs (Kingma, 2013) ) it is possible to adjust this mapping, which has the potential to simply the learning problem","reference_id":"MTM2MzUwMjAwOTU2ODc2MzEzNDk="} +{"index":2963,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2009.09761","reference_contexts":"Iterative denoising methods such as diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching (Lipman et al., 2022; Liu et al., 2022) have seen remarkable success in modelling diverse images (Rombach et al., 2022; Esser et al., 2024) , video (Ho et al., 2022; Bar-Tal et al., 2024) , audio (Kong et al., 2020) , and proteins (Abramson et al., 2024 )","reference_id":"MTI0MTM4Njg3NTI3NTA5MzI1NzY="} +{"index":2964,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2210.02747","reference_contexts":"A recent family of models, including diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching foot_0 (Lipman et al., 2022; Liu et al., 2022) models, approach the generative modelling problem by learning an ordinary differential equation (ODE) that transforms noise into data","reference_id":"MTUzODIyMjU5Njc4NDI3NDE2MTc="} +{"index":2965,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2210.02747","reference_contexts":"Iterative denoising methods such as diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching (Lipman et al., 2022; Liu et al., 2022) have seen remarkable success in modelling diverse images (Rombach et al., 2022; Esser et al., 2024) , video (Ho et al., 2022; Bar-Tal et al., 2024) , audio (Kong et al., 2020) , and proteins (Abramson et al., 2024 )","reference_id":"MTI0MTM4Njg3NTI3NTA5MzI1NzY="} +{"index":2966,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2209.03003","reference_contexts":"A recent family of models, including diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching foot_0 (Lipman et al., 2022; Liu et al., 2022) models, approach the generative modelling problem by learning an ordinary differential equation (ODE) that transforms noise into data","reference_id":"MTUzODIyMjU5Njc4NDI3NDE2MTc="} +{"index":2967,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2209.03003","reference_contexts":"A naive way to compute targets for training s θ (x t , t, d) would be to fully simulate the ODE forward with a small enough step size (Luhman & Luhman, 2021; Liu et al., 2022)","reference_id":"MzM0MTA5NjM1ODkxNzQyMjA5MQ=="} +{"index":2968,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2209.03003","reference_contexts":"Iterative denoising methods such as diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching (Lipman et al., 2022; Liu et al., 2022) have seen remarkable success in modelling diverse images (Rombach et al., 2022; Esser et al., 2024) , video (Ho et al., 2022; Bar-Tal et al., 2024) , audio (Kong et al., 2020) , and proteins (Abramson et al., 2024 )","reference_id":"MTI0MTM4Njg3NTI3NTA5MzI1NzY="} +{"index":2969,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2209.03003","reference_contexts":"In this work, we adopt the optimal transport flow-matching objective (Liu et al., 2022) for simplicity","reference_id":"MTcxNjUwODczMjY4MjMyNjk1NDI="} +{"index":2970,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2209.03003","reference_contexts":"Knowledge distillation (Luhman & Luhman, 2021) and rectified flows (Liu et al., 2022) generate a synthetic dataset by fully simulating the denoising ODE","reference_id":"NTI4NzIzNjQxNDgwMzg2NDUzMw=="} +{"index":2971,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2209.03003","reference_contexts":"In Table 1 , shortcut models display slightly better many-step generation than a base flow model-is there a way for onestep generation to improve many-step generation (rather than typically the opposite)? Extensions to the shortcut model formulation may be possible that iteratively adjust the noise-to-data mapping (e.g., with a Reflow-like procedure (Liu et al., 2022) ), or which reduce the gap between manystep and one-step generation","reference_id":"MTg2ODUzNzY3OTMwNzA2MTk0OA=="} +{"index":2972,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2209.03003","reference_contexts":"These stages introduce complexity and require either generating a large synthetic dataset (Luhman & Luhman, 2021; Liu et al., 2022) or propagating through a series of teacher and student networks (Ho et al., 2020; Meng et al., 2023)","reference_id":"MjI1MjU0MDYwNzU4ODU0NDg4Ng=="} +{"index":2973,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"1711.05101","reference_contexts":"For all runs, we utilize the AdamW optimizer (Loshchilov, 2017) with a constant learning rate of 1 * 10 -4 , and weight decay of 0.1","reference_id":"MjM0NzE2NDIyNjkyODI3NTg2Mw=="} +{"index":2974,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2101.02388","reference_contexts":"These stages introduce complexity and require either generating a large synthetic dataset (Luhman & Luhman, 2021; Liu et al., 2022) or propagating through a series of teacher and student networks (Ho et al., 2020; Meng et al., 2023)","reference_id":"MjI1MjU0MDYwNzU4ODU0NDg4Ng=="} +{"index":2975,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2101.02388","reference_contexts":"Knowledge distillation (Luhman & Luhman, 2021) and rectified flows (Liu et al., 2022) generate a synthetic dataset by fully simulating the denoising ODE","reference_id":"NTI4NzIzNjQxNDgwMzg2NDUzMw=="} +{"index":2976,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2101.02388","reference_contexts":"Reflow is comparable to knowledge distillation (Luhman & Luhman, 2021) , except the student model is trained among all t ∈ (0, 1) rather at just t = 0","reference_id":"NjU4NjQxNDk3NTg3NDE5NzgwMQ=="} +{"index":2977,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2101.02388","reference_contexts":"A naive way to compute targets for training s θ (x t , t, d) would be to fully simulate the ODE forward with a small enough step size (Luhman & Luhman, 2021; Liu et al., 2022)","reference_id":"MzM0MTA5NjM1ODkxNzQyMjA5MQ=="} +{"index":2978,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2310.04378","reference_contexts":"While such models can be used as a student for distillation (Luo et al., 2023; Geng et al., 2024) , consistency training has also been proposed to train consistency models end-to-end from scratch (Song et al., 2023; Song & Dhariwal, 2023) , and consistency models have also been studied in the flow-matching setting (Boffi et al., 2024)","reference_id":"NDI4OTUzODAwNDk0MzkxOTEwMw=="} +{"index":2979,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2202.00512","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2980,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2311.17042","reference_contexts":"Alternatives to L2 distance have been proposed for distillation targets, such as adversarial (Sauer et al., 2023) or distribution-matching (Yin et al., 2024b;a) objectives","reference_id":"MjgyMjg4MjkwODE1NDgwMTE2MQ=="} +{"index":2981,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2010.02502","reference_contexts":"A recent family of models, including diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching foot_0 (Lipman et al., 2022; Liu et al., 2022) models, approach the generative modelling problem by learning an ordinary differential equation (ODE) that transforms noise into data","reference_id":"MTUzODIyMjU5Njc4NDI3NDE2MTc="} +{"index":2982,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2010.02502","reference_contexts":"Iterative denoising methods such as diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching (Lipman et al., 2022; Liu et al., 2022) have seen remarkable success in modelling diverse images (Rombach et al., 2022; Esser et al., 2024) , video (Ho et al., 2022; Bar-Tal et al., 2024) , audio (Kong et al., 2020) , and proteins (Abramson et al., 2024 )","reference_id":"MTI0MTM4Njg3NTI3NTA5MzI1NzY="} +{"index":2983,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2310.14189","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2984,"citer_arxiv_id":"2410.12557","cited_arxiv_id":"2303.01469","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2985,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1607.06450","reference_contexts":"(2) In the second stage (Fig","reference_id":"NzM5NTA1NDkyNjAyNjE5MTcwMg=="} +{"index":2986,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1607.06450","reference_contexts":"From level-1 to level-4, the attention heads in DMTA are [1, 2, 4, 8] , and the number of channels is [48, 96, 192, 384]","reference_id":"MTMyODE0ODMzMTE1Mzk0NDg5NTQ="} +{"index":2987,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1607.06450","reference_contexts":"where ⊙ indicates element-wise multiplication, Norm denotes layer normalization [2] , W l represents linear layer, F and F ′ ∈ R Ĥ× Ŵ × Ĉ are input and output feature maps respectively, and","reference_id":"MTEzNDEzODk0MTE4NjU1NzM2Mg=="} +{"index":2988,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1704.00028","reference_contexts":"Up to now, researchers have carried out CNN's study with different perspectives and obtained more elaborate network architecture designs and learning schemes, such as residual block [29, 81, 6] , GAN [21, 65, 48] , attention [86, 66, 11, 72, 71, 68, 73] , knowledge distillation [67] , and others [26, 19, 30, 18, 76] .","reference_id":"MTgzNjA1NjkzNzc1MTgxNTMwMTA="} +{"index":2989,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1710.10196","reference_contexts":"Specifically, we train our DiffIR with the batch size of 30 and patch size of 256 on Places-Standard [88] and CelebA-HQ [27] datasets, respectively","reference_id":"NDc4OTE0MjQwNzczODY3NTg3Ng=="} +{"index":2990,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1710.10196","reference_contexts":"In testing, we use a fixed set of 2k validation and 30k testing samples from CelebA-HQ [27] and Places [88]","reference_id":"MTM3OTMzNzgyMDEzMzQ3MzA2MTQ="} +{"index":2991,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"2201.11793","reference_contexts":"Recently, DM has become increasingly influential in the field of image restoration tasks, such as superresolution [28, 52] and inpainting [40, 50, 10]","reference_id":"NTk3OTIyMjMyMDEyNDgyMDMwMA=="} +{"index":2992,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1411.1784","reference_contexts":"Besides, Palette [51] is inspired by conditional generation models [44] and proposes a conditional diffusion model for IR","reference_id":"NzMyMjU3MjY4MzA4NTE3NDA4Ng=="} +{"index":2993,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"1901.00212","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2994,"citer_arxiv_id":"2303.09472","cited_arxiv_id":"2010.04159","reference_contexts":"Presently, the transformer has been adopted in numerous vision tasks, such as image recognition [17, 60] , segmentation [62, 69, 87, 49] , object detection [5, 89] , and image restoration [7, 38, 74, 36, 8] .","reference_id":"OTY1MDU2MTg0NDcxMzY4MDMyOQ=="} +{"index":2995,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2309.16609","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2996,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2308.12966","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2997,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2311.12793","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2998,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2305.18565","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":2999,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2406.07476","reference_contexts":"The dominant paradigm in video foundation model construction relies on contrastive (Tong et al., 2022; Feichtenhofer et al., 2022; Wang et al., 2024c) or generative learning (Cheng et al., 2024; Zhang et al., 2024c) from extensive video-text pair datasets","reference_id":"MTY2NjM1MDI4MTQ1OTMxMTgzOTQ="} +{"index":3000,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2406.07476","reference_contexts":"Due to the absence of powerful video encoders in the community, existing mainstream video MLLMs (Zhang et al., 2024c; 2023; Cheng et al., 2024) still rely on established images encoder, i.e., CLIP (Radford et al., 2021) , to extract visual information frame by frame","reference_id":"MTQ4MzcyMDk5MjUwNDIzMjEzMDI="} +{"index":3001,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2406.07476","reference_contexts":"Subsequently, they integrate temporal modeling techniques, e.g., Q-former (Zhang et al., 2023) , 3D Conv (Cheng et al., 2024) , and Pooling (Zhang et al., 2024c; Xu et al., 2024) , to compress the visual tokens before feeding them with language tokens into LLMs.","reference_id":"MTYzODA2MzA0NDkwMDgzNzE4Mw=="} +{"index":3002,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2401.16420","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3003,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2406.14515","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3004,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2306.13394","reference_contexts":"For image understanding, we reported scores from popular benchmarks like MM-Vet (Yu et al., 2023) , MMBench (Liu et al., 2023b) , MMMU (Yue et al., 2024) , MME (Fu et al., 2023) , LLaVA-wild (Liu et al., 2024b) , SEED (Ge et al., 2023b) , AI2D (Kembhavi et al., 2016), and RealWorldQA (xAI, 2024)","reference_id":"MTA3NzIzMzkwMzIzNTQ4NDExMzE="} +{"index":3005,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2405.21075","reference_contexts":"Although our proposed UTR significantly mitigates temporal hacking from both data modeling and task modeling perspectives, it still has limitations in some situation, and we identify several representative examples on the VideoMME (Fu et al., 2024) benchmark","reference_id":"MTcwNzU3NjUzMzc0ODY3MDIzODQ="} +{"index":3006,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2405.21075","reference_contexts":"For video understanding, we focused on three general benchmarks: MVBench (Li et al., 2024c) , TempCompass (Liu et al., 2024c) , and VideoMME (Fu et al., 2024) , as well as four video QA benchmarks: MVSD-QA (Xu et al., 2017) , MSRVTT-QA (Xu et al., 2016) , TGIF-QA (Jang et al., 2017) , and ActivityNet-QA (Caba Heilbron et al., 2015)","reference_id":"MTI2MDE2NjgzNDA3NzM0NjM2NA=="} +{"index":3007,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2310.01218","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3008,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2310.01218","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3009,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"1711.09883","reference_contexts":"Reward hacking theory was firstly introduced in the field of RL as a special case of Goodhart's Law (Goodhart, 1984; Leike et al., 2018) , and later explored in the context of AI alignment (Leike et al., 2017)","reference_id":"MTUyMDEzNjY0MzMxODgxODMyMzA="} +{"index":3010,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"1811.07871","reference_contexts":"Reward hacking theory was firstly introduced in the field of RL as a special case of Goodhart's Law (Goodhart, 1984; Leike et al., 2018) , and later explored in the context of AI alignment (Leike et al., 2017)","reference_id":"MTUyMDEzNjY0MzMxODgxODMyMzA="} +{"index":3011,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2408.03326","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3012,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2403.11027","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3013,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2305.06355","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3014,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2311.17043","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3015,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2311.10122","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3016,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2303.05499","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3017,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2307.06281","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3018,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2403.00476","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3019,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2306.05424","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3020,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2201.03544","reference_contexts":"Subsequent works (Pan et al., 2022; Laidlaw et al., 2024) try to deal with reward hacking from different aspects","reference_id":"NTI3NTk1NDM3NTMxNDE0OTAxMA=="} +{"index":3021,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2406.16855","reference_contexts":"Exploring whether a single multimodal large language model (Yu et al., 2024; Dong et al., 2024a; Peng et al., 2024) can be utilized to handle the entire data processing and task construction pipeline is a highly promising research direction","reference_id":"MTE3NDgzMDYzMTY5MTUxMjA3NDU="} +{"index":3022,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"1707.06347","reference_contexts":"Leveraging reinforcement learning algorithms, e.g., DPO (Rafailov et al., 2024; Zhu et al., 2025) , PPO (Schulman et al., 2017) and GRPO (DeepSeek-AI et al., 2025) , for posttraining enhancement of the model's video comprehension and reasoning capabilities constitutes a key focus for future research.","reference_id":"Njg4OTM5ODA2MjI4NTAzNzk5NA=="} +{"index":3023,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2312.11805","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3024,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2407.00634","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3025,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2307.06942","reference_contexts":"Advanced techniques such as InternVID (Wang et al., 2023) and COSMO (Wang et al., 2024a) ameliorate information density to some extent through video interleave formats, yet they still struggle with the high information density of frames and fail to effectively model spatiotemporal dynamics, thus not fully addressing Principle II","reference_id":"MTA1MTU4NjE5NDAwMTE4ODU3ODM="} +{"index":3026,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2307.06942","reference_contexts":"Previous methods (Wang et al., 2023; 2024a) modeled relatively dense information by interleaving text with selected frames, yet they neglected the critical spatiotemporal dynamics","reference_id":"MTI3OTA3OTAwODIyNzQ5MDI0MjU="} +{"index":3027,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2307.06942","reference_contexts":"However, in practice, the surrogate objective rewards consistency between model predictions and human-annotated captions (Wang et al., 2024b) or curated internet content (Bain et al., 2021; Wang et al., 2023)","reference_id":"MTIzNzk1MTk5MzE1NzE4NTI5NTA="} +{"index":3028,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2307.06942","reference_contexts":"Some works try to design filtering mechanisms (Wang et al., 2024b) to obtain high-quality video data with finegrained description, while others aim to construct appropriate data structures (Wang et al., 2023; 2024a; Chen et al., 2023b) and task formats (Yu et al., 2024) to enhance modeling performance.","reference_id":"MTM0ODA4MTc1NTkyMjQ1MTM4MTM="} +{"index":3029,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2403.15377","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3030,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2401.12503","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3031,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2212.00280","reference_contexts":"Function F extracts this information, using labeled data or specialized models such as GRiT (Wu et al., 2022) and Grounding DINO (Liu et al., 2023a)","reference_id":"Mjg4MzM0MTcyNTQyNzIwNTMyNA=="} +{"index":3032,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2212.00280","reference_contexts":"This can be accomplished through expert models, e.g., GRiT (Wu et al., 2022) , Grounding DINO (Liu et al., 2023a) , and ByteTrack (Zhang et al., 2022) or directly obtained from the annotations provided by datasets.","reference_id":"MTc4NDQyNTQwNjU3MjM3NDQwNjE="} +{"index":3033,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2212.00280","reference_contexts":"To extract subject attributions from each video frame, we use the region-to-text detector GRiT (Wu et al., 2022)","reference_id":"Mjc2NjE3MzQyNzk3MDA3Nzg2OQ=="} +{"index":3034,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2212.00280","reference_contexts":"To select specific expert models, we conducted a systematic evaluation based on existing benchmarks, e.g., COCO (Lin et al., 2014) , Lvis (Gupta et al., 2019) , VG (Krishna et al., 2017) , etc., of the performance of various options, such as GRiT (Wu et al., 2022) and GroundingDINO (Liu et al., 2023a) , to identify the most suitable candidates","reference_id":"MTY3MDU2MDgxMDQ3MTk4MjgxMw=="} +{"index":3035,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2404.16994","reference_contexts":"Subsequently, they integrate temporal modeling techniques, e.g., Q-former (Zhang et al., 2023) , 3D Conv (Cheng et al., 2024) , and Pooling (Zhang et al., 2024c; Xu et al., 2024) , to compress the visual tokens before feeding them with language tokens into LLMs.","reference_id":"MTYzODA2MzA0NDkwMDgzNzE4Mw=="} +{"index":3036,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2404.16994","reference_contexts":"However, as shown in Figure 1 and previous works (Xu et al., 2024; Wang et al., 2024b) , π * often fails to generate text that accurately aligns with video content and user instructions","reference_id":"MTU4NjMxODc0OTA5NTY5ODczODI="} +{"index":3037,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2404.16994","reference_contexts":"However, recent studies have unveiled a counterintuitive \"anti-scaling law\" phenomenon (Xu et al., 2024)","reference_id":"NzM0MjcyMDIzNzg1NzI3MzgzOQ=="} +{"index":3038,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2404.16994","reference_contexts":"In the inference and evaluation phase, we essentially follow the settings of PLLaVA (Xu et al., 2024) and LLaVA-NeXT-Video (Zhang et al., 2024c ), including the system prompt for inference, the number of frames extracted, and so on, while conducting evaluations on the existing video benchmark","reference_id":"OTU0MzEyMzE5ODQyNzcyNzMxNA=="} +{"index":3039,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2407.10671","reference_contexts":"For the specific Video-UTR pipeline, we follow the general architecture in LLaVA-NeXT-Video (Zhang et al., 2024c) , which consists of a vision encoder, SigLIP-L (Zhai et al., 2023) , a large language model, QWen-2 (Yang et al., 2024) , and a modality alignment projector, 2-layer GeLU-MLP","reference_id":"MTYxOTM2MjAwNDY1NjEzMzQ4MDM="} +{"index":3040,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2308.02490","reference_contexts":"For image understanding, we reported scores from popular benchmarks like MM-Vet (Yu et al., 2023) , MMBench (Liu et al., 2023b) , MMMU (Yue et al., 2024) , MME (Fu et al., 2023) , LLaVA-wild (Liu et al., 2024b) , SEED (Ge et al., 2023b) , AI2D (Kembhavi et al., 2016), and RealWorldQA (xAI, 2024)","reference_id":"MTA3NzIzMzkwMzIzNTQ4NDExMzE="} +{"index":3041,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2306.02858","reference_contexts":"Autoregressive video-language modeling (Li et al., 2023a; Zhang et al., 2023; 2024c) , aims to replicate human video comprehension","reference_id":"OTk3NTkyMzQ3NjA0NjM4OTIxOA=="} +{"index":3042,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2306.02858","reference_contexts":"Subsequently, they integrate temporal modeling techniques, e.g., Q-former (Zhang et al., 2023) , 3D Conv (Cheng et al., 2024) , and Pooling (Zhang et al., 2024c; Xu et al., 2024) , to compress the visual tokens before feeding them with language tokens into LLMs.","reference_id":"MTYzODA2MzA0NDkwMDgzNzE4Mw=="} +{"index":3043,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2407.12772","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3044,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2309.07915","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3045,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2307.09474","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3046,"citer_arxiv_id":"2502.12081","cited_arxiv_id":"2502.04371","reference_contexts":"Leveraging reinforcement learning algorithms, e.g., DPO (Rafailov et al., 2024; Zhu et al., 2025) , PPO (Schulman et al., 2017) and GRPO (DeepSeek-AI et al., 2025) , for posttraining enhancement of the model's video comprehension and reasoning capabilities constitutes a key focus for future research.","reference_id":"Njg4OTM5ODA2MjI4NTAzNzk5NA=="} +{"index":3047,"citer_arxiv_id":"1910.04091","cited_arxiv_id":"1705.10743","reference_contexts":"It is well known that the Wasserstein distance suffers from biased gradients [Bellemare et al., 2017]","reference_id":"MTExNzg5NTQxMTUyMjI5MjQ5MjQ="} +{"index":3048,"citer_arxiv_id":"1910.04091","cited_arxiv_id":"1705.10743","reference_contexts":"However it is well known that the empiricial Wasserstein distance is a biased estimator of the Wasserstein distance over the true distributions and leads to biased gradients as discussed in [Bellemare et al., 2017] , hence SGD might fail","reference_id":"MTMxOTIxNTA4Nzg3NzY4MTYxNDg="} +{"index":3049,"citer_arxiv_id":"2104.08691","cited_arxiv_id":"2103.10385","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3050,"citer_arxiv_id":"2104.08691","cited_arxiv_id":"1808.09121","reference_contexts":"2020)over ReLU (Nair and Hinton, 2010) activations.2 The tasks are BoolQ (Clark et al., 2019) ,CB (De Marneff et al., 2019), COPA (Roemmele et al., 2011) , MultiRC (Khashabi et al., 2018) , ReCoRD (Zhang et al., 2018) , RTE (Dagan et al., 2005; Bar-Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009), WiC (Pilehvar and Camacho- Collados, 2018) , and WSC (Levesque et al., 2012) .","reference_id":"MTQ4NzI5MjQzNzcwMjk5MDExNTg="} +{"index":3051,"citer_arxiv_id":"2104.08691","cited_arxiv_id":"2002.05202","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3052,"citer_arxiv_id":"2104.08691","cited_arxiv_id":"1810.12885","reference_contexts":"2020)over ReLU (Nair and Hinton, 2010) activations.2 The tasks are BoolQ (Clark et al., 2019) ,CB (De Marneff et al., 2019), COPA (Roemmele et al., 2011) , MultiRC (Khashabi et al., 2018) , ReCoRD (Zhang et al., 2018) , RTE (Dagan et al., 2005; Bar-Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009), WiC (Pilehvar and Camacho- Collados, 2018) , and WSC (Levesque et al., 2012) .","reference_id":"MTQ4NzI5MjQzNzcwMjk5MDExNTg="} +{"index":3053,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2108.07732","reference_contexts":"The datasets for evaluation include MMLU (Hendrycks et al., 2021a) (5-shot), MMLU-Pro (Wang et al., 2024) (5-shot), GPQA (Rein et al., 2023 ) (5shot), Theorem QA (Chen et al., 2023a) (5-shot), BBH (Suzgun et al., 2023) (3-shot), HellaSwag (Zellers et al., 2019) (10-shot), Winogrande (Sakaguchi et al., 2021) (5-shot), TruthfulQA (Lin et al., 2022a ) (0-shot), ARC-C (Clark et al., 2018) (25-shot), HumanEval (Chen et al., 2021 ) (0-shot), MBPP (Austin et al., 2021 ) (0-shot), EvalPlus (Liu et al., 2023a ) (0-shot), MultiPL-E (Cassano et al., 2023 ) (0-shot on Python, C++, Java, PHP, Type-Script, C#, Bash, and JavaScript), GSM8K (Cobbe et al., 2021) (5-shot), MATH (Hendrycks et al., 2021b ) (4-shot), C-Eval (Huang et al., 2023) (5-shot), and CMMLU (Li et al., 2023) (5-shot)","reference_id":"NzgwOTEzOTE5NDk1NDYwMTM5MQ=="} +{"index":3054,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2309.16609","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3055,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2308.12966","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3056,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2212.08073","reference_contexts":"Constitutional Feedback Constitutional AI refers to the process of guiding LLMs to generate responses based on predefined sets of principles (Bai et al., 2022)","reference_id":"Mjk4MDgzNjkzMzkwMjM4ODQ5Nw=="} +{"index":3057,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2107.03374","reference_contexts":"The datasets for evaluation include MMLU (Hendrycks et al., 2021a) (5-shot), MMLU-Pro (Wang et al., 2024) (5-shot), GPQA (Rein et al., 2023 ) (5shot), Theorem QA (Chen et al., 2023a) (5-shot), BBH (Suzgun et al., 2023) (3-shot), HellaSwag (Zellers et al., 2019) (10-shot), Winogrande (Sakaguchi et al., 2021) (5-shot), TruthfulQA (Lin et al., 2022a ) (0-shot), ARC-C (Clark et al., 2018) (25-shot), HumanEval (Chen et al., 2021 ) (0-shot), MBPP (Austin et al., 2021 ) (0-shot), EvalPlus (Liu et al., 2023a ) (0-shot), MultiPL-E (Cassano et al., 2023 ) (0-shot on Python, C++, Java, PHP, Type-Script, C#, Bash, and JavaScript), GSM8K (Cobbe et al., 2021) (5-shot), MATH (Hendrycks et al., 2021b ) (4-shot), C-Eval (Huang et al., 2023) (5-shot), and CMMLU (Li et al., 2023) (5-shot)","reference_id":"NzgwOTEzOTE5NDk1NDYwMTM5MQ=="} +{"index":3058,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2107.03374","reference_contexts":"Meanwhile, Qwen2-72B, the base language model, achieves 84.2 on MMLU (Hendrycks et al., 2021a) , 37.9 on GPQA (Rein et al., 2023) , 64.6 on HumanEval (Chen et al., 2021) , 89.5 on GSM8K (Cobbe et al., 2021) , and 82.4 on BBH (Suzgun et al., 2023) .","reference_id":"MTE1ODQ3OTYxODAxNjMxNzI2MjI="} +{"index":3059,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.04132","reference_contexts":"Our focus is on comparing this model with other leading 7B models, including Llama-3-8B, which has recently demonstrated exceptional performance in the Chatbot Arena (Chiang et al., 2024)","reference_id":"NDg4MTc3NDg0MDgzOTkwNjc0OA=="} +{"index":3060,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.04132","reference_contexts":"Recently, Claude-3 Opus (Anthropic, 2024) and (OpenAI, 2024) , the updated model for ChatGPT, have ascended to the pinnacle of the Chatbot Arena (Chiang et al., 2024) in quick succession","reference_id":"OTUyMDg0OTE4NzAxMjM5MjAzNw=="} +{"index":3061,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.04132","reference_contexts":"Qwen2 outperforms competing models in evaluations of both fundamental language capabilities and instruction-tuned functionalities Specifically, Qwen2-72B-Instruct, our instruction-tuned variant, scores 9.1 on MT-Bench (Zheng et al., 2023) , 48.1 on Arena-Hard (Chiang et al., 2024) , and 35.7 on LiveCodeBench (Jain et al., 2024)","reference_id":"NDk2MDYxNDQwMDgxNDE4MzI0Ng=="} +{"index":3062,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2311.07919","reference_contexts":"In the meantime, we have unveiled the vision-language model Qwen-VL (Bai et al., 2023b) , and launched the audio-language model Qwen-Audio (Chu et al., 2023)","reference_id":"OTc0ODc4NjM5OTUwMDc1ODM3MA=="} +{"index":3063,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"1803.05457","reference_contexts":"The datasets for evaluation include MMLU (Hendrycks et al., 2021a) (5-shot), MMLU-Pro (Wang et al., 2024) (5-shot), GPQA (Rein et al., 2023 ) (5shot), Theorem QA (Chen et al., 2023a) (5-shot), BBH (Suzgun et al., 2023) (3-shot), HellaSwag (Zellers et al., 2019) (10-shot), Winogrande (Sakaguchi et al., 2021) (5-shot), TruthfulQA (Lin et al., 2022a ) (0-shot), ARC-C (Clark et al., 2018) (25-shot), HumanEval (Chen et al., 2021 ) (0-shot), MBPP (Austin et al., 2021 ) (0-shot), EvalPlus (Liu et al., 2023a ) (0-shot), MultiPL-E (Cassano et al., 2023 ) (0-shot on Python, C++, Java, PHP, Type-Script, C#, Bash, and JavaScript), GSM8K (Cobbe et al., 2021) (5-shot), MATH (Hendrycks et al., 2021b ) (4-shot), C-Eval (Huang et al., 2023) (5-shot), and CMMLU (Li et al., 2023) (5-shot)","reference_id":"NzgwOTEzOTE5NDk1NDYwMTM5MQ=="} +{"index":3064,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2110.14168","reference_contexts":"The datasets for evaluation include MMLU (Hendrycks et al., 2021a) (5-shot), MMLU-Pro (Wang et al., 2024) (5-shot), GPQA (Rein et al., 2023 ) (5shot), Theorem QA (Chen et al., 2023a) (5-shot), BBH (Suzgun et al., 2023) (3-shot), HellaSwag (Zellers et al., 2019) (10-shot), Winogrande (Sakaguchi et al., 2021) (5-shot), TruthfulQA (Lin et al., 2022a ) (0-shot), ARC-C (Clark et al., 2018) (25-shot), HumanEval (Chen et al., 2021 ) (0-shot), MBPP (Austin et al., 2021 ) (0-shot), EvalPlus (Liu et al., 2023a ) (0-shot), MultiPL-E (Cassano et al., 2023 ) (0-shot on Python, C++, Java, PHP, Type-Script, C#, Bash, and JavaScript), GSM8K (Cobbe et al., 2021) (5-shot), MATH (Hendrycks et al., 2021b ) (4-shot), C-Eval (Huang et al., 2023) (5-shot), and CMMLU (Li et al., 2023) (5-shot)","reference_id":"NzgwOTEzOTE5NDk1NDYwMTM5MQ=="} +{"index":3065,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2110.14168","reference_contexts":"Meanwhile, Qwen2-72B, the base language model, achieves 84.2 on MMLU (Hendrycks et al., 2021a) , 37.9 on GPQA (Rein et al., 2023) , 64.6 on HumanEval (Chen et al., 2021) , 89.5 on GSM8K (Cobbe et al., 2021) , and 82.4 on BBH (Suzgun et al., 2023) .","reference_id":"MTE1ODQ3OTYxODAxNjMxNzI2MjI="} +{"index":3066,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2401.06066","reference_contexts":"Recently, there has been a notable trend towards integrating both shared and routing-specific experts within MoE layers (Rajbhandari et al., 2022; Dai et al., 2024)","reference_id":"ODE1MzI1NDUyNTI3MDI2OTk0OA=="} +{"index":3067,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2401.06066","reference_contexts":"Differently, our model employs fine-grained experts (Dai et al., 2024) , creating smaller-scale experts while activating a greater number of experts simultaneously","reference_id":"NTU1MjU5NjY2MzAyNTQyMzEzMQ=="} +{"index":3068,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.07974","reference_contexts":"Specifically, we evaluate on MMLU, MMLU-Pro, GPQA, and Theorem QA for language understanding and knowledge, HumanEval, MBPP, MultiPL-E, and LiveCodeBench v1 (Jain et al., 2024) for coding, GSM8K and MATH for mathematics","reference_id":"OTk1ODE2NzUyOTM4MDY3MjY2MA=="} +{"index":3069,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.07974","reference_contexts":"Qwen2 outperforms competing models in evaluations of both fundamental language capabilities and instruction-tuned functionalities Specifically, Qwen2-72B-Instruct, our instruction-tuned variant, scores 9.1 on MT-Bench (Zheng et al., 2023) , 48.1 on Arena-Hard (Chiang et al., 2024) , and 35.7 on LiveCodeBench (Jain et al., 2024)","reference_id":"NDk2MDYxNDQwMDgxNDE4MzI0Ng=="} +{"index":3070,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2310.06825","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3071,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2401.04088","reference_contexts":"For example, transitioning from Mistral-7B (Jiang et al., 2023a) to Mixtral 8x7B (Jiang et al., 2024) , involves activating two of the eight experts at a time","reference_id":"NTQyNTU1MjI5MDMxMTU2Mzc0Mg=="} +{"index":3072,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2401.04088","reference_contexts":"In terms of the largest model of Qwen2, we compare Qwen2-72B with competitive baseline open-weight models, including Mixtral-8x22B (Jiang et al., 2024) , Llama-3-70B (AI@Meta, 2024), as well as Qwen1.5-72B (Qwen Team, 2024a) and Qwen1.5-110B (Qwen Team, 2024b).","reference_id":"MzcxMDMxNzU0NDkxNTkwNTA3MA=="} +{"index":3073,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2401.04088","reference_contexts":"These baselines include other MoE models, such as Mixtral-8x7B (Jiang et al., 2024) and Jamba (Lieber et al., 2024) , and dense models, such as Yi-1.5-34B (Young et al., 2024) Table 3 : Performance of the 30B+ dense models and 40B+ MoE models","reference_id":"MTY2MDkwNzYyOTQyMjE4MzQ0MzI="} +{"index":3074,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2306.09212","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3075,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2406.11939","reference_contexts":"Additionally, we assess the performance of human preference alignment and instruction following by evaluating on benchmarks including MT-Bench (Zheng et al., 2023) , Arena-Hard (Li et al., 2024) , AlignBench (Liu et al., 2023b) , MixEval (Ni et al., 2024) whose results approximate those of Chatbot Arena, and IFEval (Zhou et al., 2023) foot_0 for instruction following.","reference_id":"MTQ2MTc2ODU2Njc5MDMwNDAyNTc="} +{"index":3076,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.08295","reference_contexts":"This comparison also includes Mistral-7B-v0.2 (Jiang et al., 2023a) , Gemma-7B (Mesnard et al., 2024) , and our predecessor, Qwen1.5-7B (Qwen Team, 2024a).","reference_id":"Mjg4ODA4MjE2MjMzOTMwNTk2Nw=="} +{"index":3077,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.08295","reference_contexts":"Qwen2-1.5B & Qwen2-0.5B To evaluate the performance of our smaller models, specifically Qwen2-1.5B and Qwen2-0.5B, we compare them against established baselines: Phi-2 (Abdin et al., 2024), Gemma-2B (Mesnard et al., 2024) , and Qwen1.5-1.8B (Qwen Team, 2024a)","reference_id":"MjYxNzY1ODY3Mjc4MTQxODUyNA=="} +{"index":3078,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.08295","reference_contexts":"Many of these models, including Qwen (Bai et al., 2023a) , Mistral (Jiang et al., 2023a) , Gemma (Mesnard et al., 2024) , etc., have been released in an open-weight manner.","reference_id":"NjgwOTQ3ODEwMzM4OTI3MDEyOA=="} +{"index":3079,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2303.08774","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3080,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2309.00071","reference_contexts":"Moreover, we also employ YARN (Peng et al., 2023) to rescale the attention weights for better length extrapolation.","reference_id":"MTQ0MzUwMTc2NzY4MjYxNTk5OQ=="} +{"index":3081,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2309.00071","reference_contexts":"To fully leverage the model's length extrapolation potential, we adopted the YARN mechanism (Peng et al., 2023) and the Dual Chunk Attention mechanism (An et al., 2024)","reference_id":"NzQyODk3NzUwNjA1MDk2MjkzMw=="} +{"index":3082,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2309.00071","reference_contexts":"For contexts over 32K, YARN (Peng et al., 2023) was applied in this evaluation","reference_id":"NjQwNzE1NjY4NTA3ODE4NjczMQ=="} +{"index":3083,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2311.12022","reference_contexts":"The datasets for evaluation include MMLU (Hendrycks et al., 2021a) (5-shot), MMLU-Pro (Wang et al., 2024) (5-shot), GPQA (Rein et al., 2023 ) (5shot), Theorem QA (Chen et al., 2023a) (5-shot), BBH (Suzgun et al., 2023) (3-shot), HellaSwag (Zellers et al., 2019) (10-shot), Winogrande (Sakaguchi et al., 2021) (5-shot), TruthfulQA (Lin et al., 2022a ) (0-shot), ARC-C (Clark et al., 2018) (25-shot), HumanEval (Chen et al., 2021 ) (0-shot), MBPP (Austin et al., 2021 ) (0-shot), EvalPlus (Liu et al., 2023a ) (0-shot), MultiPL-E (Cassano et al., 2023 ) (0-shot on Python, C++, Java, PHP, Type-Script, C#, Bash, and JavaScript), GSM8K (Cobbe et al., 2021) (5-shot), MATH (Hendrycks et al., 2021b ) (4-shot), C-Eval (Huang et al., 2023) (5-shot), and CMMLU (Li et al., 2023) (5-shot)","reference_id":"NzgwOTEzOTE5NDk1NDYwMTM5MQ=="} +{"index":3084,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2311.12022","reference_contexts":"Meanwhile, Qwen2-72B, the base language model, achieves 84.2 on MMLU (Hendrycks et al., 2021a) , 37.9 on GPQA (Rein et al., 2023) , 64.6 on HumanEval (Chen et al., 2021) , 89.5 on GSM8K (Cobbe et al., 2021) , and 82.4 on BBH (Suzgun et al., 2023) .","reference_id":"MTE1ODQ3OTYxODAxNjMxNzI2MjI="} +{"index":3085,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2302.13971","reference_contexts":"The release of the Llama series (Touvron et al., 2023) has further ignited interests within the open-source community, particularly regarding GPT-level local LLMs","reference_id":"MTA5Mzk4MzIwOTczNzc3MzY5MjA="} +{"index":3086,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2406.01574","reference_contexts":"The datasets for evaluation include MMLU (Hendrycks et al., 2021a) (5-shot), MMLU-Pro (Wang et al., 2024) (5-shot), GPQA (Rein et al., 2023 ) (5shot), Theorem QA (Chen et al., 2023a) (5-shot), BBH (Suzgun et al., 2023) (3-shot), HellaSwag (Zellers et al., 2019) (10-shot), Winogrande (Sakaguchi et al., 2021) (5-shot), TruthfulQA (Lin et al., 2022a ) (0-shot), ARC-C (Clark et al., 2018) (25-shot), HumanEval (Chen et al., 2021 ) (0-shot), MBPP (Austin et al., 2021 ) (0-shot), EvalPlus (Liu et al., 2023a ) (0-shot), MultiPL-E (Cassano et al., 2023 ) (0-shot on Python, C++, Java, PHP, Type-Script, C#, Bash, and JavaScript), GSM8K (Cobbe et al., 2021) (5-shot), MATH (Hendrycks et al., 2021b ) (4-shot), C-Eval (Huang et al., 2023) (5-shot), and CMMLU (Li et al., 2023) (5-shot)","reference_id":"NzgwOTEzOTE5NDk1NDYwMTM5MQ=="} +{"index":3087,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2309.16039","reference_contexts":"In conjunction with these enhancements, we modified the base frequency of RoPE from 10,000 to 1,000,000 to optimize performance in long-context scenarios (Xiong et al., 2023) .","reference_id":"MTA0ODcwNzkyMjM4NjcwNDczMTA="} +{"index":3088,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2403.04652","reference_contexts":"These baselines include other MoE models, such as Mixtral-8x7B (Jiang et al., 2024) and Jamba (Lieber et al., 2024) , and dense models, such as Yi-1.5-34B (Young et al., 2024) Table 3 : Performance of the 30B+ dense models and 40B+ MoE models","reference_id":"MTY2MDkwNzYyOTQyMjE4MzQ0MzI="} +{"index":3089,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2402.05136","reference_contexts":"Llama-3-8B Yi-1.5-9B GLM-4-9B Qwen1.5-7B Qwen2-7B Kamradt, 2023) , NeedleBench (OpenCompass Contributors, 2023), and LV-Eval (Yuan et al., 2024) .","reference_id":"MTUxNzcxMDk0MDQ1NTE4OTM4Njc="} +{"index":3090,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2308.01825","reference_contexts":"Rejection Sampling For mathematical or similar tasks with definitive final answers, rejection sampling (Yuan et al., 2023) is applied to improve the quality of solutions","reference_id":"MTc3MzE1OTgyODcyOTkwODU0Mg=="} +{"index":3091,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2308.01825","reference_contexts":"Additionally, we assess the performance of human preference alignment and instruction following by evaluating on benchmarks including MT-Bench (Zheng et al., 2023) , Arena-Hard (Li et al., 2024) , AlignBench (Liu et al., 2023b) , MixEval (Ni et al., 2024) whose results approximate those of Chatbot Arena, and IFEval (Zhou et al., 2023) foot_0 for instruction following.","reference_id":"MTQ2MTc2ODU2Njc5MDMwNDAyNTc="} +{"index":3092,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2308.01825","reference_contexts":"Qwen2 outperforms competing models in evaluations of both fundamental language capabilities and instruction-tuned functionalities Specifically, Qwen2-72B-Instruct, our instruction-tuned variant, scores 9.1 on MT-Bench (Zheng et al., 2023) , 48.1 on Arena-Hard (Chiang et al., 2024) , and 35.7 on LiveCodeBench (Jain et al., 2024)","reference_id":"NDk2MDYxNDQwMDgxNDE4MzI0Ng=="} +{"index":3093,"citer_arxiv_id":"2407.10671","cited_arxiv_id":"2311.07911","reference_contexts":"Additionally, we assess the performance of human preference alignment and instruction following by evaluating on benchmarks including MT-Bench (Zheng et al., 2023) , Arena-Hard (Li et al., 2024) , AlignBench (Liu et al., 2023b) , MixEval (Ni et al., 2024) whose results approximate those of Chatbot Arena, and IFEval (Zhou et al., 2023) foot_0 for instruction following.","reference_id":"MTQ2MTc2ODU2Njc5MDMwNDAyNTc="} +{"index":3094,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1906.00910","reference_contexts":"CPC and Deep InfoMax have recently been extended in [29] and [4] respectively","reference_id":"NTUzNjY3NDY0NTA2ODI5MDE0NQ=="} +{"index":3095,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1906.00910","reference_contexts":"This strategy is also used in [4, 29, 27] and dates back to [72] .","reference_id":"MTIxMjU3NzI2OTM4NTkxMjg1NzY="} +{"index":3096,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1906.00910","reference_contexts":"On top of it, strengthening data augmentation with Ran-dAugment [14] yields better or comparable results to other state-of-the-art methods [41, 79, 87, 27, 49, 19, 29, 4]","reference_id":"MTQyOTc4NDgzNzAyOTI0NDU2MzQ="} +{"index":3097,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1801.04062","reference_contexts":"We then use the MINE estimator [5] to estimate the mutual information between the views","reference_id":"MTc3MDE1NDQwNzQ3NDU3NDk1Njc="} +{"index":3098,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1909.13719","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3099,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1911.05722","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3100,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1905.09272","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3101,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1511.06811","reference_contexts":"A powerful family of models for unsupervised representations are collected under the umbrella of \"self-supervised\" learning [64, 35, 85, 84, 78, 60, 83]","reference_id":"MzQzNjg2NTk0MTUzNjgxMjI2"} +{"index":3102,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1312.6114","reference_contexts":"This idea shows up in contemporary representation learning in the form of autoencoders [65] and generative models [40, 24] , which try to represent a data point or distribution as losslessly as possible","reference_id":"MTI1NjQ2MjQzMTE4NjI0NjMwMDg="} +{"index":3103,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1312.6114","reference_contexts":"Representations better suited to such tasks have been learnt using deep neural networks, starting with seminal techniques such as Boltzmann machines [71, 65], autoencoders [30] , variational autoencoders [40] , generative adversarial networks [24] and autoregressive models [56]","reference_id":"NTAyNDU1MDM5MzIyOTk4MjM5Mw=="} +{"index":3104,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1511.06856","reference_contexts":"As many previous unsupervised methods are evaluated with AlexNet [43] on ImageNet [16, 42, 17, 84, 54, 18, 85, 55, 22, 11, 83] , we also include the the results of CMC using this network","reference_id":"NTE2NzAyMzE0MzgxNDIwNDIyNA=="} +{"index":3105,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1912.01991","reference_contexts":"To test the compatibility of CMC with mechanisms proposed in other self-supervised learning methods, we consider combining CMC with MoCo [27] (i.e., switching from memory bank [79] to momentum encoder) and PIRL [49] (i.e., applying a second JigSaw branch)","reference_id":"MTAwNjgxNzUxOTIyMTkyMjIzMDk="} +{"index":3106,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1912.01991","reference_contexts":"On top of it, strengthening data augmentation with Ran-dAugment [14] yields better or comparable results to other state-of-the-art methods [41, 79, 87, 27, 49, 19, 29, 4]","reference_id":"MTQyOTc4NDgzNzAyOTI0NDU2MzQ="} +{"index":3107,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1601.06759","reference_contexts":"Representations better suited to such tasks have been learnt using deep neural networks, starting with seminal techniques such as Boltzmann machines [71, 65], autoencoders [30] , variational autoencoders [40] , generative adversarial networks [24] and autoregressive models [56]","reference_id":"NTAyNDU1MDM5MzIyOTk4MjM5Mw=="} +{"index":3108,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"We also investigated the strided crop strategy of CPC [57]","reference_id":"MzMyMDYxNjQzNDYxMjM3NDM4"} +{"index":3109,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"The overall proof follows a similar derivation introduced in [57] .","reference_id":"MTMyNTEzNTAxOTY2MDMzOTAyMDc="} +{"index":3110,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"Examples of such X/Y pairs are: luminance and chrominance color channels of an image [85] , patches from a single image [57] , modalities such as vision and sound [58] or the frames of a video [78]","reference_id":"MTQ5NjA0ODgzNjM4ODc2NTQxODI="} +{"index":3111,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"Some of the prior works most similar to our own (and inspirational to us) are Contrastive Predictive Coding (CPC) [57] , Deep InfoMax [31] , and Instance Discrimination [79]","reference_id":"MTM5OTU4MzUwNDc5Njg2OTAyMDk="} +{"index":3112,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"Similar to recent setups for contrastive learning [57, 25, 51] , we train this function to correctly select a single positive sample x out of a set S = {x, y 1 , y 2 , ..., y k } that contains k negative samples:","reference_id":"MTM3OTc1NTI0NDk5MDk0Mzk3NDQ="} +{"index":3113,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"[26, 81, 72, 25, 31, 57, 3, 29, 36 ] that take advantage of the co-occurence of multiple views","reference_id":"MTQ2MjM3OTMzMTM2MjA1NjA1MjQ="} +{"index":3114,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"For a fair comparison with DIM [31] and CPC [57] , we adopt the same architecture as that used in DIM and split it into two encoders, each shown as in Table 10","reference_id":"NjUwMDk0ODYxMTUwMzM3NjY2Mg=="} +{"index":3115,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"A formal proof is given by [57, 62] , which we recapitulate in supplement, showing that:","reference_id":"MjcwMDIxNTYxNDM4MTI3NTE3Mg=="} +{"index":3116,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"In this fashion, we directly perform m + 1 way softmax classification, the same as [57, 31] for a fair comparison in Sec","reference_id":"MTE3MDMzMTA5MjI0MjAzMTI1MA=="} +{"index":3117,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"1 Confusingly, the literature has previously referred to Eqn.2 as \"In-foNCE\" [57]","reference_id":"NTc3MzUwMzQ1Mzk0NDAxNTU3OA=="} +{"index":3118,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"In particular, we adapt the recently proposed method of Contrastive Predictive Coding (CPC) [57] , except we simplify it -removing the recurrent network -and generalize it -showing how to apply it to arbitrary collections of image channels, rather than just to temporal or spatial predictions","reference_id":"NzkxMDg2NjA1MzgzNzUxODgwMQ=="} +{"index":3119,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1807.03748","reference_contexts":"As proposed by previous literature [57, 31, 3, 87, 79] , the quality of such a representation is evaluated by freezing the weights of encoder and training linear classifier on top of each layer.","reference_id":"MTExMTI1NjcxNjE4MTkzNDkxODc="} +{"index":3120,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1212.0402","reference_contexts":"We train CMC on UCF101 [73] and use two CaffeNets [43] for extracting features from images and optical flows, respectively","reference_id":"MTgzNzUzMjg2NTIxMzk5ODAyMw=="} +{"index":3121,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1906.05743","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3122,"citer_arxiv_id":"1906.05849","cited_arxiv_id":"1903.12355","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3123,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.00861","reference_contexts":"Models from Anthropic (Askell et al., 2021) and OpenAI (InstructGPT (Ouyang et al., 2021) , WebGPT (Nakano et al., 2021) ) demonstrate significant progress on TruthfulQA relative to the original GPT-3 baseline","reference_id":"NTk5MTQyNzM5OTc1Nzk0MTIyNQ=="} +{"index":3124,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.00861","reference_contexts":"Appendix B.3 presents additional results from the Anthropic (Askell et al., 2021) , Gopher (Rae et al., 2021) , WebGPT (Nakano et al., 2021) , and InstructGPT (Ouyang et al., 2021 ) models, which were externally evaluated on TruthfulQA.","reference_id":"MTAyOTU0MDg0ODYwNTgzNjg5Mzg="} +{"index":3125,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.00861","reference_contexts":"The prompt is designed to encourage answers that are \"helpful, honest, and harmless\" (Askell et al., 2021) .","reference_id":"MTMyOTQ1MzExNTE3MDY3MjkzMjk="} +{"index":3126,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2101.03961","reference_contexts":"Large companies are deploying their own models (Raffel et al., 2019; Fedus et al., 2021) , and hundreds of organizations are deploying GPT-3 via APIs from OpenAI and other firms (OpenAI, 2020; Wolf et al., 2020; CohereAI, 2021; OpenAI, 2021)","reference_id":"OTYzOTI5OTkxNDc5MTgwOTcyOQ=="} +{"index":3127,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2101.00027","reference_contexts":"GPT-Neo/J (Black et al., 2021; Wang and Komatsuzaki, 2021 ) is a variant of GPT-3 with a different training set (Gao et al., 2020) .","reference_id":"MjA5MTg2MDc2ODE0OTQ4MzAzMg=="} +{"index":3128,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2009.03300","reference_contexts":"reason to focus on imitative falsehoods is that they are less likely to be covered by existing question-answering benchmarks (Clark et al., 2018; Kwiatkowski et al., 2019; Joshi et al., 2017; Hendrycks et al., 2020)","reference_id":"MTc4MDA3NjIzMzgyOTE4MDM5NA=="} +{"index":3129,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2009.03300","reference_contexts":"Numerous NLP benchmarks test models on factual questions (Bhakthavatsalam et al., 2021; Clark et al., 2018; Hendrycks et al., 2020; Talmor et al., 2019)","reference_id":"MzQ4MzU5MTY1NTgwMzkyNDUwNQ=="} +{"index":3130,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2103.03874","reference_contexts":"While most benchmarks are multiple choice, some require models to generate short (single-phrase) answers (Hendrycks et al., 2021; Lewis et al., 2020) .","reference_id":"ODI2NzEzNTY4ODg2ODIyMDEwOA=="} +{"index":3131,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2001.08361","reference_contexts":"This \"inverse scaling\" trend contrasts with most tasks in NLP, where performance improves with model size (Brown et al., 2020; Kaplan et al., 2020)","reference_id":"MTM2MTMyNjg4NDk1NTk0OTA3MTE="} +{"index":3132,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2001.08361","reference_contexts":"TruthfulQA is intended as a zero-shot benchmark (Brown et al., 2020; Wei et al., 2021)","reference_id":"MTA0MzY1Mzk1NzczOTA2NDEzNTI="} +{"index":3133,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2001.08361","reference_contexts":"Another reason is that scaling laws suggest that scaling up models will reduce perplexity on the training distribution (Kaplan et al., 2020)","reference_id":"NDgzNDE0MzgwMjcwMzYzMDI4Ng=="} +{"index":3134,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2001.08361","reference_contexts":"GPT-3 fails to reliably generalize from its training data about multiplication (Brown et al., 2020)","reference_id":"MTQ3OTIxNjYzNjExMjczMjA5NTQ="} +{"index":3135,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2103.14659","reference_contexts":"The problem of imitative falsehoods is similar to models learning to imitate offensive or prejudiced language (Kenton et al., 2021; Bender et al., 2021 )","reference_id":"ODUzNjQ5MDQ4OTY5MTEzMTg0Mw=="} +{"index":3136,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2005.11401","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3137,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2009.06807","reference_contexts":"Increasing the safety and alignment of pre-trained models remains a challenging problem (Dinan et al., 2020; Tamkin et al., 2021; Xu et al., 2020; Solaiman and Dennison, 2021; McGuffie and Newhouse, 2020) .","reference_id":"MTAyODc5NTQ1MzI5NTIyNjc1MjU="} +{"index":3138,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.09332","reference_contexts":"Models from Anthropic (Askell et al., 2021) and OpenAI (InstructGPT (Ouyang et al., 2021) , WebGPT (Nakano et al., 2021) ) demonstrate significant progress on TruthfulQA relative to the original GPT-3 baseline","reference_id":"NTk5MTQyNzM5OTc1Nzk0MTIyNQ=="} +{"index":3139,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.09332","reference_contexts":"Appendix B.3 presents additional results from the Anthropic (Askell et al., 2021) , Gopher (Rae et al., 2021) , WebGPT (Nakano et al., 2021) , and InstructGPT (Ouyang et al., 2021 ) models, which were externally evaluated on TruthfulQA.","reference_id":"MTAyOTU0MDg0ODYwNTgzNjg5Mzg="} +{"index":3140,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.09332","reference_contexts":"WebGPT is a GPT-3 based model that is given access to a text-based web browser and search engine that it can use to answer questions (Nakano et al., 2021) .","reference_id":"NDA3NzA0NDU3MTcwOTQ0MTE0MA=="} +{"index":3141,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2203.02155","reference_contexts":"Models from Anthropic (Askell et al., 2021) and OpenAI (InstructGPT (Ouyang et al., 2021) , WebGPT (Nakano et al., 2021) ) demonstrate significant progress on TruthfulQA relative to the original GPT-3 baseline","reference_id":"NTk5MTQyNzM5OTc1Nzk0MTIyNQ=="} +{"index":3142,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2203.02155","reference_contexts":"Appendix B.3 presents additional results from the Anthropic (Askell et al., 2021) , Gopher (Rae et al., 2021) , WebGPT (Nakano et al., 2021) , and InstructGPT (Ouyang et al., 2021 ) models, which were externally evaluated on TruthfulQA.","reference_id":"MTAyOTU0MDg0ODYwNTgzNjg5Mzg="} +{"index":3143,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2203.02155","reference_contexts":"InstructGPT is a GPT-3 based model that is finetuned with human preferences to follow natural language instructions (Ouyang et al., 2021) .","reference_id":"MTAwMTUzODU5NTMxODgxNjQ5MTM="} +{"index":3144,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2105.11447","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3145,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.11446","reference_contexts":"Gopher is a 280-billion parameter model whose pre-training data was more heavily filtered for high-quality, scientific sources (Rae et al., 2021) .","reference_id":"NTE2OTI3MjQzMjUyMTMyMTE0OQ=="} +{"index":3146,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2112.11446","reference_contexts":"Appendix B.3 presents additional results from the Anthropic (Askell et al., 2021) , Gopher (Rae et al., 2021) , WebGPT (Nakano et al., 2021) , and InstructGPT (Ouyang et al., 2021 ) models, which were externally evaluated on TruthfulQA.","reference_id":"MTAyOTU0MDg0ODYwNTgzNjg5Mzg="} +{"index":3147,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"1910.10683","reference_contexts":"UnifiedQA (Khashabi et al., 2020 ) is a T5 model (Raffel et al., 2019) fine-tuned on diverse QA tasks","reference_id":"MTcxNzg4MDE1ODM1Njk0MTEyNzU="} +{"index":3148,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"1910.10683","reference_contexts":"Large companies are deploying their own models (Raffel et al., 2019; Fedus et al., 2021) , and hundreds of organizations are deploying GPT-3 via APIs from OpenAI and other firms (OpenAI, 2020; Wolf et al., 2020; CohereAI, 2021; OpenAI, 2021)","reference_id":"OTYzOTI5OTkxNDc5MTgwOTcyOQ=="} +{"index":3149,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2104.07567","reference_contexts":"These range from subtle inaccuracies to wild hallucinations (Shuster et al., 2021; Zhou et al., 2021; Krishna et al., 2021)","reference_id":"MTIyODk3NzA5NzU2NTQ1NzEwMTc="} +{"index":3150,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2104.07567","reference_contexts":"Concepts related to truthfulness in natural language generation include factuality, veracity, and avoiding hallucinations (Shuster et al., 2021; Zhou et al., 2021)","reference_id":"MzUxMjc0Nzk5ODM3MTY3NjE4Mw=="} +{"index":3151,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2009.01325","reference_contexts":"Truthfulness is relevant to many applications including generating news stories (Kreps et al., 2020; Zellers et al., 2019) , summarization (Gabriel et al., 2021; Maynez et al., 2020; Stiennon et al., 2020; Wang et al., 2020) , conversational dialog (Shuster et al., 2021; Roller et al., 2021) , and question answering (Dou et al., 2021; Krishna et al., 2021; Lewis et al., 2020; Logan IV et al., 2019)","reference_id":"MTcxNDQxMzA1NDMwNjA1OTE4MTY="} +{"index":3152,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2009.01325","reference_contexts":"Models could be fine-tuned on a set of examples chosen to demonstrate truthfulness (Solaiman and Dennison, 2021) or fine-tuned by reinforcement learning from human feedback (Stiennon et al., 2020)","reference_id":"MTM0ODEwNzE4ODM1MzEyMDUwMTQ="} +{"index":3153,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2102.02503","reference_contexts":"This could lead to deception and distrust (Tamkin et al., 2021) .","reference_id":"MTQzMzU0NjY3MzMwNTk3MTYzMTY="} +{"index":3154,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2102.02503","reference_contexts":"Increasing the safety and alignment of pre-trained models remains a challenging problem (Dinan et al., 2020; Tamkin et al., 2021; Xu et al., 2020; Solaiman and Dennison, 2021; McGuffie and Newhouse, 2020) .","reference_id":"MTAyODc5NTQ1MzI5NTIyNjc1MjU="} +{"index":3155,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"2010.07079","reference_contexts":"Increasing the safety and alignment of pre-trained models remains a challenging problem (Dinan et al., 2020; Tamkin et al., 2021; Xu et al., 2020; Solaiman and Dennison, 2021; McGuffie and Newhouse, 2020) .","reference_id":"MTAyODc5NTQ1MzI5NTIyNjc1MjU="} +{"index":3156,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"1905.12616","reference_contexts":"Truthfulness is relevant to many applications including generating news stories (Kreps et al., 2020; Zellers et al., 2019) , summarization (Gabriel et al., 2021; Maynez et al., 2020; Stiennon et al., 2020; Wang et al., 2020) , conversational dialog (Shuster et al., 2021; Roller et al., 2021) , and question answering (Dou et al., 2021; Krishna et al., 2021; Lewis et al., 2020; Logan IV et al., 2019)","reference_id":"MTcxNDQxMzA1NDMwNjA1OTE4MTY="} +{"index":3157,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"1905.12616","reference_contexts":"If models can generate plausible false statements in ways that are not easily identifiable, they could be used to deceive humans via disinformation or fraud (Zellers et al., 2019; Schuster et al., 2019)","reference_id":"NjExODc1MjEyNTAyNDY3NzM4MA=="} +{"index":3158,"citer_arxiv_id":"2109.07958","cited_arxiv_id":"1912.08777","reference_contexts":"Paraphrases were auto-generated using a PEGASUSbased paraphrasing model (Zhang et al., 2019; Rajauria, 2020) , then filtered and edited to ensure that the meaning of the question was preserved","reference_id":"MjQzOTAxMTA3MDMyNTgxNjIzNw=="} +{"index":3159,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2103.10951","reference_contexts":"One type of approach towards controllable image generation specifies the underlying content of an image utilizing text through GANs [53, 54, 2] , VQ-VAEs [40] , or diffusion models [33]","reference_id":"MzU1Mzk0Njk5MDMyODY2NjY="} +{"index":3160,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2009.00713","reference_contexts":"While diffusion models have been developed for image generation [47] , they have further proven successful in the generation of waveforms [3] , 3D shapes [55] , decision making [18] , and text [1] , suggesting that our proposed composition operators may further be applied in such domains.","reference_id":"MTE5MjA0NTI4NDU0MzAyMDA5ODU="} +{"index":3161,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2012.01316","reference_contexts":"Energy-Based Models (EBMs) [10, 9, 12, 37] are a class of generative models where the data distribution is modeled using an unnormalized probability density","reference_id":"MTEyMzUyNDA0MTUzMjUyMjQwNzI="} +{"index":3162,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2012.01316","reference_contexts":"We train energy-based models using the codebase from [9] , where we encode discrete labels and continuous labels using an embedding layer and a linear layer, respectively","reference_id":"MTYxODYxODIxOTY3Nzk3NzgyNg=="} +{"index":3163,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1903.08689","reference_contexts":"Energy-Based Models (EBMs) [10, 9, 12, 37] are a class of generative models where the data distribution is modeled using an unnormalized probability density","reference_id":"MTEyMzUyNDA0MTUzMjUyMjQwNzI="} +{"index":3164,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1903.08689","reference_contexts":"Incorporating additional structures into diffusion models from EBMs [10] , such as a conservative score field, can be a promising direction towards compositions of separately trained diffusion models.","reference_id":"NjcxNjQxMzgxMTQ1OTk3MTUxNQ=="} +{"index":3165,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1903.08689","reference_contexts":"The denoising procedure can be seen as parameterizing the gradients of the data distribution [48] , which is similar to EBMs [27, 10, 37, 12, 11]","reference_id":"MzI1Nzk1MDQ4ODU1ODg5MDcxNg=="} +{"index":3166,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1903.08689","reference_contexts":"A gradient based MCMC procedure, Langevin dynamics [10] , is then used to sample from the unnormalized probability distribution to iteratively refine the generated image x:","reference_id":"MTI2Mzk4MTU4MDIyMDY1NzAzNDQ="} +{"index":3167,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1710.10196","reference_contexts":"In the setting where only one concept c 1 is conditioned for sampling, the above equation will reduce to classifier-free guidance [16] : εθ (x t , t|c 1 ) = ϵ θ (x t , t) + w ϵ θ (x t , t|c 1 ) -ϵ θ (x t , t) , (20) where the temperature scaling w > 1.","reference_id":"MjUyMDQ1MTkwODU2MDYxMTk5MQ=="} +{"index":3168,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1710.10196","reference_contexts":"We use 30, 000 image-label pairs from CelebA-HQ [20] to train a classifier to evaluate the generated images","reference_id":"OTM1OTA1OTU2MzE3Njc1OTcxMw=="} +{"index":3169,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1412.6980","reference_contexts":"We use the Adam optimizer [25] with β 1 = 0 and β 2 = 0.99 to train the StyleGAN2 models","reference_id":"ODIzOTY5MDcwMDY3OTYwMjQ5OQ=="} +{"index":3170,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1412.6980","reference_contexts":"We use the Adam optimizer [25] with a learning rate of 10 -4","reference_id":"MTMyOTUwOTk2MjYwNTU3NDY0OTQ="} +{"index":3171,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1412.6980","reference_contexts":"We use the Adam optimizer [25] with β 1 = 0 and β 2 = 0.99 to train the models.","reference_id":"NTM5ODM0ODc0OTU4MzA5OTA2NA=="} +{"index":3172,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1711.05101","reference_contexts":"We use the AdamW optimizer [30] with β 1 = 0.9 and β 2 = 0.999","reference_id":"OTE3MDY3NjQ5MTM0NTk5NjEwNA=="} +{"index":3173,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2204.13807","reference_contexts":"However, such systems are not fully compositional and generate incorrect images when given more complex descriptions [31, 49]","reference_id":"MTMwMTc1NjQ5MzEwMzcyNzIzNg=="} +{"index":3174,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"We develop Composed GLIDE (Ours), a version of GLIDE [33] that utilizes our compositional operators to combine textual descriptions, without further training","reference_id":"NTI4NjgzNDY2MDQxMDU1MTc4NA=="} +{"index":3175,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"We use the small GLIDE model released by [33] in our experiments","reference_id":"MTQ5NjQ5ODg4MjA4OTg1NTcwNA=="} +{"index":3176,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"First, we illustrate how our approach may be applied to large pre-trained diffusion models, such as Stable Diffusion [42] and GLIDE [33] , to compose multiple text descriptions","reference_id":"NjAyMzIyNjQxMzE2MTkyOTEzMg=="} +{"index":3177,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"To generate FFHQ images, we first generate images at 64 × 64 resolution and then upsample the images to 256 × 256 using a sampler provided by [33]","reference_id":"MTQ4MDk5MDgzMjA3MTkxOTg1OTU="} +{"index":3178,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"Diffusion models have recently shown great promise in image generation tasks [6] , enabling effective image editing [32, 24] , text conditioning [33, 41, 13] , and image inpainting [43]","reference_id":"ODA5NDI1MTA3NjIzNjgzOTM2"} +{"index":3179,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"Our approach is implemented based on the code from [35, 33]","reference_id":"MTY0Nzg2NTgwMTQ2MjYyMzgzMDY="} +{"index":3180,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"One type of approach towards controllable image generation specifies the underlying content of an image utilizing text through GANs [53, 54, 2] , VQ-VAEs [40] , or diffusion models [33]","reference_id":"MzU1Mzk0Njk5MDMyODY2NjY="} +{"index":3181,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"We first show the image generation results of the pre-trained diffusion model, GLIDE [33] , in Figure 3","reference_id":"MTM2Njc2NTUxOTA1NDczOTYyMzA="} +{"index":3182,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"GLIDE [33] is a recently released text-conditioned diffusion model for image generation","reference_id":"NzkxNTg0OTUzNDE4Nzc3NDI5"} +{"index":3183,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2112.10741","reference_contexts":"This is because the pre-trained diffusion model, GLIDE [33] , is trained to avoid generating human images","reference_id":"MTQxMTMzMjgwNzIxNDM3NTI1NDM="} +{"index":3184,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2212.08751","reference_contexts":"We first use Point-E [34] , the pre-trained 3D point cloud generation model, to generate the point clouds of an object based on the text description","reference_id":"OTUwMDk3NzA0NDQ3NDQ4NDY2Mg=="} +{"index":3185,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2212.08751","reference_contexts":"We provide qualitative results of composing the pre-trained text-to-3D diffusion model, Point-E [34] , to generate interesting 3D hybrid objects.","reference_id":"MzQ2MjA1ODExOTg5NDA0MDY1Mw=="} +{"index":3186,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1903.12370","reference_contexts":"Energy-Based Models (EBMs) [10, 9, 12, 37] are a class of generative models where the data distribution is modeled using an unnormalized probability density","reference_id":"MTEyMzUyNDA0MTUzMjUyMjQwNzI="} +{"index":3187,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"1903.12370","reference_contexts":"The denoising procedure can be seen as parameterizing the gradients of the data distribution [48] , which is similar to EBMs [27, 10, 37, 12, 11]","reference_id":"MzI1Nzk1MDQ4ODU1ODg5MDcxNg=="} +{"index":3188,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2204.06125","reference_contexts":"Existing text-conditioned diffusion models such as DALL-E 2 [39] have recently made remarkable strides towards compositional generation, and are capable of generating photorealistic images given textual descriptions","reference_id":"NDQ5MjI3NDA5NDUzNDAyMDUyMw=="} +{"index":3189,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2111.05826","reference_contexts":"Diffusion models have recently shown great promise in image generation tasks [6] , enabling effective image editing [32, 24] , text conditioning [33, 41, 13] , and image inpainting [43]","reference_id":"ODA5NDI1MTA3NjIzNjgzOTM2"} +{"index":3190,"citer_arxiv_id":"2206.01714","cited_arxiv_id":"2011.13456","reference_contexts":"The denoising procedure can be seen as parameterizing the gradients of the data distribution [48] , which is similar to EBMs [27, 10, 37, 12, 11]","reference_id":"MzI1Nzk1MDQ4ODU1ODg5MDcxNg=="} +{"index":3191,"citer_arxiv_id":"2210.09477","cited_arxiv_id":"2204.02491","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3192,"citer_arxiv_id":"1904.09728","cited_arxiv_id":"1811.01088","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3193,"citer_arxiv_id":"1904.09728","cited_arxiv_id":"1806.03822","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3194,"citer_arxiv_id":"1904.09728","cited_arxiv_id":"1907.10641","reference_contexts":"We also apply a lightweight form of adversarial filtering to make the task more challenging by using a deep stylistic classifier to remove easier examples on the dev and test sets (Sakaguchi et al., 2019)","reference_id":"MTU5MTM1NjA4MzA0MTAzMjYwMjI="} +{"index":3195,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"1801.01973","reference_contexts":"For inference, we run the diffusion process by 50 steps for each image with PNDM [22] noise scheduler","reference_id":"MTY5MzU0MzA1Njg0MDY4OTAzMQ=="} +{"index":3196,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"1504.00325","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3197,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2212.05032","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3198,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"1811.12231","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3199,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2212.09611","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3200,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2207.12598","reference_contexts":"We use the default guidance scale of 7.5 for classifier-free guidance [14] .","reference_id":"NTY0NTQ4NjkyMjkwNDYyMDc0Nw=="} +{"index":3201,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2207.12598","reference_contexts":"During inference, the special identifier is used as the negative prompt for classifier-free guidance [14] to avoid generating non-preferred images.","reference_id":"MTIzNTQzMTI3NDYwNDMxNTExNzc="} +{"index":3202,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"1412.6980","reference_contexts":"The LoRA weights are trained for 10k iterations with the AdamW [19] optimizer with a learning rate of 1 × 10 -5 and a weight decay of 1 × 10 -2 , which is kept constant during training","reference_id":"MTA0NTI4MjI4NzgwNzcxNzQxNTc="} +{"index":3203,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"1412.6980","reference_contexts":"The model is trained by the AdamW optimizer [19] with a learning rate of 1.7 × 10 -5 for 1 epoch","reference_id":"NTc1OTUwMTYyODQ4NzQ2Nzc4NQ=="} +{"index":3204,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2302.12192","reference_contexts":"[12] and [21] use similar methodology to improve text-to-image models, which are highly related to our work","reference_id":"Mjk2ODcyNTQ0MDQ4NzczMDE2NQ=="} +{"index":3205,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2202.09778","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3206,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"1511.02793","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3207,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2204.06125","reference_contexts":"In each pair, the images are generated by DALL•E [30] and Stable Diffusion [32] with the same text prompt","reference_id":"MTU5MDgyNzIwNzkyMDc5MTQ1NDc="} +{"index":3208,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2204.06125","reference_contexts":"Several following works, including DALL•E 2 [30] , GLIDE [26], Imagen [35] , ERNIE-ViLG [9, 47], Stable Diffusion [32] , bring the magic of textto-image generation to the public attention","reference_id":"MTY3NzEyMzI2MTc2ODMzODE2NzY="} +{"index":3209,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2204.06125","reference_contexts":"The recent progress in diffusion models [26, 30, 35, 32] has enabled impressive advancements in text-to-image generation, with many models now being deployed in realworld applications such as DALL•E [30] and Stable Diffusion [32]","reference_id":"NDg4NjAwODc3NTU2NzIyNjQ3"} +{"index":3210,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2204.06125","reference_contexts":"Some recent studies [26, 30, 3 ] use the CLIP [29] model as a proxy for human judgment to evaluate the alignment between generated images and text prompts","reference_id":"MTAwNjIyOTY3Mjk4MTY5NDEyOTM="} +{"index":3211,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2204.06125","reference_contexts":"Moreover, it can capture users' intentions by encoding text prompts, making it a plausible choice for evaluating the alignment between a prompt and a generated image [26, 30, 3, 32]","reference_id":"MTQ2Mzk0NjI4NDQyMTA3NDYyODY="} +{"index":3212,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2210.14896","reference_contexts":"The prompts are randomly sampled from DiffusionDB [45] , which is a large database of images and prompts sourced from the Stable Foundation Discord channel","reference_id":"MjQ4ODM0Mzc0MTQ0ODQ3MTAxMA=="} +{"index":3213,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2210.14896","reference_contexts":"DiffusionDB [45] is a largescale open-source database collected from the Stable Foundation Discord channel, containing the text prompt and parameters for each image","reference_id":"NjAxMDkwNDMyMDc5MDU0NzgxMg=="} +{"index":3214,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2210.14896","reference_contexts":"We construct the training data from the \"large first 1m\" split of DiffusionDB [45] , and a subset of the pre-train dataset of Stable Diffusion (LAION-5B) for regularization","reference_id":"Nzk2NDkyOTQ1NTE4NTE4Nzk5Nw=="} +{"index":3215,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2210.14896","reference_contexts":"In this study, we randomly sample 100 user-provided prompts from DiffusionDB [45]","reference_id":"MTQyNDQ0MjI1OTQ1MTc2NzQzMzA="} +{"index":3216,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2302.13848","reference_contexts":"DreamBooth [33] and ELITE [46] explore customizing Stable Diffusion to a certain object","reference_id":"MTMyOTcxMzU1OTE3OTE4MzYyNDA="} +{"index":3217,"citer_arxiv_id":"2303.14420","cited_arxiv_id":"2112.15283","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3218,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2112.10684","reference_contexts":"Table 1 : A list of LMs used in the experiments: GPT-2 (Radford et al., 2019) , MetaICL (Min et al., 2021b) , GPT-J (Wang and Komatsuzaki, 2021) , fairseq LMs (Artetxe et al., 2021) and GPT-3 (Brown et al., 2020)","reference_id":"MTMzMjEzNDMxMzY1MDg3MDAzNjk="} +{"index":3219,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2112.10684","reference_contexts":"is consistent over 12 different models including the GPT-3 family (Radford et al., 2019; Min et al., 2021b; Wang and Komatsuzaki, 2021; Artetxe et al., 2021; Brown et al., 2020)","reference_id":"MTYyNTAyOTM5NTk4MDk0ODgwNA=="} +{"index":3220,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2010.11982","reference_contexts":"Prior work proposes better ways of formulating the problem (Zhao et al., 2021; Holtzman et al., 2021; Min et al., 2021a) , better ways of choosing labeled examples for the demonstrations (Liu et al., 2021; Lu et al., 2021; Rubin et al., 2021) , meta-training with an explicit in-context learning objective (Chen et al., 2021; Min et al., 2021b) , and learning to follow instructions as a variant of in-context learning (Mishra et al., 2021b; Efrat and Levy, 2020; Wei et al., 2022a; Sanh et al., 2022)","reference_id":"MTczNjUwOTEyMDg0NTM1NTM1ODI="} +{"index":3221,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2010.11982","reference_contexts":"Prior work has found it promising to train the model that reads the natural language description of the task (called instructions) and performs a new task at inference (Mishra et al., 2021b; Efrat and Levy, 2020; Wei et al., 2022a; Sanh et al., 2022)","reference_id":"Njg2MjMxMTg0OTMwNDQyMzkyNQ=="} +{"index":3222,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2101.00027","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3223,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2101.06804","reference_contexts":"Prior work proposes better ways of formulating the problem (Zhao et al., 2021; Holtzman et al., 2021; Min et al., 2021a) , better ways of choosing labeled examples for the demonstrations (Liu et al., 2021; Lu et al., 2021; Rubin et al., 2021) , meta-training with an explicit in-context learning objective (Chen et al., 2021; Min et al., 2021b) , and learning to follow instructions as a variant of in-context learning (Mishra et al., 2021b; Efrat and Levy, 2020; Wei et al., 2022a; Sanh et al., 2022)","reference_id":"MTczNjUwOTEyMDg0NTM1NTM1ODI="} +{"index":3224,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2101.06804","reference_contexts":"Despite incontext learning consistently outperforming zeroshot inference on a wide range of tasks (Zhao et al., 2021; Liu et al., 2021) , there is little understanding of how it works and which aspects of the demonstrations contribute to end task performance.","reference_id":"NzAyNjI2MjQyNTgxNjgzNDE1Nw=="} +{"index":3225,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2101.06804","reference_contexts":"First, using the demonstrations with gold labels significantly improves the performance over no demonstrations, 5 as it has been consistently found in much of prior work (Brown et al., 2020; Zhao et al., 2021; Liu et al., 2021)","reference_id":"MTMwMjgxMDE2Nzk3Nzc4MDU0Njg="} +{"index":3226,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"1907.11692","reference_contexts":"Large language models have been key to strong performance in a wide range of downstream tasks (Devlin et al., 2019; Radford et al., 2019; Liu et al., 2019; Raffel et al., 2020; Lewis et al., 2020)","reference_id":"OTU4MDE4Nzk0NjU2ODUzNjMyNA=="} +{"index":3227,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2106.13353","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3228,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2104.08786","reference_contexts":"At the same time, some work reports brittleness and oversensitivity for in-context learning (Lu et al., 2021; Zhao et al., 2021; Mishra et al., 2021a) .","reference_id":"MTQ2Nzc2NDczODIyODU4NjY1Mjk="} +{"index":3229,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2104.08786","reference_contexts":"Prior work proposes better ways of formulating the problem (Zhao et al., 2021; Holtzman et al., 2021; Min et al., 2021a) , better ways of choosing labeled examples for the demonstrations (Liu et al., 2021; Lu et al., 2021; Rubin et al., 2021) , meta-training with an explicit in-context learning objective (Chen et al., 2021; Min et al., 2021b) , and learning to follow instructions as a variant of in-context learning (Mishra et al., 2021b; Efrat and Levy, 2020; Wei et al., 2022a; Sanh et al., 2022)","reference_id":"MTczNjUwOTEyMDg0NTM1NTM1ODI="} +{"index":3230,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2209.07686","reference_contexts":"Since the first version of our paper, Madaan and Yazdanbakhsh (2022) conducted a similar analysis with the chain of thought prompting (Wei et al., 2022b) which generates a rationale to perform complex tasks such as math problems","reference_id":"NDY2MDQzOTAzMTQ0NTE5ODM5OA=="} +{"index":3231,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2108.04106","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3232,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2109.07830","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3233,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2104.08773","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3234,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2202.07206","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3235,"citer_arxiv_id":"2202.12837","cited_arxiv_id":"2201.11903","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3236,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2303.08774","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3237,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2304.08465","reference_contexts":"We evaluate our pipeline against several baselines, including image inversion methods like Null-text Inversion [42] and PnP Inversion [27] ; inpainting methods such as BrushNet [28] and PowerPaint [76] ; and zero-shot image editing methods like InfEdit [69] and MasaCtrl [12]","reference_id":"MTE2NjUzMzE0MjU5NDEzOTQ5NTY="} +{"index":3238,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2403.03206","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3239,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2302.10893","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3240,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2208.01626","reference_contexts":"Next, we use prompt-to-prompt (P2P) [24] control with off-the-shelf T2I diffusion models to produce source and target (edited) images from captions and editing instructions","reference_id":"OTIwNjk3OTIwOTc5NTk0OTEyMA=="} +{"index":3241,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2208.01626","reference_contexts":"The model (Middle) For free-form editing, we use the collected images as anchors, and invoke regular diffusion followed by prompt-to-prompt (P2P) [24] control to produce source and target images; (Bottom) For region-based editing, we first produce an editing region based on the instruction, then invoke a modified inpainting diffusion pipeline [44, 5] to produce the images.","reference_id":"MTYxNjc3NzMxMTU1NjUwOTUyNjk="} +{"index":3242,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2208.01626","reference_contexts":"Due to the challenges of obtaining massive instruction-based image editing data, previous research has widely adopted text-to-image (T2I) models [49, 48, 50, 51, 56, 24] to produce both the source images and target (edited) images","reference_id":"NzUzNDUwNDYwOTM4NzQ4MTk5"} +{"index":3243,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2208.01626","reference_contexts":"• Null Text Inversion: inverts the source image with DDIM [63] trajectory and then performs editing during the denoising process with text-image cross-attention control [24] .","reference_id":"MzA1NzgwNDcwMzUwMzU2NDUwOQ=="} +{"index":3244,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2208.01626","reference_contexts":"After producing I s , we invoke prompt-to-prompt (P2P) control [24] with the target caption I t to produce the target image I t on the same z T from real image anchor I ⋆","reference_id":"MTAwOTE2MjkyNTgzMDA1ODg0NDA="} +{"index":3245,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2208.01626","reference_contexts":"For automatically annotated datasets, InstructPix2Pix utilizes prompt-to-prompt [24] to generate data pairs based on the captions form LAION-Aesthetics [59] and edited captions generated by GPT-3 [11] , but biases in the generative model and insufficient information in web-crawled captions make the generated image samples fail to represent the editing instructions, as shown in the Table 1","reference_id":"MTY2ODU3MjUyNjI2NTg4NTE4Njc="} +{"index":3246,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2208.01626","reference_contexts":"Prompt-to-Prompt [24] injecting the input caption attention maps to the target ones","reference_id":"MTA2MDk3MDUzNDg3NTUwMzU2NTQ="} +{"index":3247,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2308.00352","reference_contexts":"Large language models have demonstrated remarkable abilities in various areas, such as agents [18, 25, 61, 35] and tool uage [57]","reference_id":"MjI5MDcxMjUxMDM2MDM0NTIyMA=="} +{"index":3248,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2404.09990","reference_contexts":"HQ-Edit [26] 197,350 6","reference_id":"MTA0NTUwMjYwNDYzNzk3NDI2NTU="} +{"index":3249,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2404.09990","reference_contexts":"HQ-Edit [26] uses advanced models like GPT-4 [2, 1] and DALL-E 3 [6] for generating image editing pairs, but may fail to preserve fine-grained details and realism in the target image.","reference_id":"MTcxNDk3MzcxMTMyNDM4NDg2MTk="} +{"index":3250,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2310.01506","reference_contexts":"We evaluate our pipeline against several baselines, including image inversion methods like Null-text Inversion [42] and PnP Inversion [27] ; inpainting methods such as BrushNet [28] and PowerPaint [76] ; and zero-shot image editing methods like InfEdit [69] and MasaCtrl [12]","reference_id":"MTE2NjUzMzE0MjU5NDEzOTQ5NTY="} +{"index":3251,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2210.09276","reference_contexts":"Imagic [29] supports complex textual instructions via text embedding optimization and model fine-tuning","reference_id":"ODc4NDY3Njc3Nzg2OTA4NzQ5Mg=="} +{"index":3252,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2210.09276","reference_contexts":"Powerful Large-scale diffusion models have significantly facilitated text-based image editing [29, 55, 33, 14, 40]","reference_id":"MTg0MzUzOTYyNDc0NDIzMjc2MjM="} +{"index":3253,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2304.02643","reference_contexts":"Subsequently, we perform SAM [30] on these bounding boxes to derive fine-grained object masks, expanding them to create contour masks","reference_id":"NTM4ODY3OTgxMjk2MzI3NjYwMg=="} +{"index":3254,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2304.02643","reference_contexts":"Then we employ GroundingDINO [38] and SAM [31] to obtain bounding boxes and fine-grained mask of this target object","reference_id":"MTAxMTM5MjYxOTk4NDc3NjY5NTg="} +{"index":3255,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2303.05499","reference_contexts":"Then we employ GroundingDINO [38] and SAM [31] to obtain bounding boxes and fine-grained mask of this target object","reference_id":"MTAxMTM5MjYxOTk4NDc3NjY5NTg="} +{"index":3256,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2303.05499","reference_contexts":"To generate region-based edited images, we use the Grounding DINO [38] to obtain bounding boxes of editing areas in real images, serving as coarse-grained masks","reference_id":"ODYyNTU3NDcwNzYyOTQ3MTIyNw=="} +{"index":3257,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2307.11410","reference_contexts":"Powerful Large-scale diffusion models have significantly facilitated text-based image editing [29, 55, 33, 14, 40]","reference_id":"MTg0MzUzOTYyNDc0NDIzMjc2MjM="} +{"index":3258,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2204.06125","reference_contexts":"MagicBrush [71] manually annotated by hearing online labors using DALL-E 2 [50] for crafting dataset, also limited in dataset size","reference_id":"Mzk1NTcxMzM2MjcxNTY0NjE4OA=="} +{"index":3259,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2204.06125","reference_contexts":"It provides annotator defined instructions and editing masks, as well as the ground truth images generated by DALLE-2 [50] for evaluation, allowing for more effective metric assessment of the Ours w/ ULTRAEDIT MagicBrush [71] InstructPix2Pix [10] TrueSkill score 25.5 ± 0.8 23.7 ± 0.7 22.6 ± 0.7","reference_id":"ODA1ODgzMjQxODU0NTczMDM2NA=="} +{"index":3260,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2204.06125","reference_contexts":"Due to the challenges of obtaining massive instruction-based image editing data, previous research has widely adopted text-to-image (T2I) models [49, 48, 50, 51, 56, 24] to produce both the source images and target (edited) images","reference_id":"NzUzNDUwNDYwOTM4NzQ4MTk5"} +{"index":3261,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"Emu Edit [60] , trained on 10 million proprietary multi-task data, demonstrates state-of-the-art performance.","reference_id":"NDI0MjQ2NTU2MDI0NDc3ODEwMw=="} +{"index":3262,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"Note: As also reported in [60] , we found the MagicBrush benchmark itself introduces biases towards its training set, leading to unfairly high results of models trained on its training set, i.e","reference_id":"MTczMTU4MTI3MTA2MDIxMDU2Nzc="} +{"index":3263,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"The instruction-based image editing methods include: InstructPix2Pix [10] , HIVE [73] , MagicBrush [71] , and Emu Edit [60]","reference_id":"OTEzNTc4MDI4NzUwOTE1NDE5Mg=="} +{"index":3264,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"Additionally, we compared the performance of various models trained using our dataset with Emu Edit [60] on the Emu test benchmark","reference_id":"NzEwNzc5ODgxMzM4OTY1OTgyOA=="} +{"index":3265,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"We follow the settings of Emu Edit [60] to train a diffusion model using various scales and types of data from our dataset, then evaluate the trained model across multiple benchmarks to assess the effectiveness of our dataset in advancing its image editing capabilities","reference_id":"MjAxMTAyMzUyODUyNjY1NjIyNQ=="} +{"index":3266,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"The instruction-guided image editing methods include: InstructPix2Pix [10] , HIVE [73] , MagicBrush [71] , and Emu Edit [60]","reference_id":"MTE0ODUwNTc5ODE2NzY4Njg0MTc="} +{"index":3267,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"However, instead of synthesizing all images using T2I models as in prior works [10, 60] , ULTRAEDIT adopts real images as anchors to mitigate the biases within these T2I models","reference_id":"MTUxNjg4NjIzOTU3MDk0Nzc0MTQ="} +{"index":3268,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"Consequently, consistent with the Emu Edit [60] , we utilize the L1 distance, CLIP image similarity and DINO similarity between the source images and edited images to measure the the model's ability of preserving elements from the source image","reference_id":"OTI3ODQ0ODAxMzEyOTAxOTM0Nw=="} +{"index":3269,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2311.10089","reference_contexts":"Compared to several prior works on the same front [10, 60, 71] , we venture into tackling some of their drawbacks, which prevent the model from robustly interpreting and executing instructions","reference_id":"OTM2NTI0MDc5NTk5ODAwOTQ5Mw=="} +{"index":3270,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2212.06909","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3271,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2312.04965","reference_contexts":"We evaluate our pipeline against several baselines, including image inversion methods like Null-text Inversion [42] and PnP Inversion [27] ; inpainting methods such as BrushNet [28] and PowerPaint [76] ; and zero-shot image editing methods like InfEdit [69] and MasaCtrl [12]","reference_id":"MTE2NjUzMzE0MjU5NDEzOTQ5NTY="} +{"index":3272,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2303.09618","reference_contexts":"Instructpix2pix [10] and HIVE [73] both are trained on generated editing data to handle user-written instructions during inference","reference_id":"MTE3NTc2MjI1NjU5NzI0MTkxNDc="} +{"index":3273,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2303.09618","reference_contexts":"The instruction-guided image editing methods include: InstructPix2Pix [10] , HIVE [73] , MagicBrush [71] , and Emu Edit [60]","reference_id":"MTE0ODUwNTc5ODE2NzY4Njg0MTc="} +{"index":3274,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2303.09618","reference_contexts":"The instruction-based image editing methods include: InstructPix2Pix [10] , HIVE [73] , MagicBrush [71] , and Emu Edit [60]","reference_id":"OTEzNTc4MDI4NzUwOTE1NDE5Mg=="} +{"index":3275,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2306.03514","reference_contexts":"Given an imageinstruction pair ⟨I ⋆ , T e ⟩, we first detect all objects within I ⋆ using recognize-anything [74] and prompt LLM with the object list , soruce caption T s , target captionT t and editing instruction T e to identify the object to be edited","reference_id":"NjM4NzMwOTQ5MzI3MDk3NDQzOA=="} +{"index":3276,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2306.03514","reference_contexts":"For generating region-based image editing data, we first employ the recognize-anything [74] to identify objects in the source images","reference_id":"MTE2MTA0NzU0NTEwNzY5MDk3ODQ="} +{"index":3277,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2309.07915","reference_contexts":"Like previous works [34, 13, 71, 75] , we gathered source data from various public data sources with image caption data, as illustrated in Table 7 , which includes diverse images with either manually annotated captions or detailed captions generated by advanced image captioning models [13, 58] .","reference_id":"MjkwOTYyNzc0MzU2NDMwMjUyMg=="} +{"index":3278,"citer_arxiv_id":"2407.05282","cited_arxiv_id":"2312.03594","reference_contexts":"We evaluate our pipeline against several baselines, including image inversion methods like Null-text Inversion [42] and PnP Inversion [27] ; inpainting methods such as BrushNet [28] and PowerPaint [76] ; and zero-shot image editing methods like InfEdit [69] and MasaCtrl [12]","reference_id":"MTE2NjUzMzE0MjU5NDEzOTQ5NTY="} +{"index":3279,"citer_arxiv_id":"1808.09121","cited_arxiv_id":"1301.3781","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3280,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1412.6071","reference_contexts":"It is only outperformed by the fractional max-pooling approach (Graham, 2015) which uses a much larger network (on the order of 50M parameters).","reference_id":"MTAyNjEzMjYwODI5NTMwOTg3MzU="} +{"index":3281,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1412.6071","reference_contexts":"After performing our experiments we became aware of recent results by Graham (2015) who report a new state of the art on CIFAR-10/100 with data augmentation","reference_id":"MTgxODk5MTIzNzAwNTYyNjA5NjE="} +{"index":3282,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1412.6071","reference_contexts":"The complete model architecture for the large All-CNN derived from the spatially sparse network of Benjamin Graham (see Graham (2015) for an explanation) is givenin Table 5","reference_id":"ODY1OTA5OTc4NjM1NDczMDM2NA=="} +{"index":3283,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1408.5093","reference_contexts":"We performed all experiments using the Caffe (Jia et al., 2014) framework.","reference_id":"MTUzNDAwMjgzNDM5MDQ3MTg5Nw=="} +{"index":3284,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1409.1556","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3285,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1312.6034","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3286,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1409.4842","reference_contexts":"The second difference of the network model we consider to standard CNNs is that -similar to models recently used for achieving state-of-the-art performance in the ILSVRC-2012 competition (Simonyan & Zisserman, 2014; Szegedy et al., 2014) -we make use of small convolutional layers with k < 5 which can greatly reduce the number of parameters in a network and thus serve as a form of regularization","reference_id":"NTM2MjIyMzgyMTUxNzMxNTEwOQ=="} +{"index":3287,"citer_arxiv_id":"1412.6806","cited_arxiv_id":"1409.4842","reference_contexts":"Most notably the top entries in the 2014 ImageNet challenge deviated from the standard design principles by either introducing multiple convolutions in between pooling layers (Simonyan & Zisserman, 2014) or by building heterogeneous modules performing convolutions and pooling at multiple scales in each layer (Szegedy et al., 2014)","reference_id":"MTA0Mjk1MDE0NjEzMTI2MjY1MjU="} +{"index":3288,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2206.02779","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3289,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2211.14305","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3290,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2108.02938","reference_contexts":"Avrahami et al., 2022c; Choi et al., 2021; Mokady et al., 2022; Couairon et al., 2022; Kong et al., 2023; Kwon et al., 2022) or by performing lightweight model finetuning (Ruiz et al., 2022; Kawar et al., 2022; Kim et al., 2022; Valevski et al., 2022)","reference_id":"OTQyODcyNzM1NzI3Mzk0Mjc0Mg=="} +{"index":3291,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2210.11427","reference_contexts":"Avrahami et al., 2022c; Choi et al., 2021; Mokady et al., 2022; Couairon et al., 2022; Kong et al., 2023; Kwon et al., 2022) or by performing lightweight model finetuning (Ruiz et al., 2022; Kawar et al., 2022; Kim et al., 2022; Valevski et al., 2022)","reference_id":"OTQyODcyNzM1NzI3Mzk0Mjc0Mg=="} +{"index":3292,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2208.01626","reference_contexts":"Previously, these methods have concentrated on specific tasks and designed a tailored methodology (e.g., replacing objects in an image, manipulating style, or controlling layout (Tumanyan et al., 2022; Hertz et al., 2022; Avrahami et al., 2022a) ).","reference_id":"MTUxMjAzMDcwOTU5MTc4MjExMzY="} +{"index":3293,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2208.01626","reference_contexts":"Recent works (Tumanyan et al., 2022; Hertz et al., 2022) rely on architectural properties and insights about the internal features of the pretrained model, and tailor image editing techniques accordingly","reference_id":"MTE5OTYwMTQ4NzM2NTY5NTI4OTA="} +{"index":3294,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2208.01626","reference_contexts":"(ii) On the other side of the spectrum, numerous methods proposed to implicitly control the generated content by manipulating the generation process of a pre-trained model (Kwon & Ye, 2022; Meng et al., 2021; Tumanyan et al., 2022; Hertz et al., 2022; Figure 2","reference_id":"NTYzNTI2MTc5Mjk3NjEwMzcxMQ=="} +{"index":3295,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2207.12598","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3296,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2210.02303","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3297,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2210.09276","reference_contexts":"Avrahami et al., 2022c; Choi et al., 2021; Mokady et al., 2022; Couairon et al., 2022; Kong et al., 2023; Kwon et al., 2022) or by performing lightweight model finetuning (Ruiz et al., 2022; Kawar et al., 2022; Kim et al., 2022; Valevski et al., 2022)","reference_id":"OTQyODcyNzM1NzI3Mzk0Mjc0Mg=="} +{"index":3298,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2209.15264","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3299,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2210.10960","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3300,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2202.09778","reference_contexts":"The number of denoising steps, and the parameterization of the transformation varies among different works (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020; Lu et al., 2022a; b; Liu et al., 2022)","reference_id":"NTQ2MTg0ODYwNTc5MDkyNDQ2Nw=="} +{"index":3301,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2206.00927","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3302,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2211.01095","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3303,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2108.01073","reference_contexts":"(ii) On the other side of the spectrum, numerous methods proposed to implicitly control the generated content by manipulating the generation process of a pre-trained model (Kwon & Ye, 2022; Meng et al., 2021; Tumanyan et al., 2022; Hertz et al., 2022; Figure 2","reference_id":"NTYzNTI2MTc5Mjk3NjEwMzcxMQ=="} +{"index":3304,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2211.09794","reference_contexts":"Avrahami et al., 2022c; Choi et al., 2021; Mokady et al., 2022; Couairon et al., 2022; Kong et al., 2023; Kwon et al., 2022) or by performing lightweight model finetuning (Ruiz et al., 2022; Kawar et al., 2022; Kim et al., 2022; Valevski et al., 2022)","reference_id":"OTQyODcyNzM1NzI3Mzk0Mjc0Mg=="} +{"index":3305,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2112.10741","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3306,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2204.06125","reference_contexts":"These models have been successfully used in various domains, including images (Dhariwal & Nichol, 2021; Nichol & Dhariwal, 2021; Saharia et al., 2022b; Ramesh et al., 2022; Rombach et al., 2022) , video (Ho et al., 2022; Singer et al., 2022) , 3D scenes (Múller et al., 2022) , and motion sequences (Yuan et al., 2022; Tevet et al., 2022) .","reference_id":"OTMwOTExNDUxMTk1MTIzNzYyNQ=="} +{"index":3307,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2204.06125","reference_contexts":"Text-to-image generative models have emerged as a \"disruptive technology\", demonstrating unprecedented capabilities in synthesizing high-quality and diverse images from text prompts, where diffusion models are currently established as state-of-the-art (Saharia et al., 2022b; Ramesh et al., 2022; Rombach et al., 2022; Croitoru et al., 2022)","reference_id":"MTU4OTYwMjU2MzI5NjcyMDM1NDQ="} +{"index":3308,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2204.06125","reference_contexts":"The most prominent example of conditional diffusion models is recent text-to-image diffusion models, which have demonstrated groundbreaking synthesis capabilities (Nichol et al., 2021; Saharia et al., 2022b; Ramesh et al., 2022; Nichol et al., 2021; Rombach et al., 2022; Sheynin et al., 2022)","reference_id":"MTAzNzQ1OTg4Mzg5MDY1OTEwNzQ="} +{"index":3309,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2208.12242","reference_contexts":"Avrahami et al., 2022c; Choi et al., 2021; Mokady et al., 2022; Couairon et al., 2022; Kong et al., 2023; Kwon et al., 2022) or by performing lightweight model finetuning (Ruiz et al., 2022; Kawar et al., 2022; Kim et al., 2022; Valevski et al., 2022)","reference_id":"OTQyODcyNzM1NzI3Mzk0Mjc0Mg=="} +{"index":3310,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2205.11487","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3311,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2210.08402","reference_contexts":"To further assess the quality of our results, we evaluated two CLIP-based scores: (i) text-image CLIP score (Radford et al., 2021) measured by the cosine similarity between the text prompt and the image embeddings, and (ii) CLIP aesthetic (Schuhmann et al., 2022) measured by a linear estimator on top of CLIP predicting the aesthetic quality of the images.","reference_id":"NDMzMzEwODgxNzQyMzExMTgyNg=="} +{"index":3312,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2204.02849","reference_contexts":"The most prominent example of conditional diffusion models is recent text-to-image diffusion models, which have demonstrated groundbreaking synthesis capabilities (Nichol et al., 2021; Saharia et al., 2022b; Ramesh et al., 2022; Nichol et al., 2021; Rombach et al., 2022; Sheynin et al., 2022)","reference_id":"MTAzNzQ1OTg4Mzg5MDY1OTEwNzQ="} +{"index":3313,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2010.02502","reference_contexts":"The number of denoising steps, and the parameterization of the transformation varies among different works (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020; Lu et al., 2022a; b; Liu et al., 2022)","reference_id":"NTQ2MTg0ODYwNTc5MDkyNDQ2Nw=="} +{"index":3314,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2209.14916","reference_contexts":"These models have been successfully used in various domains, including images (Dhariwal & Nichol, 2021; Nichol & Dhariwal, 2021; Saharia et al., 2022b; Ramesh et al., 2022; Rombach et al., 2022) , video (Ho et al., 2022; Singer et al., 2022) , 3D scenes (Múller et al., 2022) , and motion sequences (Yuan et al., 2022; Tevet et al., 2022) .","reference_id":"OTMwOTExNDUxMTk1MTIzNzYyNQ=="} +{"index":3315,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2211.12572","reference_contexts":"Previously, these methods have concentrated on specific tasks and designed a tailored methodology (e.g., replacing objects in an image, manipulating style, or controlling layout (Tumanyan et al., 2022; Hertz et al., 2022; Avrahami et al., 2022a) ).","reference_id":"MTUxMjAzMDcwOTU5MTc4MjExMzY="} +{"index":3316,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2211.12572","reference_contexts":"Recent works (Tumanyan et al., 2022; Hertz et al., 2022) rely on architectural properties and insights about the internal features of the pretrained model, and tailor image editing techniques accordingly","reference_id":"MTE5OTYwMTQ4NzM2NTY5NTI4OTA="} +{"index":3317,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2211.12572","reference_contexts":"(ii) On the other side of the spectrum, numerous methods proposed to implicitly control the generated content by manipulating the generation process of a pre-trained model (Kwon & Ye, 2022; Meng et al., 2021; Tumanyan et al., 2022; Hertz et al., 2022; Figure 2","reference_id":"NTYzNTI2MTc5Mjk3NjEwMzcxMQ=="} +{"index":3318,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2210.09477","reference_contexts":"Avrahami et al., 2022c; Choi et al., 2021; Mokady et al., 2022; Couairon et al., 2022; Kong et al., 2023; Kwon et al., 2022) or by performing lightweight model finetuning (Ruiz et al., 2022; Kawar et al., 2022; Kim et al., 2022; Valevski et al., 2022)","reference_id":"OTQyODcyNzM1NzI3Mzk0Mjc0Mg=="} +{"index":3319,"citer_arxiv_id":"2302.08113","cited_arxiv_id":"2212.02500","reference_contexts":"These models have been successfully used in various domains, including images (Dhariwal & Nichol, 2021; Nichol & Dhariwal, 2021; Saharia et al., 2022b; Ramesh et al., 2022; Rombach et al., 2022) , video (Ho et al., 2022; Singer et al., 2022) , 3D scenes (Múller et al., 2022) , and motion sequences (Yuan et al., 2022; Tevet et al., 2022) .","reference_id":"OTMwOTExNDUxMTk1MTIzNzYyNQ=="} +{"index":3320,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2103.10951","reference_contexts":"Many works have already used CLIP successfully for computer vision tasks that require the understanding of some auxiliary text, such as generating or editing an image based on a natural language condition [5, 14, 28]","reference_id":"MTA1OTE5NzY1NTk1MTQyOTUzMjQ="} +{"index":3321,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1504.00325","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3322,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1411.5654","reference_contexts":"Early works have used LSTM variants [8, 38, 39] , while recent works [16, 26] adopted the improved transformer architecture [36]","reference_id":"Nzg0Njk0NTEyMTY1MjUyMTIwMA=="} +{"index":3323,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1810.04805","reference_contexts":"larly to BERT [11]","reference_id":"MTYzMTk0MTg1OTcwOTI4MDQ1ODM="} +{"index":3324,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1810.04805","reference_contexts":"The entire model is then pre-trained to perform prediction utilizing the BERT [11] architecture","reference_id":"MTc0MDg2ODUzMzY5NjYxOTUyOTY="} +{"index":3325,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1810.04805","reference_contexts":"Built upon the transformer, one of the most notable works is BERT [11] , demonstrating the dominance of the newly introduced paradigm","reference_id":"Mzg5NTUzMTc0MTk1NjkzMzA1OA=="} +{"index":3326,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2010.11929","reference_contexts":"Moreover, recent models apply self-attention [16, 43] or use an expressive visual Transformer [12] as an encoder [23]","reference_id":"Mzg2NzEyOTkzNzU5MDg5Njc3Ng=="} +{"index":3327,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2108.00946","reference_contexts":"Many works have already used CLIP successfully for computer vision tasks that require the understanding of some auxiliary text, such as generating or editing an image based on a natural language condition [5, 14, 28]","reference_id":"MTA1OTE5NzY1NTk1MTQyOTUzMjQ="} +{"index":3328,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1412.6980","reference_contexts":"For optimization, we use AdamW [18] with weight decay fix as introduced by Loshchilov et al","reference_id":"MTYyODg4NTg0NTUwNDE4ODE5NDY="} +{"index":3329,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2101.00190","reference_contexts":"[20] , which demonstrates the ability to efficiently adapt a language model for new tasks by concatenating a learned prefix","reference_id":"ODk2NjEwMTE0NzkyNjAzMDg5OQ=="} +{"index":3330,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2101.00190","reference_contexts":"Li and Liang [20] showed that increasing the size of the prefix length, up to a certain value, improves the performance of the model in an underlying task","reference_id":"MTMyNDg3NzI5NDQ0MTM3MDk0OTg="} +{"index":3331,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2101.00190","reference_contexts":"Our attempt to adjust a frozen language model is inspired by the work of Li and Liang [20]","reference_id":"MTY1NTQ2MTAzMzEwODc0NDE0NjM="} +{"index":3332,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1711.05101","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3333,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1908.02265","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3334,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2103.00020","reference_contexts":"To extract visual information from an image x i , we use the visual encoder of a pre-trained CLIP [29] model","reference_id":"ODQwOTAxNTQ0NjAwMzM2MTYwNQ=="} +{"index":3335,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2103.00020","reference_contexts":"[29] presented a novel approach, known as CLIP, to jointly represent images and text descriptions","reference_id":"MTI1Mzk0MzQxNDMxNzgxMTU4OTg="} +{"index":3336,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2103.00020","reference_contexts":"[29]","reference_id":"MTQxODM1Mjg2NzM2NTIzOTM5ODQ="} +{"index":3337,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1908.07490","reference_contexts":"Most close to ours, are works that employ vision-andlanguage pre-training to create a shared latent space of both vision and text [19, 25, 35, 46, 47]","reference_id":"NTc5OTk1OTI0NjY0MjUyMjg4Mg=="} +{"index":3338,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"1908.07490","reference_contexts":"Many approaches have been proposed for image captioning [4, 9, 13, 19, 34, 35, 42, 44, 47]","reference_id":"NzI5OTg2OTYwODQwOTE1NjI5Mw=="} +{"index":3339,"citer_arxiv_id":"2111.09734","cited_arxiv_id":"2108.10904","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3340,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2305.10403","reference_contexts":"In Figure 2 , we compare the results of our evaluation with the closest sized PaLM (Chowdhery et al., 2022) and PaLM 2 models (Anil et al., 2023) .","reference_id":"MTM0MTM1ODcyMjQyNTAyMTY0MzI="} +{"index":3341,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2305.10403","reference_contexts":"As with our recent Gemini release, these are informed by prior academic literature on language model risks (Weidinger et al., 2021) , findings from similar prior exercises conducted across the industry (Anil et al., 2023) , ongoing engagement with experts internally and externally, and unstructured attempts to discover new model vulnerabilities.","reference_id":"NzA0ODQxMzI1NTg5MDE1MzQ2NA=="} +{"index":3342,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2305.10403","reference_contexts":"We focus on discoverable memorization, which serves as a reasonable upper-bound on the memorization of a model (Nasr et al., 2023) and has been the common definition used in several studies (Anil et al., 2023; Carlini et al., 2022; Kudugunta et al., 2023) .","reference_id":"MTU1OTc4MTE1NTk1MzU5NTg4NTI="} +{"index":3343,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2108.07732","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021)","reference_id":"ODI1ODM3NDY0ODI3OTc3MzM3Mg=="} +{"index":3344,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2108.07732","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3345,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1911.11641","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3346,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2202.07646","reference_contexts":"We focus on discoverable memorization, which serves as a reasonable upper-bound on the memorization of a model (Nasr et al., 2023) and has been the common definition used in several studies (Anil et al., 2023; Carlini et al., 2022; Kudugunta et al., 2023) .","reference_id":"MTU1OTc4MTE1NTk1MzU5NTg4NTI="} +{"index":3347,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2107.03374","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021)","reference_id":"ODI1ODM3NDY0ODI3OTc3MzM3Mg=="} +{"index":3348,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2107.03374","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3349,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1905.10044","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021)","reference_id":"ODI1ODM3NDY0ODI3OTc3MzM3Mg=="} +{"index":3350,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1905.10044","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3351,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2110.14168","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021)","reference_id":"ODI1ODM3NDY0ODI3OTc3MzM3Mg=="} +{"index":3352,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2110.14168","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3353,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1810.04805","reference_contexts":"Gemma also builds on Google's long history of open models and ecosystems, including Word2Vec (Mikolov et al., 2013) , the Transformer (Vaswani et al., 2017) , BERT (Devlin et al., 2018) , and T5 (Raffel et al., 2019) and T5X (Roberts et al., 2022) .","reference_id":"MTMzODY2NTMwNzQ4ODgwNzM5NzI="} +{"index":3354,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2009.03300","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021)","reference_id":"ODI1ODM3NDY0ODI3OTc3MzM3Mg=="} +{"index":3355,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2009.03300","reference_contexts":"On MMLU (Hendrycks et al., 2020) , Gemma 7B outperforms all OSS alternatives at the same or smaller scale; it also outperforms several larger models, including LLaMA2 13B","reference_id":"ODM0MDIwODc4OTk4NTEwNTM="} +{"index":3356,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2210.17546","reference_contexts":"However, to better capture potential paraphrased memorizations, we include approximate memorization (Ippolito et al., 2022)","reference_id":"NjU5MjEyMjg3ODgxMjQwNDUx"} +{"index":3357,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1705.03551","reference_contexts":"ing comprehension (Joshi et al., 2017) , and more.","reference_id":"MjAwOTIwMTI2MzgyMDczMjc3Mw=="} +{"index":3358,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1301.3781","reference_contexts":"Gemma also builds on Google's long history of open models and ecosystems, including Word2Vec (Mikolov et al., 2013) , the Transformer (Vaswani et al., 2017) , BERT (Devlin et al., 2018) , and T5 (Raffel et al., 2019) and T5X (Roberts et al., 2022) .","reference_id":"MTMzODY2NTMwNzQ4ODgwNzM5NzI="} +{"index":3359,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2311.17035","reference_contexts":"Recent work has shown that aligned models may be vulnerable to new adversarial attacks that can bypass alignment (Nasr et al., 2023)","reference_id":"MTYwMjgyMTQ1NTc2NzY1NDYzMjg="} +{"index":3360,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1606.06031","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3361,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1910.10683","reference_contexts":"Gemma also builds on Google's long history of open models and ecosystems, including Word2Vec (Mikolov et al., 2013) , the Transformer (Vaswani et al., 2017) , BERT (Devlin et al., 2018) , and T5 (Raffel et al., 2019) and T5X (Roberts et al., 2022) .","reference_id":"MTMzODY2NTMwNzQ4ODgwNzM5NzI="} +{"index":3362,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1907.10641","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021)","reference_id":"ODI1ODM3NDY0ODI3OTc3MzM3Mg=="} +{"index":3363,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1907.10641","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3364,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1904.09728","reference_contexts":"We measure Gemma models' performance on domains including physical reasoning (Bisk et al., 2019) , social reasoning (Sap et al., 2019) , question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , coding (Austin et al., 2021; Chen et al., 2021) , mathematics (Cobbe et al., 2021) , commonsense reasoning (Sakaguchi et al., 2019) , language modeling (Paperno et al., 2016) , read-Model Safety Instr","reference_id":"MTc2OTYyODYyNDUwOTU3NDE5MTM="} +{"index":3365,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1911.02150","reference_contexts":"Multi-Query Attention (Shazeer, 2019)","reference_id":"Mjk2MzA0OTYwMjU3NTIzMTAyMA=="} +{"index":3366,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2002.05202","reference_contexts":"GeGLU Activations (Shazeer, 2020)","reference_id":"MTAxNTk5MDk0NzA2NTcwMTA0ODA="} +{"index":3367,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2104.09864","reference_contexts":"(Su et al., 2021)","reference_id":"MTE5NzkzMDA4OTMyMjgxMDA1NjY="} +{"index":3368,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1409.3215","reference_contexts":"Like Gemini, Gemma builds on recent work on sequence models (Sutskever et al., 2014) and transformers (Vaswani et al., 2017) , deep learning methods based on neural networks (LeCun et al., 2015) , and techniques for large-scale training on distributed systems (Barham et al., 2022; Dean et al., 2012; Roberts et al., 2023)","reference_id":"MTE5MTc5ODM5MTU4NTMwOTExNDc="} +{"index":3369,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1706.03762","reference_contexts":"Like Gemini, Gemma builds on recent work on sequence models (Sutskever et al., 2014) and transformers (Vaswani et al., 2017) , deep learning methods based on neural networks (LeCun et al., 2015) , and techniques for large-scale training on distributed systems (Barham et al., 2022; Dean et al., 2012; Roberts et al., 2023)","reference_id":"MTE5MTc5ODM5MTU4NTMwOTExNDc="} +{"index":3370,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1706.03762","reference_contexts":"The Gemma model architecture is based on the transformer decoder (Vaswani et al., 2017)","reference_id":"MTIxMDU1NTQxMjUyOTA5OTUzMjY="} +{"index":3371,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2201.11903","reference_contexts":"Our LM-based judges employ a number of known strategies, such as chain-of-thought prompting (Wei et al., 2022) , rubrics and constitutions (Bai et al., 2022) , to be aligned with human preferences.","reference_id":"NzkwMzY1MjMwMjIwNjQyNjczOA=="} +{"index":3372,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2112.04359","reference_contexts":"In addition to bringing benefits to the AI development ecosystem, we are aware that malicious uses of LLMs, such as the creation of deepfake imagery, AI-generated disinformation, and illegal and disturbing material can cause harm on both an individual and institutional levels (Weidinger et al., 2021)","reference_id":"MTM5MDM1OTc1NjU1OTAzMTg1NzA="} +{"index":3373,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2112.04359","reference_contexts":"As with our recent Gemini release, these are informed by prior academic literature on language model risks (Weidinger et al., 2021) , findings from similar prior exercises conducted across the industry (Anil et al., 2023) , ongoing engagement with experts internally and externally, and unstructured attempts to discover new model vulnerabilities.","reference_id":"NzA0ODQxMzI1NTg5MDE1MzQ2NA=="} +{"index":3374,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"2105.04663","reference_contexts":"This simplifies the development process by enabling a single Python process to orchestrate the entire training run; we also leverage the GSPMD partitioner (Xu et al., 2021) for the training step computation and the MegaScale XLA compiler (XLA, 2019) .","reference_id":"OTA2NzI1NDcxOTg3MDAwNzYzMw=="} +{"index":3375,"citer_arxiv_id":"2403.08295","cited_arxiv_id":"1910.07467","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3376,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"1301.3781","reference_contexts":"These language models have evolved, from learning individual word vectors with single-layer models (1) , to more complex language generation architectures such as recurrent neural networks (2) and most recently transformers (3) (4) (5)","reference_id":"MTYwMzAxMTIxMDA5NDYzNjI0NTM="} +{"index":3377,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"2005.14165","reference_contexts":"We used a similar setup to the original work (4) , in which we present GPT-3 with the following analogy: \"audacious is to boldness as Muslim is to...\" and ask GPT-3 to complete the analogy in a zero-shot setting","reference_id":"MTA1NjEzOTc3OTEwNDg5NDkyODA="} +{"index":3378,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"2005.14165","reference_contexts":"These language models have evolved, from learning individual word vectors with single-layer models (1) , to more complex language generation architectures such as recurrent neural networks (2) and most recently transformers (3) (4) (5)","reference_id":"MTYwMzAxMTIxMDA5NDYzNjI0NTM="} +{"index":3379,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"2005.14165","reference_contexts":"When provided with a single correctly-answered example, that accuracy increases to 59%, better than the average college student, who correctly answers such questions 57% of the time (4, 11)","reference_id":"NTQzNDczOTcyNTY5MDcyNjI5NA=="} +{"index":3380,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"2001.04451","reference_contexts":"These language models have evolved, from learning individual word vectors with single-layer models (1) , to more complex language generation architectures such as recurrent neural networks (2) and most recently transformers (3) (4) (5)","reference_id":"MTYwMzAxMTIxMDA5NDYzNjI0NTM="} +{"index":3381,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"2004.09456","reference_contexts":"As a result, even though the various language models have different architectures, since they are trained on similar corpora of text, they often learn similar biases (6) .","reference_id":"NTkwMDc0ODYzMDA5MzgyOTg4OA=="} +{"index":3382,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"2004.09456","reference_contexts":"Religious bias was considered in (6) , where the authors curated a dataset to measure a range of associations and stereotypes learned by a language model, including examples of anti-Muslim or Islamophobic sentiment.","reference_id":"MTUyODUwODY2ODE1OTU4NzEzMzY="} +{"index":3383,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"1909.01326","reference_contexts":"Previous work has explored the biases that are ultimately learned by language models, with the primary focus being racial and gender stereotypes (7) (8) (9) (10)","reference_id":"MjQzODE1MjczMDYxODgzODg1MQ=="} +{"index":3384,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"1904.03035","reference_contexts":"Previous work has explored the biases that are ultimately learned by language models, with the primary focus being racial and gender stereotypes (7) (8) (9) (10)","reference_id":"MjQzODE1MjczMDYxODgzODg1MQ=="} +{"index":3385,"citer_arxiv_id":"2101.05783","cited_arxiv_id":"1908.07125","reference_contexts":"This is inspired by \"adversarial triggers\" (14) , which are short sequences of words that change language models outputs","reference_id":"NzEzMjExMjk0NTI1NzEzNDQ0MQ=="} +{"index":3386,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1906.00910","reference_contexts":"PIRL-c2x achieves a top-1 accuracy of 67.4%, which is close to the accuracy obtained by AMDIM [4] with a model that has 6× more parameters.","reference_id":"MTA2NTYxNzcwNjY3NjA2ODI2MDY="} +{"index":3387,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1907.02544","reference_contexts":"A variety of work has studied reconstructing images from a small, intermediate representation, e.g., using sparse coding [50] , adversarial training [11, 12, 43] , autoencoders [42, 55, 67] , or probabilistic versions thereof [59] .","reference_id":"MTMwOTI2NjY5NDExODAxNTA1NDg="} +{"index":3388,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"Examples of image transformations used include rotations [18] , affine transformations [31, 49, 57, 76] , and jigsaw transformations [46]","reference_id":"NDE1MzAyOTc5MTcxMzY4NDIzMA=="} +{"index":3389,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"We performed an experiment evaluating the performance of PIRL using the Rotation pretext task [18]","reference_id":"MTIzMzkwODA2NDIzMzQ0OTA0NjA="} +{"index":3390,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"Prior pretext tasks include image colorization [8, 28, 35, 36, 77, 78] , orientation prediction [18] , affine transform prediction [76] , predicting contextual image patches [9], reordering image patches [5, 19, 46, 48] , counting visual primitives [47] , or their combinations [10]","reference_id":"OTM3NzcwNjEwMjk1MjUxOTM5NQ=="} +{"index":3391,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"We contrast our loss function to losses [9, 18, 44, 46, 76 ] that aim to learn image representations v I = φ θ (I) that are covariant to image transformations t ∈ T by minimizing:","reference_id":"MTUwNjU5NDQwNTIwNTA5MTEzNDE="} +{"index":3392,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"Our motivation for this design of g(v I t ) is the desire to remain as close as possible to the covariant pretext task of [18, 19, 46]","reference_id":"MTc4MzIyNzU1MzgyNjE4NTI0NTA="} +{"index":3393,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"ImageNet (Rotation) outperform those trained using the Rotation pretext task of [18]","reference_id":"OTkyMjgwOTY3MzU3Mzg3NjcyMQ=="} +{"index":3394,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"Data pre-processing We base our Rotation implementation on [18, 33]","reference_id":"MTc5NTgxMzkxMDI4MDk4MTg5OTQ="} +{"index":3395,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"The same is true for a range of other pretext tasks that have recently been studied [9, 18, 44, 76]","reference_id":"MTMwNjE0MTQwNjEyMjA5NDE3NTc="} +{"index":3396,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"Many pretext tasks for self-supervised learning [18, 46, 76] involve transforming an image I, computing a representation of the transformed image, and predicting properties of transformation t from that representation","reference_id":"MTMzMjU4NTExMzAxODM3ODYwMzg="} +{"index":3397,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"On the challenging iNaturalist dataset, which has over 8, 000 classes, we obtain a gain of 11% in top-1 accuracy over the prior best result [18]","reference_id":"MTU5OTQ0ODE4MTUxMzEzMDU3"} +{"index":3398,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"To demonstrate that PIRL is more generally applicable, we also experiment with the Rotation pretext task [18] and with a combination of both tasks in Section 4.3","reference_id":"ODMwNjcxNjYyNzQ0NTQ5NzAxOQ=="} +{"index":3399,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1803.07728","reference_contexts":"The set T may contain transformations such as a re-shuffling of patches in the image [46] , image rotations [18] , etc","reference_id":"MjYxNTk0ODY2MjI2NjEyOTIwOA=="} +{"index":3400,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1905.09272","reference_contexts":"Indeed, PIRL even outperforms most self-supervised learn- ers that use much larger models [26, 51] .","reference_id":"MTQyODEyOTI2NjY5MTMxMTk5ODE="} +{"index":3401,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1905.09272","reference_contexts":"This result is a substantial improvement over prior self-supervised approaches that obtain slightly worse performance than fully supervised baselines despite using orders of magnitude more curated training data [19] or much larger backbone models [26]","reference_id":"MjMxNDk4MzUyNzAwMzM0MzIzMw=="} +{"index":3402,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1905.09272","reference_contexts":"We perform semi-supervised image classification experiments on ImageNet following the experimental setup of [26, 72, 75]","reference_id":"MTI4OTMyMDAyNjUyMjkyNTIyMjc="} +{"index":3403,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1905.09272","reference_contexts":"PIRL is related to approaches that learn invariant image Finally, PIRL is also related to approaches that use a contrastive loss [22] in predictive learning [23, 24, 26, 51, 61, 64]","reference_id":"ODc4Njg1ODgzODYyNDMwNDgyMQ=="} +{"index":3404,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1808.06670","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3405,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1901.09005","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3406,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1608.03983","reference_contexts":"We train our models using mini-batch SGD using the cosine learning rate decay [40] scheme with an initial learning rate of 1.2×10 -1 and a final learning rate of 1.2 × 10 -4","reference_id":"ODM3MjY0NzQ3MTY0OTcxMDM4OQ=="} +{"index":3407,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1608.03983","reference_contexts":"We optimize the models using mini-batch SGD with the cosine learning rate decay [40] scheme with an initial learning rate of 1.2 × 10 -1 and a final learning rate of 1.2 × 10 -4 , momentum of 0.9 and a weight decay of 10 -4","reference_id":"NDk2NjgzNjM1NDM3MDU0MjU4Mg=="} +{"index":3408,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1807.03748","reference_contexts":"Indeed, PIRL even outperforms most self-supervised learn- ers that use much larger models [26, 51] .","reference_id":"MTQyODEyOTI2NjY5MTMxMTk5ODE="} +{"index":3409,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1807.03748","reference_contexts":"These prior approaches predict missing parts of the data, e.g., future frames in videos [23, 51] , or operate on multiple views [64]","reference_id":"NjczNTQxNDk5ODEzNjc3Mjc4Nw=="} +{"index":3410,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1807.03748","reference_contexts":"Prior work has found that it is important to use a large number of negatives in the NCE loss of Equation 4 [51, 72]","reference_id":"MTU5MDAyNzM4ODIwNjQ3Mjg1ODY="} +{"index":3411,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1807.03748","reference_contexts":"PIRL is related to approaches that learn invariant image Finally, PIRL is also related to approaches that use a contrastive loss [22] in predictive learning [23, 24, 26, 51, 61, 64]","reference_id":"ODc4Njg1ODgzODYyNDMwNDgyMQ=="} +{"index":3412,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1807.03748","reference_contexts":"In Table 1 , we use the VOC07+12 training split and the VOC07 test set as used in prior work [19, 51]","reference_id":"MTM5MjY0NDU2MTA2NzkzNDE2NDQ="} +{"index":3413,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1906.05743","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3414,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1503.01817","reference_contexts":"To study how much changes in the data distribution im- pact PIRL, we pre-train models on uncurated images from the unlabeled YFCC dataset [63]","reference_id":"MTMyMTAzOTU1Mjk2NjcyMDEzOTQ="} +{"index":3415,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1906.05849","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3416,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1904.12848","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3417,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1905.03670","reference_contexts":"To compare PIRL with larger models, we also performed experiments in which we followed [33, 75] by doubling the number of channels in ResNet-50; see \"PIRL-c2x\" in Figure 2","reference_id":"NjIyMDc5MjUwODM3MjgzMjc4OA=="} +{"index":3418,"citer_arxiv_id":"1912.01991","cited_arxiv_id":"1905.03670","reference_contexts":"We perform semi-supervised image classification experiments on ImageNet following the experimental setup of [26, 72, 75]","reference_id":"MTI4OTMyMDAyNjUyMjkyNTIyMjc="} +{"index":3419,"citer_arxiv_id":"2203.17189","cited_arxiv_id":"1902.08295","reference_contexts":"Previous Google-released systems based on TensorFlow include Tensor2Tensor (Vaswani et al., 2018) , Lingvo (Shen et al., 2019) , and the Mesh TensorFlow (Shazeer et al., 2018) -based T5 (Raffel et al., 2020) .","reference_id":"OTQ1MzI1NTIzNjY2NTYyODg4"} +{"index":3420,"citer_arxiv_id":"2203.17189","cited_arxiv_id":"1909.08053","reference_contexts":"These options correspond to previously described parallelism techniques: 2D parameter partitioning is also known as ZeRO-3 (Rajbhandari et al., 2020) or fully sharded data parallelism; 1D activation partitioning is also known as Megatron (Shoeybi et al., 2019) and is the default in the Mesh TensorFlow Transformer (Shazeer et al., 2018) ; and 2D activation partitioning is the \"fully sharded\" case described in Xu et al","reference_id":"MzI5MTE1MjEyNDk1Mzg0MjI4OQ=="} +{"index":3421,"citer_arxiv_id":"2203.17189","cited_arxiv_id":"2201.11990","reference_contexts":"Comparable projects from other research groups include model libraries such as fairseq (Ott et al., 2019) , large-scale parallelism libraries such as FairScale (Baines et al., 2021) , and libraries that include both kinds of functionality such as DeepSpeed (Rasley et al., 2020) and Megatron (Smith et al., 2022)","reference_id":"MTU2MjY5NDc4NTQyOTY3ODcyMDQ="} +{"index":3422,"citer_arxiv_id":"2203.17189","cited_arxiv_id":"2201.08239","reference_contexts":"We also provide a model configuration (without any checkpoints) for a decoder-only architecture that is compatible with LaMDA (Thoppilan et al., 2022) .","reference_id":"MzAxMjI1MzM1NDk3MjE0NTAwNQ=="} +{"index":3423,"citer_arxiv_id":"2203.17189","cited_arxiv_id":"1803.07416","reference_contexts":"Previous Google-released systems based on TensorFlow include Tensor2Tensor (Vaswani et al., 2018) , Lingvo (Shen et al., 2019) , and the Mesh TensorFlow (Shazeer et al., 2018) -based T5 (Raffel et al., 2020) .","reference_id":"OTQ1MzI1NTIzNjY2NTYyODg4"} +{"index":3424,"citer_arxiv_id":"2203.17189","cited_arxiv_id":"2105.04663","reference_contexts":"We use the XLA GSPMD partitioner (Xu et al., 2021) to automatically shard the computation graph and use jax.pjit as a frontend to interact with GSPMD, providing our own high-level API to simplify configuration","reference_id":"MTMzOTc3MzM0NjY1OTAxMzIzOTk="} +{"index":3425,"citer_arxiv_id":"2203.17189","cited_arxiv_id":"2105.04663","reference_contexts":"JAX (Bradbury et al., 2018; Frostig et al., 2018) is uniquely positioned to provide such benefits; its NumPy-like (Harris et al., 2020 ) API makes it easy to understand and develop, while the jax.pjit API backed by XLA GSPMD (Xu et al., 2021) provides a powerful and efficient compiler-based programming model for parallelism","reference_id":"MTUwNjE1NDY3NDgyODU3NzcxMA=="} +{"index":3426,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1607.06450","reference_contexts":"One of the most successful approaches to stabilizing neural networks are constraints on activations, and gradients (Pascanu et al., 2013; Ioffe and Szegedy, 2015; Salimans and Kingma, 2016; Ba et al., 2016)","reference_id":"MTI5MDQ5MTQ1NjMzMTMxOTQ5MTc="} +{"index":3427,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2005.14165","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3428,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2202.01169","reference_contexts":"words, sentences, image patches, etc.), but alternative schemes have been proposed (Lewis et al., 2021; Roller et al., 2021; Zuo et al., 2021; Clark et al., 2022) .","reference_id":"OTY3MDYwMzI3NTg2NTcxMDg1Mg=="} +{"index":3429,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2202.01169","reference_contexts":"The remarkable success of scale in natural language processing (Kaplan et al., 2020; Brown et al., 2020) has reinvigorated MoE research evidenced by a surge of recent work (Lepikhin et al., 2020; Fedus et al., 2021; Yang et al., 2021; Kim et al., 2021; Du et al., 2021; Artetxe et al., 2021; Zuo et al., 2021; Clark et al., 2022)","reference_id":"NDM1MjQ1Mjg3Mjc2NjE1NjgzNg=="} +{"index":3430,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1803.05457","reference_contexts":"Question answering is probed with the SQuAD dataset (Rajpurkar et al., 2016) as well as on grade-school science questions in ARC Easy and ARC Reasoning Challenge (Clark et al., 2018)","reference_id":"Mzg3NTE1OTUyMzg1MTkwODk2Mw=="} +{"index":3431,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1810.04805","reference_contexts":"Sentinel tokens denote masked sequences in the span-corruption objective (Fedus et al., 2018; Devlin et al., 2018)","reference_id":"NzA1OTg2MzA4NzczNDQ0ODU="} +{"index":3432,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1810.04805","reference_contexts":"Each model is pre-trained for 20k steps on mC4 using a masked language modeling objective (Fedus et al., 2018; Devlin et al., 2018) .","reference_id":"MTQ2MzIwNDQxNjU1ODQzMzU1NTU="} +{"index":3433,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2010.11929","reference_contexts":"(2021) was introduced in Vision Transformers (Dosovitskiy et al., 2020) for image classification","reference_id":"NzI0ODU2ODYzNTA4MzQ1MDg0Mw=="} +{"index":3434,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1312.4314","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3435,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1801.07736","reference_contexts":"Sentinel tokens denote masked sequences in the span-corruption objective (Fedus et al., 2018; Devlin et al., 2018)","reference_id":"NzA1OTg2MzA4NzczNDQ0ODU="} +{"index":3436,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1801.07736","reference_contexts":"Each model is pre-trained for 20k steps on mC4 using a masked language modeling objective (Fedus et al., 2018; Devlin et al., 2018) .","reference_id":"MTQ2MzIwNDQxNjU1ODQzMzU1NTU="} +{"index":3437,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"selective precision) (Fedus et al., 2021)","reference_id":"MTYwNzIyMjMxNzA1NzUzMjU0MA=="} +{"index":3438,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"These approaches have yielded state-of-the-art translation models (Lepikhin et al., 2020) , 4-7x pre-training speed-ups (Fedus et al., 2021; Artetxe et al., 2021) , and GPT-3 level one-shot performance using 1/3 the energy training cost (Du et al., 2021)","reference_id":"MTI1MTU0NjYyMDk1OTM0MzYwODE="} +{"index":3439,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"The remarkable success of scale in natural language processing (Kaplan et al., 2020; Brown et al., 2020) has reinvigorated MoE research evidenced by a surge of recent work (Lepikhin et al., 2020; Fedus et al., 2021; Yang et al., 2021; Kim et al., 2021; Du et al., 2021; Artetxe et al., 2021; Zuo et al., 2021; Clark et al., 2022)","reference_id":"NDM1MjQ1Mjg3Mjc2NjE1NjgzNg=="} +{"index":3440,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"And in the span of the last twelve months, a milestone of efficiently training trillion parameter deep neural networks has been achieved by multiple groups (Fedus et al., 2021; Yang et al., 2021; Du et al., 2021) , and most recently, Lin et al","reference_id":"MTYzOTg4NjU3NDMwNDQxNTU1NTg="} +{"index":3441,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"One side note is that the recent significant successes of sparse expert models have often been in settings with a lot of data and no distribution shift -two examples being language modeling/span corruption and machine translation (Shazeer et al., 2017; Lepikhin et al., 2020; Kim et al., 2021; Fedus et al., 2021) .","reference_id":"MTU4OTQwODU2NzcwNTc1MjAyMzQ="} +{"index":3442,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"We generalize top-1 routing (Fedus et al., 2021; Roller et al., 2021) and top-2 (Shazeer et al., 2017; Lepikhin et al., 2020) to study top-n routing where each token is processed by at most n experts","reference_id":"MTcxNzg2MDI4NjA3OTE4MjU5Mjk="} +{"index":3443,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"Further, we leave as future work techniques that adapt large language models through reinforcement learning (Ouyang et al., 2022) 4.1 HYPOTHESIS: A GENERALIZATION PROBLEM Sparse models have performed remarkably well in the regime of large datasets, but have sometimes performed poorly when fine-tuning (Fedus et al., 2021; Artetxe et al., 2021)","reference_id":"MTY3NDE2NDgxODU0NTM2OTU4Mw=="} +{"index":3444,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.03961","reference_contexts":"To improve hardware utilization, most implementations of sparse models have static batch sizes for each expert (Shazeer et al., 2017; 2018; Lepikhin et al., 2020; Fedus et al., 2021)","reference_id":"MTQwNjE2NDEzNDQzODcyOTMxMzY="} +{"index":3445,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1606.08415","reference_contexts":"Shazeer (2020) extends this to other variants and presents a GELU-Linear (Hendrycks and Gimpel, 2016) FFN layer as a replacement the usual ReLU (Nair and Hinton, 2010) FFN in Transformer.","reference_id":"MTcxOTU2NzAyMjMxNTgzMjUwNjM="} +{"index":3446,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1705.03551","reference_contexts":"(2020) , we evaluate the knowledge of our models by fine-tuning on three closed-book question answer datasets: Natural Questions (Kwiatkowski et al., 2019) , Web Questions (Berant et al., 2013) and Trivia QA (Joshi et al., 2017)","reference_id":"OTI3NDYwOTcyMjc1ODQ2OTcyOQ=="} +{"index":3447,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2001.08361","reference_contexts":"As seen previously, improvements in small-scale settings may fail to generalize when scaled up and therefore trends should always be monitored and re-assessed at increasing scale (Kaplan et al., 2020) .","reference_id":"NDEzNzY2NzI3OTUwODUwNjU3Nw=="} +{"index":3448,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2001.08361","reference_contexts":"The remarkable success of scale in natural language processing (Kaplan et al., 2020; Brown et al., 2020) has reinvigorated MoE research evidenced by a surge of recent work (Lepikhin et al., 2020; Fedus et al., 2021; Yang et al., 2021; Kim et al., 2021; Du et al., 2021; Artetxe et al., 2021; Zuo et al., 2021; Clark et al., 2022)","reference_id":"NDM1MjQ1Mjg3Mjc2NjE1NjgzNg=="} +{"index":3449,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1808.06226","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3450,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2006.16668","reference_contexts":"These approaches have yielded state-of-the-art translation models (Lepikhin et al., 2020) , 4-7x pre-training speed-ups (Fedus et al., 2021; Artetxe et al., 2021) , and GPT-3 level one-shot performance using 1/3 the energy training cost (Du et al., 2021)","reference_id":"MTI1MTU0NjYyMDk1OTM0MzYwODE="} +{"index":3451,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2006.16668","reference_contexts":"The MoE top-2 routing algorithm (Shazeer et al., 2017; 2018; Lepikhin et al., 2020) works as follows: first the router finds the expert that is assigned the higher router score (gate 1 ) and always sends the token to that expert","reference_id":"MTA5OTIyOTc1ODIzMzEyMjY5NjI="} +{"index":3452,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2006.16668","reference_contexts":"To improve hardware utilization, most implementations of sparse models have static batch sizes for each expert (Shazeer et al., 2017; 2018; Lepikhin et al., 2020; Fedus et al., 2021)","reference_id":"MTQwNjE2NDEzNDQzODcyOTMxMzY="} +{"index":3453,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2006.16668","reference_contexts":"The remarkable success of scale in natural language processing (Kaplan et al., 2020; Brown et al., 2020) has reinvigorated MoE research evidenced by a surge of recent work (Lepikhin et al., 2020; Fedus et al., 2021; Yang et al., 2021; Kim et al., 2021; Du et al., 2021; Artetxe et al., 2021; Zuo et al., 2021; Clark et al., 2022)","reference_id":"NDM1MjQ1Mjg3Mjc2NjE1NjgzNg=="} +{"index":3454,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2006.16668","reference_contexts":"We generalize top-1 routing (Fedus et al., 2021; Roller et al., 2021) and top-2 (Shazeer et al., 2017; Lepikhin et al., 2020) to study top-n routing where each token is processed by at most n experts","reference_id":"MTcxNzg2MDI4NjA3OTE4MjU5Mjk="} +{"index":3455,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2006.16668","reference_contexts":"(2021) for Pareto efficiency, differing from other work which use a larger and more expensive 2.0 capacity factor (Lepikhin et al., 2020; Du et al., 2021) .","reference_id":"MjQwNjMyMTQxNjE3OTIwMTYzMw=="} +{"index":3456,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2104.08691","reference_contexts":"Promising new techniques have emerged as an alternative, including few-shot inference (Brown et al., 2020) , prefix tuning (Li and Liang, 2021), prompt tuning (Lester et al., 2021) , and adapter modules (Houlsby et al., 2019) -however, a quality gap still persists compared to fine-tuning","reference_id":"NjQ2MTQ1MjU2NDIzODc1MzE4OA=="} +{"index":3457,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2103.16716","reference_contexts":"BASE layers (Lewis et al., 2021) recasts token routing as a linear assignment problem -removing the need for load balancing auxiliary losses","reference_id":"NTU4MjAyOTczNzU5NDY4MzE0NQ=="} +{"index":3458,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2103.16716","reference_contexts":"words, sentences, image patches, etc.), but alternative schemes have been proposed (Lewis et al., 2021; Roller et al., 2021; Zuo et al., 2021; Clark et al., 2022) .","reference_id":"OTY3MDYwMzI3NTg2NTcxMDg1Mg=="} +{"index":3459,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2101.00190","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3460,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1710.03740","reference_contexts":"As in most modern distributed Transformers we train with mixed precision (Micikevicius et al., 2017) foot_3","reference_id":"MTY0NTkyNDg2NTE1MzkzOTk0NTk="} +{"index":3461,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2102.11972","reference_contexts":"This quality gain was corroborated in later work (Narang et al., 2021) .","reference_id":"ODA2MjY2NDA5MjE4OTE0MDA1Mw=="} +{"index":3462,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1808.08745","reference_contexts":"The CNN-DM (Hermann et al., 2015) and BBC XSum (Narayan et al., 2018) datasets are used to measure the ability to summarize articles","reference_id":"MTc3MjQ0NTkxMjkxNTI3NzE4NDk="} +{"index":3463,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1511.06807","reference_contexts":"We next explore a hypothesis that adding noise into the model can improve training stability (Neelakantan et al., 2015)","reference_id":"OTQwMDI5NTgyMTU2Njg1Nzg1MA=="} +{"index":3464,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.14599","reference_contexts":"And finally, we test our model's natural language inference capabilities on the Adversarial NLI Benchmark (Nie et al., 2019) .","reference_id":"NjkxOTA3NTMzOTU3MTA3NTI1Ng=="} +{"index":3465,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2104.10350","reference_contexts":"And despite a shocking number of parameters, sparse models reduce the carbon footprint for training large neural networks by an order of magnitude (Patterson et al., 2021)","reference_id":"MjA5OTE0NTc4Nzc0NDQ0Nzg5"} +{"index":3466,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"Sparse expert neural networks showcase the advantage of sheer scale and offer an efficient alternative to the static neural network architectures commonly used today (Raffel et al., 2019; Brown et al., 2020; Rae et al., 2021)","reference_id":"NDA3NTU1MzU3MTEyMzQ5NDYxNg=="} +{"index":3467,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"We design a large-scale stability study of sparse models FLOP-matched to the T5-XL version (Raffel et al., 2019) pre-trained on the multilingual corpus mC4 (Xue et al., 2020)","reference_id":"NDQ0ODg5NjE2NDkxNjgyMjU4Mg=="} +{"index":3468,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"(2021) showed a 4x pre-train speed-up over T5-XXL (Raffel et al., 2019) and Du et al","reference_id":"NTEyMzA0NTY0OTYxOTU3MzcxMQ=="} +{"index":3469,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"We pre-train for 1.5T tokens on a mixture of English-only C4 dataset (Raffel et al., 2019) and the dataset from GLaM (Du et al., 2021) summarized in Appendix E","reference_id":"MzEyODAwODU2OTcyNjYxMjcxNg=="} +{"index":3470,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"Given our improvements to training stability, fine-tuning and model design, we start by validating a sparse model approximately FLOP-matched to T5-Large (Raffel et al., 2019)","reference_id":"ODA3OTUzNDg4NDA4NTgxNTA5Ng=="} +{"index":3471,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"(2021) observed that a sparse 1.6T parameter model achieved a 4x pre-training speed-up over the prior state-of-the-art (Raffel et al., 2019) , but lagged smaller models when fine-tuned on common benchmarks like SuperGLUE","reference_id":"ODMzMTAzNzk1NzkwODYyNjExMg=="} +{"index":3472,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"Table 10 summarizes the quality of a dense T5-Large (L) model and sparse model with approximately the same number of FLOPs pre-trained for 500k steps with a 1M batch size (524B tokens) on the C4 dataset (Raffel et al., 2019)","reference_id":"MTU3NTAyODkxNTY4NzgwODcyMTY="} +{"index":3473,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"Each model was pre-trained on 500B tokens from the C4 corpus (Raffel et al., 2019)","reference_id":"NzM2NDM0MDczNjU1MTI4Nzc4OA=="} +{"index":3474,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.10683","reference_contexts":"We consider our ST-MoE-L model pre-trained either on the monolingual C4 corpus (Raffel et al., 2019) or on the multilingual mC4 corpus (Xue et al., 2020)","reference_id":"Mzg3NjgyNDU2Mzc5MjQ4NjMwNA=="} +{"index":3475,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1606.05250","reference_contexts":"Question answering is probed with the SQuAD dataset (Rajpurkar et al., 2016) as well as on grade-school science questions in ARC Easy and ARC Reasoning Challenge (Clark et al., 2018)","reference_id":"Mzg3NTE1OTUyMzg1MTkwODk2Mw=="} +{"index":3476,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2002.08910","reference_contexts":"(2021) , striking gains are observed in closed book question answering (Roberts et al., 2020) .","reference_id":"MTE5MTAxODU0OTI1NzQ3Nzg5Mw=="} +{"index":3477,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2106.04426","reference_contexts":"Hash layers (Roller et al., 2021) shows random fixed routing (per hash functions) led to competitive performance with learned routing","reference_id":"MTc1Mzg1MzQ1OTU5MjU0NTA4MQ=="} +{"index":3478,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2106.04426","reference_contexts":"We generalize top-1 routing (Fedus et al., 2021; Roller et al., 2021) and top-2 (Shazeer et al., 2017; Lepikhin et al., 2020) to study top-n routing where each token is processed by at most n experts","reference_id":"MTcxNzg2MDI4NjA3OTE4MjU5Mjk="} +{"index":3479,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2106.04426","reference_contexts":"words, sentences, image patches, etc.), but alternative schemes have been proposed (Lewis et al., 2021; Roller et al., 2021; Zuo et al., 2021; Clark et al., 2022) .","reference_id":"OTY3MDYwMzI3NTg2NTcxMDg1Mg=="} +{"index":3480,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1911.02150","reference_contexts":"The compute-tomemory ratio (operational intensity) can serve as an estimate of the efficiency of different operations (Williams et al., 2009; Shazeer, 2019)","reference_id":"NjIyNzA1MjUxODEwMjk3NjQyMA=="} +{"index":3481,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1701.06538","reference_contexts":"The MoE top-2 routing algorithm (Shazeer et al., 2017; 2018; Lepikhin et al., 2020) works as follows: first the router finds the expert that is assigned the higher router score (gate 1 ) and always sends the token to that expert","reference_id":"MTA5OTIyOTc1ODIzMzEyMjY5NjI="} +{"index":3482,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1701.06538","reference_contexts":"Finally, an auxiliary load balancing loss encourages tokens to be roughly evenly distributed across the experts (Shazeer et al., 2017)","reference_id":"MTU5Mzg5OTQxMzIzNTM3ODQxNTk="} +{"index":3483,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1701.06538","reference_contexts":"Rather than applying the same parameters to all inputs, sparse expert networks dynamically select which parameters to use for each input (Shazeer et al., 2017)","reference_id":"MTE2NDM2NjUzMDY0MTM5ODE2NTE="} +{"index":3484,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1701.06538","reference_contexts":"One side note is that the recent significant successes of sparse expert models have often been in settings with a lot of data and no distribution shift -two examples being language modeling/span corruption and machine translation (Shazeer et al., 2017; Lepikhin et al., 2020; Kim et al., 2021; Fedus et al., 2021) .","reference_id":"MTU4OTQwODU2NzcwNTc1MjAyMzQ="} +{"index":3485,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1701.06538","reference_contexts":"We generalize top-1 routing (Fedus et al., 2021; Roller et al., 2021) and top-2 (Shazeer et al., 2017; Lepikhin et al., 2020) to study top-n routing where each token is processed by at most n experts","reference_id":"MTcxNzg2MDI4NjA3OTE4MjU5Mjk="} +{"index":3486,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1701.06538","reference_contexts":"To improve hardware utilization, most implementations of sparse models have static batch sizes for each expert (Shazeer et al., 2017; 2018; Lepikhin et al., 2020; Fedus et al., 2021)","reference_id":"MTQwNjE2NDEzNDQzODcyOTMxMzY="} +{"index":3487,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2109.10686","reference_contexts":"Furthermore, this work emphasizes jointly optimizing both the upstream pre-training and the downstream fine-tuning metrics to avoid discrepancies (Tay et al., 2021)","reference_id":"MTMxMjcyMDk0NzExMTIxODg3MTI="} +{"index":3488,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2112.12731","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3489,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1609.08144","reference_contexts":"One might think that the router z-loss is a convoluted method replaceable by clipping logits (Wu et al., 2016)","reference_id":"ODMyMzEzNzA1OTA2OTAwMzEzOQ=="} +{"index":3490,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2010.11934","reference_contexts":"We consider our ST-MoE-L model pre-trained either on the monolingual C4 corpus (Raffel et al., 2019) or on the multilingual mC4 corpus (Xue et al., 2020)","reference_id":"Mzg3NjgyNDU2Mzc5MjQ4NjMwNA=="} +{"index":3491,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"2010.11934","reference_contexts":"We design a large-scale stability study of sparse models FLOP-matched to the T5-XL version (Raffel et al., 2019) pre-trained on the multilingual corpus mC4 (Xue et al., 2020)","reference_id":"NDQ0ODg5NjE2NDkxNjgyMjU4Mg=="} +{"index":3492,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1910.07467","reference_contexts":"Our second example is the scale parameter in root mean square (RMS) normalization (Zhang and Sennrich, 2019)","reference_id":"NDI5MjM3MjQzNDI0NDgzMjU3OQ=="} +{"index":3493,"citer_arxiv_id":"2202.08906","cited_arxiv_id":"1810.12885","reference_contexts":"We illustrate this problem through two tasks in SuperGLUE (Wang et al., 2019 ) -Commitment Bank (De Marneffe et al., 2019) and ReCORD (Zhang et al., 2018)","reference_id":"MTQzODg3Njc4NDk1OTUwMjE0OTc="} +{"index":3494,"citer_arxiv_id":"2102.09690","cited_arxiv_id":"2012.15723","reference_contexts":"Independent and concurrent work also shows that when finetuning masked language models on few examples, the choice of training examples can impact results (Schick & Schütze, 2020; Gao et al., 2020)","reference_id":"NzIwMDM4NjA4MDI5MjkxNzU5Mg=="} +{"index":3495,"citer_arxiv_id":"2102.09690","cited_arxiv_id":"2012.15723","reference_contexts":"Nevertheless, if these disadvantages are acceptable or avoidable, finetuning can improve accuracy over in-context learning in some cases (Schick & Schütze, 2020; Gao et al., 2020)","reference_id":"Njg0MjU4OTYyOTgxODcyNDkzOA=="} +{"index":3496,"citer_arxiv_id":"2102.09690","cited_arxiv_id":"2012.15723","reference_contexts":"Thus, one should tune the prompt's format and examples to achieve the best possible performance (Brown et al., 2020; Gao et al., 2020)","reference_id":"MTc3NTIzNTEyMDA1MTU1MTM0OA=="} +{"index":3497,"citer_arxiv_id":"2102.09690","cited_arxiv_id":"2009.07118","reference_contexts":"Independent and concurrent work also shows that when finetuning masked language models on few examples, the choice of training examples can impact results (Schick & Schütze, 2020; Gao et al., 2020)","reference_id":"NzIwMDM4NjA4MDI5MjkxNzU5Mg=="} +{"index":3498,"citer_arxiv_id":"2102.09690","cited_arxiv_id":"2009.07118","reference_contexts":"Nevertheless, if these disadvantages are acceptable or avoidable, finetuning can improve accuracy over in-context learning in some cases (Schick & Schütze, 2020; Gao et al., 2020)","reference_id":"Njg0MjU4OTYyOTgxODcyNDkzOA=="} +{"index":3499,"citer_arxiv_id":"2102.09690","cited_arxiv_id":"1806.02847","reference_contexts":"Few-shot Learning with Language Models Recent work uses LMs to solve NLP tasks, e.g., for story cloze prediction (Schwartz et al., 2017) , knowledge base completion (Petroni et al., 2019) , and Winograd schemas (Trinh & Le, 2018)","reference_id":"MzQ5NTk0OTAwNDUzNTE0NTQzNA=="} +{"index":3500,"citer_arxiv_id":"2102.09690","cited_arxiv_id":"1901.11373","reference_contexts":"Few-shot learning-the ability to learn tasks with limited examples-is an important aspect of intelligence (Lake et al., 2015; Yogatama et al., 2019)","reference_id":"MTY0NTIxNzQ0ODI0ODA5OTc4NDE="} +{"index":3501,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"1508.05326","reference_contexts":"To ensure the versatility of mPLUG-DocOwl, we collect diverse document understanding datasets with different task formats, including Visual Question Answering (VQA) [Antol et al., 2015] , Information Extraction (IE), Natural Language Inference (NLI) [Bowman et al., 2015] , and Image Captioning (IC)","reference_id":"MTUwODUyMTMyMzk4Njc2NzU2NjM="} +{"index":3502,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.08485","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3503,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2305.07895","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3504,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2211.05100","reference_contexts":"Large language models (LLMs) like ChatGPT [OpenAI, 2022] , BLOOM [Scao et al., 2022] , and LLaMA [Touvron et al., 2023] have undergone rapid development to enable the realization of general artificial intelligence, boasting impressive zero-shot capabilities across diverse linguistic applications","reference_id":"MTExOTk2NjM5Nzc3NDU0MTAyMw=="} +{"index":3505,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2302.13971","reference_contexts":"Large language models (LLMs) like ChatGPT [OpenAI, 2022] , BLOOM [Scao et al., 2022] , and LLaMA [Touvron et al., 2023] have undergone rapid development to enable the realization of general artificial intelligence, boasting impressive zero-shot capabilities across diverse linguistic applications","reference_id":"MTExOTk2NjM5Nzc3NDU0MTAyMw=="} +{"index":3506,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2303.04671","reference_contexts":"Systematic collaboration approaches, such as Visual ChatGPT [Wu et al., 2023] and MM-REACT [Yang et al., 2023] , leverage various vision experts or tools to express visual information with text descriptions","reference_id":"OTc2NzYwODQ0OTcyMDA4MzYwOA=="} +{"index":3507,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.01196","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3508,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2302.00402","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3509,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2303.11381","reference_contexts":"Systematic collaboration approaches, such as Visual ChatGPT [Wu et al., 2023] and MM-REACT [Yang et al., 2023] , leverage various vision experts or tools to express visual information with text descriptions","reference_id":"OTc2NzYwODQ0OTcyMDA4MzYwOA=="} +{"index":3510,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"The architecture of mPLUG-DocOwl is based on a popular multi-modal language model, mPLUG-Owl [Ye et al., 2023] , which comprises a pre-trained visual foundation model, a visual abstractor, and a language foundation model","reference_id":"MTA5MDE5MTc3MTM1MDI2Mjg3NjY="} +{"index":3511,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"Other training hyper-parameters are the same as mPLUG-Owl [Ye et al., 2023] .","reference_id":"MTYzOTE0NTIzNDcwMDM3OTQ1Ng=="} +{"index":3512,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"To enhance the model's ability of language comprehension and multi-modal open-ended conversation, we follow mPLUG-Owl [Ye et al., 2023] to introduce language-only and general vision-and-language instruction tuning data [Taori et al., 2023 , Vicuna, 2023 , Xu et al., 2023a , Liu et al., 2023a] .","reference_id":"MTUwOTY5MjMwMjk3MDI5MTE5NjU="} +{"index":3513,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"By performing unified instruction tuning for Document Understanding upon the mPLUG-Owl [Ye et al., 2023] , we further propose a modularized MLLM [Li et al., 2022 , Xu et al., 2023b] , namely mPLUG-DocOwl","reference_id":"MTYxNzA5NDc5NjkwOTQwMTU4Nzc="} +{"index":3514,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"mPLUG-Owl [Ye et al., 2023] performs instruction tuning with a unified format as \"Human:{question} AI:{answer}\"","reference_id":"MzcyMjU5Mzc1ODA2NDE5NjI0NA=="} +{"index":3515,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"On the other hand, some approaches, such as MiniGPT-4 [Zhu et al., 2023] , LLaVA [Liu et al., 2023a] , and mPLUG-Owl [Ye et al., 2023] , leverage LLMs to build unified models for multi-modality with limited connected parameters","reference_id":"MTMzMzMwNTEzMTc4NTM4ODAzMjE="} +{"index":3516,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"With the LLM as the language decoder, Multimodal large language models (MLLMs) such as MiniGPT-4 [Zhu et al., 2023] , LLaVA [Liu et al., 2023a] , and mPLUG-Owl [Ye et al., 2023] have demonstrated remarkable zero-shot performance in various open-ended vision-and-language tasks","reference_id":"MTIzOTYwNTA0OTA5MDM4OTQ5ODc="} +{"index":3517,"citer_arxiv_id":"2307.02499","cited_arxiv_id":"2304.14178","reference_contexts":"We compare mPLUG-DocOwl with other popular mult-modal large language models, including mPLUG-Owl [Ye et al., 2023] and Mini-GPT4 [Zhu et al., 2023] , on LLMDoc","reference_id":"ODgzNzYzMjg1MzYyMjUyOTAwOQ=="} +{"index":3518,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1312.6229","reference_contexts":"Since then there has been a notable shift towards CNNs in many areas of computer vision [3, 4, 5, 6, 7, 8]","reference_id":"MzQ5Mjc1OTEwNTYwMjAzNDY2Mg=="} +{"index":3519,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1409.1556","reference_contexts":"Since then there has been a notable shift towards CNNs in many areas of computer vision [3, 4, 5, 6, 7, 8]","reference_id":"MzQ5Mjc1OTEwNTYwMjAzNDY2Mg=="} +{"index":3520,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1409.1556","reference_contexts":"Network depth is a major determinant of model expressiveness, both in theory [9, 10] and in practice [5, 7, 8]","reference_id":"MTIyNjU5OTQwMzA5MjE2NTM2MTk="} +{"index":3521,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1409.1556","reference_contexts":"Whereas AlexNet had 5 convolutional layers [1] , the VGG network and GoogLeNet in 2014 had 19 and 22 layers respectively [5, 7] , and most recently the ResNet architecture featured 152 layers [8] .","reference_id":"MTY1Mjk4MDc3MjE4MjI0MDU1NjY="} +{"index":3522,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1412.6806","reference_contexts":"Since then there has been a notable shift towards CNNs in many areas of computer vision [3, 4, 5, 6, 7, 8]","reference_id":"MzQ5Mjc1OTEwNTYwMjAzNDY2Mg=="} +{"index":3523,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1412.6806","reference_contexts":"to massive data augmentation [6, 25]","reference_id":"NzM2NzIwOTgwNjI2MDAzNTI0Mw=="} +{"index":3524,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1412.6806","reference_contexts":"Horizontal flipping and translation by 4 pixels are the two standard data augmentation techniques adopted in our experiments, following the common practice [6, 13, 20, 21, 24, 26, 30]","reference_id":"MzY5MDUzNjYzOTQ5Nzg0ODcxMA=="} +{"index":3525,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"[8] , the ResNets with constant depth of 1202 layers yields a test error of 6.67%, which is worse than the 110-layer constant depth ResNet","reference_id":"NTc1MTMyOTAyMTg0NTI1NDY1Mg=="} +{"index":3526,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"Network depth is a major determinant of model expressiveness, both in theory [9, 10] and in practice [5, 7, 8]","reference_id":"MTIyNjU5OTQwMzA5MjE2NTM2MTk="} +{"index":3527,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"Whereas AlexNet had 5 convolutional layers [1] , the VGG network and GoogLeNet in 2014 had 19 and 22 layers respectively [5, 7] , and most recently the ResNet architecture featured 152 layers [8] .","reference_id":"MTY1Mjk4MDc3MjE4MjI0MDU1NjY="} +{"index":3528,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"[8] .","reference_id":"MTI0Mzc3OTA4MDUxNjY5NDAzMjM="} +{"index":3529,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"Residual networks (ResNets) [8] simplify Highway Networks by shortcutting (mostly) with identity functions","reference_id":"NDY1MjE4NzQwMDk5NzE5MTk3MQ=="} +{"index":3530,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"Even on modern computers with multiple state-of-the-art GPUs, architectures like the 152-layer ResNet require several weeks to converge on the ImageNet dataset [8] .","reference_id":"Mjc5Njk3MTUwMDYzNDcyNzI1MA=="} +{"index":3531,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"Since then there has been a notable shift towards CNNs in many areas of computer vision [3, 4, 5, 6, 7, 8]","reference_id":"MzQ5Mjc1OTEwNTYwMjAzNDY2Mg=="} +{"index":3532,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"We resolve this conflict by creating deep Residual Network [8] architectures (with hundreds or even thousands of layers) with sufficient modeling capacity; however, during training we shorten the network significantly by randomly removing a substantial fraction of layers independently for each sample or mini-batch","reference_id":"MTkyNDY5NTA0ODA4NzkyNDAwMA=="} +{"index":3533,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"[8]","reference_id":"MTI4ODc5MzEyMTAxNTAxMDE4MjE="} +{"index":3534,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"[8] to build a 152-layer ResNet with 50 bottleneck residual blocks","reference_id":"NzA2Mzk2NjM3ODg1NDY0NDM3Nw=="} +{"index":3535,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"Recently, several new architectures attempt to circumvent this problem through direct identity mappings between layers, which allow the network to pass on features unimpededly from earlier layers to later layers [8, 15] .","reference_id":"NDQxNTQxNTc3NDgzMjMzOTI1NA=="} +{"index":3536,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"[8] , we construct our network as the functional composition of L residual blocks (ResBlocks), each encoding the update rule (1)","reference_id":"MTE4MTA4NDA1NTUwMzE2MzY1Nzc="} +{"index":3537,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1512.03385","reference_contexts":"[8] tried to learn CIFAR-10 using an aggressively deep ResNet with 1202 layers","reference_id":"NTIzNjM2OTQwOTI2Mjg2NjU2"} +{"index":3538,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1409.5185","reference_contexts":"Several approaches exist to reduce this effect in practice, for example through careful initialization [12] , hidden layer supervision [13] , or, recently, Batch Normalization [14] .","reference_id":"Njc3OTE4NTQyNzg5NTE1MjMzMQ=="} +{"index":3539,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1409.5185","reference_contexts":"Horizontal flipping and translation by 4 pixels are the two standard data augmentation techniques adopted in our experiments, following the common practice [6, 13, 20, 21, 24, 26, 30]","reference_id":"MzY5MDUzNjYzOTQ5Nzg0ODcxMA=="} +{"index":3540,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1502.03167","reference_contexts":"We also observe that similar to Dropout [16] , training with stochastic depth acts as a regularizer, even in the presence of Batch Normalization [14]","reference_id":"MTM2MjIwNDMxNjI0MzE2Nzc5OQ=="} +{"index":3541,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1502.03167","reference_contexts":"A notable recent contribution towards training of very deep networks is Batch Normalization [14] , which standardizes the mean and variance of hidden layers with respect to each mini-batch","reference_id":"MTgyODY1NjE3MDU0MDM0MTM1ODY="} +{"index":3542,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1502.03167","reference_contexts":"Several approaches exist to reduce this effect in practice, for example through careful initialization [12] , hidden layer supervision [13] , or, recently, Batch Normalization [14] .","reference_id":"Njc3OTE4NTQyNzg5NTE1MjMzMQ=="} +{"index":3543,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1502.03167","reference_contexts":"Anecdotally, Dropout loses effectiveness when used in combination with Batch Normalization [14, 23]","reference_id":"OTI5MjA3NzAxOTQ4OTk0Mjg1OA=="} +{"index":3544,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1505.00387","reference_contexts":"Diminishing feature reuse during forward propagation (also known as loss in information flow [15] ) refers to the analogous problem to vanishing gradients in the forward direction","reference_id":"MTIyMjIyODU5ODYzMDMxODc2NTI="} +{"index":3545,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1505.00387","reference_contexts":"Highway Networks [15] allow earlier representations to flow unimpededly to later layers through parameterized skip connections known as \"information highways\", which can cross several layers at once","reference_id":"OTQ4OTIyODA3OTA4OTI1NjIyNg=="} +{"index":3546,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1302.4389","reference_contexts":"Many follow up works have been empirically successful, such as DropConnect [20] , Maxout [21] and DropIn [22]","reference_id":"MTY1NzkzNzQ5NTE1NDQzNDAxMjU="} +{"index":3547,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1302.4389","reference_contexts":"Horizontal flipping and translation by 4 pixels are the two standard data augmentation techniques adopted in our experiments, following the common practice [6, 13, 20, 21, 24, 26, 30]","reference_id":"MzY5MDUzNjYzOTQ5Nzg0ODcxMA=="} +{"index":3548,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1312.4400","reference_contexts":"Horizontal flipping and translation by 4 pixels are the two standard data augmentation techniques adopted in our experiments, following the common practice [6, 13, 20, 21, 24, 26, 30]","reference_id":"MzY5MDUzNjYzOTQ5Nzg0ODcxMA=="} +{"index":3549,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1412.6071","reference_contexts":"to massive data augmentation [6, 25]","reference_id":"NzM2NzIwOTgwNjI2MDAzNTI0Mw=="} +{"index":3550,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1412.6830","reference_contexts":"Horizontal flipping and translation by 4 pixels are the two standard data augmentation techniques adopted in our experiments, following the common practice [6, 13, 20, 21, 24, 26, 30]","reference_id":"MzY5MDUzNjYzOTQ5Nzg0ODcxMA=="} +{"index":3551,"citer_arxiv_id":"1603.09382","cited_arxiv_id":"1603.05027","reference_contexts":"Because ImageNet is a very complicated and large dataset, the model complexity required could potentially be much more than that of the 152-layer ResNet [35]","reference_id":"OTMzNTc4NzM0ODc5ODA4ODExNQ=="} +{"index":3552,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2110.14168","reference_contexts":"A surge of datasets has been proposed to facilitate the research in this domain (Upadhyay & Chang, 2017; Amini et al., 2019; Miao et al., 2020; Cobbe et al., 2021)","reference_id":"MzEwNjU0Mzg1NjUyOTE3MzY2OQ=="} +{"index":3553,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"1810.04805","reference_contexts":"To get the contextualized representation of the given problem and candidate examples, we use the BERT (Devlin et al., 2018) [CLS] token representation as the problem encoding","reference_id":"MTM2MjM0ODgxMDQ0NTc4MDEyOTE="} +{"index":3554,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"1810.04805","reference_contexts":"The policy network is built on top of the language model BERT (Devlin et al., 2018) with fixed parameters, followed by a one-layer linear neural network with learnable parameters","reference_id":"MjY3OTkxMjM5ODExMTgzMjc5MQ=="} +{"index":3555,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2209.14375","reference_contexts":"In recent work that focuses on aligning language models with human values (Ouyang et al., 2022; Qiu et al., 2022; Glaese et al., 2022) , policy gradient has been used to optimize language models with rewards learned from human feedback and preference","reference_id":"MTgwMDY2NzQ2ODg0OTQxODU2MDE="} +{"index":3556,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"1412.6980","reference_contexts":"For fine-tuning the UnifiedQA and TAPEX baselines, we use the Adam optimizer (Kingma & Ba, 2014) with an initial learning rate of 5e-5","reference_id":"MjI2Nzk2MzYxODM2MTk1NTU4MQ=="} +{"index":3557,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2205.11916","reference_contexts":"Inspired by the recent progress achieved by GPT-3 in solving MWPs (Wei et al., 2022; Wang et al., 2022; Kojima et al., 2022) , we evaluate TABMWP using GPT-3 models in zero-shot and few-shot learning manners.","reference_id":"MTA4NzM3MDY0NTQ0Mzk2OTM1MzU="} +{"index":3558,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"1509.02971","reference_contexts":"Many policy gradient algorithms have been proposed in the past decade (Silver et al., 2014; Lillicrap et al., 2015; Mnih et al., 2016; Schulman et al., 2017; Barth-Maron et al., 2018)","reference_id":"MTU3OTU1Nzk0MzU3OTM4ODI0NjM="} +{"index":3559,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2212.10535","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3560,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2203.02155","reference_contexts":"In recent work that focuses on aligning language models with human values (Ouyang et al., 2022; Qiu et al., 2022; Glaese et al., 2022) , policy gradient has been used to optimize language models with rewards learned from human feedback and preference","reference_id":"MTgwMDY2NzQ2ODg0OTQxODU2MDE="} +{"index":3561,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"1707.06347","reference_contexts":"Many policy gradient algorithms have been proposed in the past decade (Silver et al., 2014; Lillicrap et al., 2015; Mnih et al., 2016; Schulman et al., 2017; Barth-Maron et al., 2018)","reference_id":"MTU3OTU1Nzk0MzU3OTM4ODI0NjM="} +{"index":3562,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2203.11171","reference_contexts":"Inspired by the recent progress achieved by GPT-3 in solving MWPs (Wei et al., 2022; Wang et al., 2022; Kojima et al., 2022) , we evaluate TABMWP using GPT-3 models in zero-shot and few-shot learning manners.","reference_id":"MTA4NzM3MDY0NTQ0Mzk2OTM1MzU="} +{"index":3563,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2203.11171","reference_contexts":"Unlike random selection (Wei et al., 2022; Wang et al., 2022) , brute-force search, or retrieval-based selection (Liu et al., 2022a) , PROMPTPG learns to construct the prompt dynamically given the candidate pool when interacting with the GPT-3 API.","reference_id":"NTIxNDMwODQ2ODU0NTk5MjExNg=="} +{"index":3564,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2203.11171","reference_contexts":"The in-context examples can be randomly (Wei et al., 2022; Wang et al., 2022) or retrieval-based selected (Liu et al., 2022a) from the training set","reference_id":"MTUyMzg4MDI2OTc4NTYyNTAyMTc="} +{"index":3565,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2203.11171","reference_contexts":"Inspired the success of the large pre-trained language model GPT-3 (Brown et al., 2020) in solving math word problems (Wei et al., 2022; Wang et al., 2022) , we first build a strong baseline using few-shot GPT-3 on TABMWP","reference_id":"MTM2NzMzNjI1Nzc5NjY3NjYwNzI="} +{"index":3566,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2203.11171","reference_contexts":"Provided with a few in-context examples of math word problems as the context, GPT-3 can generate the answer for a test problem, and shows impressive performance across different MWP datasets (Wei et al., 2022; Wang et al., 2022)","reference_id":"NzIzMjExNTA2MjY5MDUwODMxMA=="} +{"index":3567,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2201.11903","reference_contexts":"In the few-shot setting, we follow the standard prompting (Wei et al., 2022) where in-context examples are randomly selected from the training data as demonstrations for the text example","reference_id":"NDg1Mzg2NjYzMjkzNjI2MzQxNg=="} +{"index":3568,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2201.11903","reference_contexts":"Inspired by the recent progress achieved by GPT-3 in solving MWPs (Wei et al., 2022; Wang et al., 2022; Kojima et al., 2022) , we evaluate TABMWP using GPT-3 models in zero-shot and few-shot learning manners.","reference_id":"MTA4NzM3MDY0NTQ0Mzk2OTM1MzU="} +{"index":3569,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2201.11903","reference_contexts":"Unlike random selection (Wei et al., 2022; Wang et al., 2022) , brute-force search, or retrieval-based selection (Liu et al., 2022a) , PROMPTPG learns to construct the prompt dynamically given the candidate pool when interacting with the GPT-3 API.","reference_id":"NTIxNDMwODQ2ODU0NTk5MjExNg=="} +{"index":3570,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2201.11903","reference_contexts":"The in-context examples can be randomly (Wei et al., 2022; Wang et al., 2022) or retrieval-based selected (Liu et al., 2022a) from the training set","reference_id":"MTUyMzg4MDI2OTc4NTYyNTAyMTc="} +{"index":3571,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2201.11903","reference_contexts":"Inspired the success of the large pre-trained language model GPT-3 (Brown et al., 2020) in solving math word problems (Wei et al., 2022; Wang et al., 2022) , we first build a strong baseline using few-shot GPT-3 on TABMWP","reference_id":"MTM2NzMzNjI1Nzc5NjY3NjYwNzI="} +{"index":3572,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2201.11903","reference_contexts":"Provided with a few in-context examples of math word problems as the context, GPT-3 can generate the answer for a test problem, and shows impressive performance across different MWP datasets (Wei et al., 2022; Wang et al., 2022)","reference_id":"NzIzMjExNTA2MjY5MDUwODMxMA=="} +{"index":3573,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"2201.11903","reference_contexts":"We further implement the zero-shot GPT-3 model, the few-shot GPT-3 model, and their chain-of-thought (CoT) reasoning variants (Wei et al., 2022)","reference_id":"MTQ0NzYzOTY4MDY1OTk1NjU2NDI="} +{"index":3574,"citer_arxiv_id":"2209.14610","cited_arxiv_id":"1709.00103","reference_contexts":"Datasets like WTQ (Pasupat & Liang, 2015) , WikiSQL (Zhong et al., 2017) , and SQA (Iyyer et al., 2017) contain semi-structured tables from Wikipedia, while Spider (Yu et al., 2018) collects structured tables sourced from databases","reference_id":"MTQ0NzAxNzMxNjE3MTM4NDE1Njk="} +{"index":3575,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.18223","reference_contexts":"R ECENT years have seen the remarkable progress of LLMs [1] , [2], [3], [4], [5]","reference_id":"MTA3NDM3NDc5OTY5MjE2MDczNA=="} +{"index":3576,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.18223","reference_contexts":"Coarse-grained caption data share some typical traits in common: (1) The data volume is large since samples are generally sourced from the internet","reference_id":"MTg0MDY1NDc1MDM5MDk2MzYwMTg="} +{"index":3577,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.18223","reference_contexts":"The following parts of the survey are structured as such: the survey starts with a comprehensive review of the essential aspects of MLLMs, including (1) Mainstream architectures ( §2); (2) A full recipe of training strategy and data ( §3); (3) Common practices of performance evaluation ( §4)","reference_id":"Njg2MjI3ODQ2MDk2MDExNjMz"} +{"index":3578,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.13971","reference_contexts":"R ECENT years have seen the remarkable progress of LLMs [1] , [2], [3], [4], [5]","reference_id":"MTA3NDM3NDc5OTY5MjE2MDczNA=="} +{"index":3579,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.13971","reference_contexts":"LLaMA series [5] , [57] and Vicuna family [4] are representative open-sourced LLMs that have attracted much academic attention","reference_id":"OTU0ODk4Mjg2MTUzMjkwMjI5MA=="} +{"index":3580,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.13971","reference_contexts":"For example, MobileVLM series [63] , [64] use downscaled LLaMA [5] (termed as MobileLLaMA 1.4B/2.7B), enabling efficient inference on mobile processors.","reference_id":"MTIxNjk1MTgyNzAxOTgxOTQ3MDQ="} +{"index":3581,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.03277","reference_contexts":"By scaling up data size and model size, these LLMs raise extraordinary emergent abilities, typically including instruction following [5] , [6] , In-Context Learning (ICL) [7] , and Chain of Thought (CoT) [8]","reference_id":"MTY5NTk2NDYwNTQxMzA3MzYzNzA="} +{"index":3582,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2201.11903","reference_contexts":"By scaling up data size and model size, these LLMs raise extraordinary emergent abilities, typically including instruction following [5] , [6] , In-Context Learning (ICL) [7] , and Chain of Thought (CoT) [8]","reference_id":"MTY5NTk2NDYwNTQxMzA3MzYzNzA="} +{"index":3583,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2201.11903","reference_contexts":"As the pioneer work [8] points out, CoT is \"a series of intermediate reasoning steps\", which has been proven to be effective in complex reasoning tasks [8] , [182] , [183]","reference_id":"MTU4MDgyOTM3MDg5NTc3OTIwMzE="} +{"index":3584,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.02643","reference_contexts":"Similarly, Osprey [29] supports point input by utilizing a segmentation model [9]","reference_id":"MTUzNDcyNTAzNTA3ODg1Njk4ODM="} +{"index":3585,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.02643","reference_contexts":"Concurrently, Large Vision Models (LVMs) can see clearly [9] , [10] , [11] , [12] , but commonly lag in reasoning.","reference_id":"MTQ2MTkyNjg3MzgzODQ1Njk0NTY="} +{"index":3586,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2203.03605","reference_contexts":"Concurrently, Large Vision Models (LVMs) can see clearly [9] , [10] , [11] , [12] , but commonly lag in reasoning.","reference_id":"MTQ2MTkyNjg3MzgzODQ1Njk0NTY="} +{"index":3587,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.07193","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3588,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2108.10904","reference_contexts":"Prior to MLLM, there have been a lot of works devoted to multimodality, which can be divided into discriminative [13] , [14] , [15] and generative [16] , [17] , [18] paradigms","reference_id":"NDU5NTMwMDIzMTk3MDgyMDA3Mw=="} +{"index":3589,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2109.01652","reference_contexts":"(2) MLLM uses new training paradigms to unleash its full potential, such as using multimodal instruction tuning [19] , [20] to encourage the model to follow new instructions","reference_id":"MTEwNTA3MjM2MDA2NTY1NjAwODc="} +{"index":3590,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2109.01652","reference_contexts":"A closely related technique to ICL is instruction-tuning (see §3.2), which is shown empirically to enhance the ICL ability [19] .","reference_id":"OTM4MzQ3OTkxMjI5Mjc5MjIwOA=="} +{"index":3591,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2109.01652","reference_contexts":"This simple yet effective idea has sparked the success of subsequent NLP works, such as ChatGPT [2], InstructGPT [95] , FLAN [19] , [56] , and OPT-IML [96] .","reference_id":"MTYyNTE4NjkwNzkzOTIyNjAxOTU="} +{"index":3592,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2109.01652","reference_contexts":"The image is from [19] .","reference_id":"MTI3MzM2MTA2MjUwNjI0NjM3Nw=="} +{"index":3593,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"The training objective is typically the original auto-regressive objective used to train LLMs [20] , [37] , [71] , [101] , based on which the MLLM is encouraged to predict the next token of the response","reference_id":"MzE4NTI2NzMxMTQ5NzQ4MjkwNw=="} +{"index":3594,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"In this case, the responses can be naturally judged by benchmark metrics [20] , [60] , [70] , [76], [101] , [102] , [103] , [104]","reference_id":"ODgzODA3MDA1Njc1NDgzMzc0OQ=="} +{"index":3595,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"The instruction template is flexible and subject to manual designs [20] , [25], [98] , as exemplified in Table 5","reference_id":"MTM0NzAyMjQ5Mjg3MDI0MDAyMzE="} +{"index":3596,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"LLaVA [20] extends the approach to the multimodal field by translating images into text of captions and bounding boxes, and prompting text-only GPT-4 to generate new data with the guidance of requirements and demonstrations","reference_id":"MTM0NjE2MzEzNzQwNzUyNjQzMTE="} +{"index":3597,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"LLaVA [20] proposes to score the responses via text-only GPT-4 in terms of different aspects, such as helpfulness and accuracy","reference_id":"MTEwOTY2NTEyMzAyNDkxMDIyMzM="} +{"index":3598,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"For example, LLaVA-Med [158] injects medical knowledge into vanilla LLaVA [20] and develops an assistant specialized in medical image understanding and question answering.","reference_id":"MTIwMDczNTQxNDQ1NjQ4NzE3Njk="} +{"index":3599,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"(2) MLLM uses new training paradigms to unleash its full potential, such as using multimodal instruction tuning [19] , [20] to encourage the model to follow new instructions","reference_id":"MTEwNTA3MjM2MDA2NTY1NjAwODc="} +{"index":3600,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"In contrast, some methods simply use a MLP-based interface to bridge the modality gap [20] , [37] , [71] , [72]","reference_id":"MTgyMjI0OTM4MDI3Njk3NDAwMw=="} +{"index":3601,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"visual encoders and LLMs) frozen and train a learnable interface [20] , [35] , [72]","reference_id":"ODU2MjI1NjkxMTg2MDAyMTk4OQ=="} +{"index":3602,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"For example, LLaVA-RLHF [112] collects human preference data and tunes a model with fewer hallucinations based on LLaVA [20]","reference_id":"MTE2ODAxMTU0NDQ4MzI5MDg2OTk="} +{"index":3603,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.08485","reference_contexts":"Preliminary research on MLLMs focuses on text content generation grounded in text prompts and image [20] , [24] /video [25] , [26] /audio [27]","reference_id":"MTIwMjYwNTg1ODQ3MDQxODM1ODU="} +{"index":3604,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.10592","reference_contexts":"Armed with the two traits, MLLM exhibits new capabilities, such as writing website code based on images [21] , understanding the deep meaning of a meme [22] , and OCR-free math reasoning [23] .","reference_id":"OTk4MjAzODEyMTUxNjU1MjcxNQ=="} +{"index":3605,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.10592","reference_contexts":"Following this idea, subsequent works such as MiniGPT-4 [21] , Chat-Bridge [104] , GPT4Tools [107] , and DetGPT [72] develop different datasets catering for different needs","reference_id":"NDY4OTA1NzYzNDY3NTIwNDc0"} +{"index":3606,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.10592","reference_contexts":"For example, MiniGPT-4 [21] adopts an EVA-CLIP [47] , [48] (ViT-G/14) encoder, which is trained with improved training techniques","reference_id":"NjMzMzMwMzAyNTUzNjI3MDE4Mw=="} +{"index":3607,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.10592","reference_contexts":"Specifically, some works [21] , [35] , [60] , [70] , [102] , [105] hand-craft a pool of candidate instructions and sample one of them during training","reference_id":"NjgyNjY5OTg0NTAyODExNTE4NA=="} +{"index":3608,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"Armed with the two traits, MLLM exhibits new capabilities, such as writing website code based on images [21] , understanding the deep meaning of a meme [22] , and OCR-free math reasoning [23] .","reference_id":"OTk4MjAzODEyMTUxNjU1MjcxNQ=="} +{"index":3609,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"In terms of applications in multimodality, M-ICL is mainly used in two scenarios: (1) solving various visual reasoning tasks [22] , [74] , [178] , [179] , [180] and (2) teaching LLMs to use external tools [169] , [170] , [181]","reference_id":"NDY3OTkzNzQwMzQwOTA3NjYzMw=="} +{"index":3610,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"For example, given an image, MM-REACT [22] can interpret the meaning beneath the surface, such as explaining why a meme is funny","reference_id":"MzE3MjY5NDQ5ODAxNzc0NjE0OQ=="} +{"index":3611,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3612,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"Specifically, in the ICL setting, LLMs learn from a few examples along with an optional instruction and extrapolate to new questions, thereby solving complex and unseen tasks in a few-shot manner [22] , [169] , [170]","reference_id":"MTcwNTMzNTIzMjQ5MTk4NTcwOTI="} +{"index":3613,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"According to the setting, the reasoning systems can be further categorized into few-shot models [22] , [169] , [170] , [181] and zero-shot models [79] , [197]","reference_id":"MTA2ODE2MTk2NTU0NDI5MDAyMTE="} +{"index":3614,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"Similarly, some works [22] , [188] prompt models with descriptions of the task and tool usage to decompose complex tasks into sub-tasks.","reference_id":"MTI3NzM3MTk3OTQyNDA0MDEwMTU="} +{"index":3615,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"The former configuration requires LLMs to decide on their own when to halt the reasoning chains [22] , [116] , [169] , [170] , [185] , [188] , while the latter setting stops the chains with a pre-defined length [79] , [184] , [186] , [187] .","reference_id":"MTc5NTAyOTY4NDYxNjQxNDczMDQ="} +{"index":3616,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.11381","reference_contexts":"In contrast, the predictingbased pattern requires extending the reasoning chains given conditions such as instructions and previous reasoning history [22] , [116] , [169] , [170] , [185] , [188]","reference_id":"NDc3NDg5Nzk4NDQ5MzIzMDEx"} +{"index":3617,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.03378","reference_contexts":"Armed with the two traits, MLLM exhibits new capabilities, such as writing website code based on images [21] , understanding the deep meaning of a meme [22] , and OCR-free math reasoning [23] .","reference_id":"OTk4MjAzODEyMTUxNjU1MjcxNQ=="} +{"index":3618,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.01390","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3619,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.06355","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3620,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.02858","reference_contexts":"Preliminary research on MLLMs focuses on text content generation grounded in text prompts and image [20] , [24] /video [25] , [26] /audio [27]","reference_id":"MTIwMjYwNTg1ODQ3MDQxODM1ODU="} +{"index":3621,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.02858","reference_contexts":"A common and feasible solution is to leverage a group of learnable query tokens to extract information in a query-based manner [69], which first has been implemented in BLIP-2 [59] , and subsequently inherited by a variety of work [26] , [60] , [70]","reference_id":"OTg1NzM2MjA0NjMxNzYzNDA0Nw=="} +{"index":3622,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.15195","reference_contexts":"Finer control on user prompts is developed to support specific regions through boxes [28] or a certain object through a click [29]","reference_id":"NTE5NTA2MTY5Njk5MjM1OTk4Ng=="} +{"index":3623,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.15195","reference_contexts":"On the input side, models that support finer control from user prompts are developed progressively, evolving from image to region [28] , [138] , [139] and even pixels [29] , [140] , [141]","reference_id":"MTAzMzAxMzIyMzY3MjUwMTk1OTI="} +{"index":3624,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.03905","reference_contexts":"(2) Enhanced support on input and output modalities [30] , [31] , such as image, video, audio, and point cloud","reference_id":"Nzc4NjgxODMxMTk1MTM3MTAzNQ=="} +{"index":3625,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.03905","reference_contexts":"ImageBind-LLM [30] uses the Image-Bind [55] encoder, which supports encoding image, text, audio, depth, thermal, and Inertial Measurement Unit (IMU) data","reference_id":"MzA1ODIyMTU4MDU2MzE5NDgyMA=="} +{"index":3626,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.16058","reference_contexts":"(2) Enhanced support on input and output modalities [30] , [31] , such as image, video, audio, and point cloud","reference_id":"Nzc4NjgxODMxMTk1MTM3MTAzNQ=="} +{"index":3627,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.05519","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3628,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.12966","reference_contexts":"Some methods [34] , [81] , [82] also unfreeze more modules (e.g","reference_id":"NjI4MjgzMTAxNjUyNDIyMTQ4"} +{"index":3629,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.12966","reference_contexts":"Take Qwen-VL [34] as an example, the parameter size of the Q-Former is about 0.08B, accounting for less than 1% of the whole parameters, while the encoder and the LLM account for about 19.8% (1.9B) and 80.2% (7.7B), respectively","reference_id":"MTU4MTkyNDg4NDU4ODUyNTM2NjU="} +{"index":3630,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.12966","reference_contexts":"Taking a similar approach, Qwen-VL [34] is developed from the bilingual LLM Qwen [58] and supports both Chinese and English","reference_id":"NzY0MDQ1MDk1MjkzMTA4MDMyOQ=="} +{"index":3631,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.12966","reference_contexts":"Notably, many works have empirically verified that using higher resolution can achieve remarkable performance gains [34] , [50] , [51] , [52]","reference_id":"MTA2ODA2NzU0NzYxMzU3MzI2MTM="} +{"index":3632,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.12966","reference_contexts":"Chinese) with relatively limited training corpus [33] , [34]","reference_id":"MTI1NTA2OTU3MDAwNTg0MzgzNjY="} +{"index":3633,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"The evaluation settings are typically zero-shot [60] , [102] , [104] , [105] or finetuning [20] , [35] , [60] , [70] , [76], [101] , [103] , [105]","reference_id":"MzExNDU1MDUwNjQ5MzQ4NTE3OQ=="} +{"index":3634,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"Subsequent works follow this idea and prompt ChatGPT [81] or GPT-4 [35] , [70] , [101] , [104] , [105] to rate results [35] , [70] , [81] , [101] , [104] or judge which one is better [103] .","reference_id":"NjE0MTA2MzE2Njk3MTA4MTU2Mw=="} +{"index":3635,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"visual encoders and LLMs) frozen and train a learnable interface [20] , [35] , [72]","reference_id":"ODU2MjI1NjkxMTg2MDAyMTk4OQ=="} +{"index":3636,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"Specifically, some works [21] , [35] , [60] , [70] , [102] , [105] hand-craft a pool of candidate instructions and sample one of them during training","reference_id":"NjgyNjY5OTg0NTAyODExNTE4NA=="} +{"index":3637,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"Some studies transfer the strong capabilities of MLLMs to other domains such as medical image understanding [35] , [36] , [37] and document parsing [38] , [39] , [40]","reference_id":"MTYzNzkyOTg4NjQwNTg1MTYyNDg="} +{"index":3638,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"A main issue of applying text-only GPT-4 as an evaluator is that the judge is only based on image-related text content, such as captions or bounding box coordinates, without accessing the image [35]","reference_id":"MTIxMjgwOTk5NjA3NDc4NzQ1NDA="} +{"index":3639,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"document understanding [38] , [39] , [156] , [157] and medical domains [35] , [36] , [37]","reference_id":"MTE5NjYwMDg2MTQyNjA4NjQzOQ=="} +{"index":3640,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.10415","reference_contexts":"The training objective is typically the original auto-regressive objective used to train LLMs [20] , [37] , [71] , [101] , based on which the MLLM is encouraged to predict the next token of the response","reference_id":"MzE4NTI2NzMxMTQ5NzQ4MjkwNw=="} +{"index":3641,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.10415","reference_contexts":"In contrast, some methods simply use a MLP-based interface to bridge the modality gap [20] , [37] , [71] , [72]","reference_id":"MTgyMjI0OTM4MDI3Njk3NDAwMw=="} +{"index":3642,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.10415","reference_contexts":"Note that the instruction template can also be generalized to the case of multi-round conversations [20] , [37] , [71] , [98] .","reference_id":"OTQzNDEzNzc4MzExOTQ3OTAx"} +{"index":3643,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.10415","reference_contexts":"Some studies transfer the strong capabilities of MLLMs to other domains such as medical image understanding [35] , [36] , [37] and document parsing [38] , [39] , [40]","reference_id":"MTYzNzkyOTg4NjQwNTg1MTYyNDg="} +{"index":3644,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.10415","reference_contexts":"document understanding [38] , [39] , [156] , [157] and medical domains [35] , [36] , [37]","reference_id":"MTE5NjYwMDg2MTQyNjA4NjQzOQ=="} +{"index":3645,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.02499","reference_contexts":"document understanding [38] , [39] , [156] , [157] and medical domains [35] , [36] , [37]","reference_id":"MTE5NjYwMDg2MTQyNjA4NjQzOQ=="} +{"index":3646,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.02499","reference_contexts":"Some studies transfer the strong capabilities of MLLMs to other domains such as medical image understanding [35] , [36] , [37] and document parsing [38] , [39] , [40]","reference_id":"MTYzNzkyOTg4NjQwNTg1MTYyNDg="} +{"index":3647,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2403.04473","reference_contexts":"document understanding [38] , [39] , [156] , [157] and medical domains [35] , [36] , [37]","reference_id":"MTE5NjYwMDg2MTQyNjA4NjQzOQ=="} +{"index":3648,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2403.04473","reference_contexts":"Some studies transfer the strong capabilities of MLLMs to other domains such as medical image understanding [35] , [36] , [37] and document parsing [38] , [39] , [40]","reference_id":"MTYzNzkyOTg4NjQwNTg1MTYyNDg="} +{"index":3649,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.14824","reference_contexts":"embodied agents [41] , [42] and GUI agents [43] , [44] , [45]","reference_id":"ODczMDM5NDY0MTc1MDQ4ODI3NQ=="} +{"index":3650,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.13771","reference_contexts":"user-friendly assistants specially designed for Graphical User Interface (GUI), as exemplified by CogAgent [44] , AppAgent [43] , and Mobile-Agent [45]","reference_id":"MTU3NDE0NTIxNTI2MTEzMjQzMDg="} +{"index":3651,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.13771","reference_contexts":"embodied agents [41] , [42] and GUI agents [43] , [44] , [45]","reference_id":"ODczMDM5NDY0MTc1MDQ4ODI3NQ=="} +{"index":3652,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.08914","reference_contexts":"Similarly, CogAgent [44] uses a dual-encoder mechanism, where two encoders process high and low-resolution images, respectively","reference_id":"MTYzMDc0MzA4Njk0NDk0Nzc5MjM="} +{"index":3653,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.08914","reference_contexts":"user-friendly assistants specially designed for Graphical User Interface (GUI), as exemplified by CogAgent [44] , AppAgent [43] , and Mobile-Agent [45]","reference_id":"MTU3NDE0NTIxNTI2MTEzMjQzMDg="} +{"index":3654,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.08914","reference_contexts":"embodied agents [41] , [42] and GUI agents [43] , [44] , [45]","reference_id":"ODczMDM5NDY0MTc1MDQ4ODI3NQ=="} +{"index":3655,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2401.16158","reference_contexts":"user-friendly assistants specially designed for Graphical User Interface (GUI), as exemplified by CogAgent [44] , AppAgent [43] , and Mobile-Agent [45]","reference_id":"MTU3NDE0NTIxNTI2MTEzMjQzMDg="} +{"index":3656,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2401.16158","reference_contexts":"embodied agents [41] , [42] and GUI agents [43] , [44] , [45]","reference_id":"ODczMDM5NDY0MTc1MDQ4ODI3NQ=="} +{"index":3657,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.15389","reference_contexts":"For example, MiniGPT-4 [21] adopts an EVA-CLIP [47] , [48] (ViT-G/14) encoder, which is trained with improved training techniques","reference_id":"NjMzMzMwMzAyNTUzNjI3MDE4Mw=="} +{"index":3658,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2310.03744","reference_contexts":"[50] , [61] find that simply scaling up LLM from 7B to 13B brings comprehensive improvement on various benchmarks","reference_id":"NDAzODMwMzE2NzA4NTE1NjMwMg=="} +{"index":3659,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2310.03744","reference_contexts":"Notably, many works have empirically verified that using higher resolution can achieve remarkable performance gains [34] , [50] , [51] , [52]","reference_id":"MTA2ODA2NzU0NzYxMzU3MzI2MTM="} +{"index":3660,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2310.03744","reference_contexts":"For example, LLaVA series adopts one/two linear MLP [20] , [50] to project visual tokens and align the feature dimension with word embeddings.","reference_id":"MTY4MDk1MTU5NDk5NDI2MTM1Mzg="} +{"index":3661,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.06607","reference_contexts":"Notably, many works have empirically verified that using higher resolution can achieve remarkable performance gains [34] , [50] , [51] , [52]","reference_id":"MTA2ODA2NzU0NzYxMzU3MzI2MTM="} +{"index":3662,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2403.09611","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3663,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.07575","reference_contexts":"For example, Monkey [51] and SPHINX [53] divide a large image into smaller patches and send sub-images together with a downsampled highresolution image to the image encoder, where the subimages and the low-resolution image capture local and global features, respectively","reference_id":"MTI3NDAwNjk0NTE5MTY2ODYzNzQ="} +{"index":3664,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2210.11416","reference_contexts":"This simple yet effective idea has sparked the success of subsequent NLP works, such as ChatGPT [2], InstructGPT [95] , FLAN [19] , [56] , and OPT-IML [96] .","reference_id":"MTYyNTE4NjkwNzkzOTIyNjAxOTU="} +{"index":3665,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2210.11416","reference_contexts":"Among them, Flan-T5 [56] series are relatively early LLMs used in works like BLIP-2 [59] and InstructBLIP [60]","reference_id":"NjM5NTMzMjk5OTMxNjQ3NTA5OA=="} +{"index":3666,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.09288","reference_contexts":"LLaMA series [5] , [57] and Vicuna family [4] are representative open-sourced LLMs that have attracted much academic attention","reference_id":"OTU0ODk4Mjg2MTUzMjkwMjI5MA=="} +{"index":3667,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.16609","reference_contexts":"In contrast, Qwen [58] is a bilingual LLM that supports Chinese and English well.","reference_id":"MTY1NDQ1MjEwNzg1MjQ2ODI4ODc="} +{"index":3668,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.16609","reference_contexts":"Taking a similar approach, Qwen-VL [34] is developed from the bilingual LLM Qwen [58] and supports both Chinese and English","reference_id":"NzY0MDQ1MDk1MjkzMTA4MDMyOQ=="} +{"index":3669,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2301.12597","reference_contexts":"A common and feasible solution is to leverage a group of learnable query tokens to extract information in a query-based manner [69], which first has been implemented in BLIP-2 [59] , and subsequently inherited by a variety of work [26] , [60] , [70]","reference_id":"OTg1NzM2MjA0NjMxNzYzNDA0Nw=="} +{"index":3670,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2301.12597","reference_contexts":"Among them, Flan-T5 [56] series are relatively early LLMs used in works like BLIP-2 [59] and InstructBLIP [60]","reference_id":"NjM5NTMzMjk5OTMxNjQ3NTA5OA=="} +{"index":3671,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.06500","reference_contexts":"A common and feasible solution is to leverage a group of learnable query tokens to extract information in a query-based manner [69], which first has been implemented in BLIP-2 [59] , and subsequently inherited by a variety of work [26] , [60] , [70]","reference_id":"OTg1NzM2MjA0NjMxNzYzNDA0Nw=="} +{"index":3672,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.06500","reference_contexts":"In this case, the responses can be naturally judged by benchmark metrics [20] , [60] , [70] , [76], [101] , [102] , [103] , [104]","reference_id":"ODgzODA3MDA1Njc1NDgzMzc0OQ=="} +{"index":3673,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.06500","reference_contexts":"Hence, abundant works [60] , [70] , [76], [82] , [101] , [102] , [103] , [104] have utilized existing highquality datasets to construct instruction-formatted datasets","reference_id":"MTA4MzU3Nzg1Mzc2MzEwMjIxOTY="} +{"index":3674,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.06500","reference_contexts":"Among them, Flan-T5 [56] series are relatively early LLMs used in works like BLIP-2 [59] and InstructBLIP [60]","reference_id":"NjM5NTMzMjk5OTMxNjQ3NTA5OA=="} +{"index":3675,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.06500","reference_contexts":"A short answer is TABLE 6: Instruction templates for VQA datasets, cited from [60]","reference_id":"MjQzMDQ2MzMwOTMzMzgxNjExNA=="} +{"index":3676,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.09958","reference_contexts":"[62] see a similar phenomenon by scaling up LLMs from 13B to 35B and 65B/70B, where the larger model size brings consistent gains on benchmarks specifically designed for MLLMs","reference_id":"NTkyODI4NTE4OTExOTAzOTI2Mw=="} +{"index":3677,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.16886","reference_contexts":"MobileVLM [63] explores developing small-size variants of MLLMs for resource-limited scenarios","reference_id":"NzAzNDgzMzgwODYyOTY2ODc4NA=="} +{"index":3678,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.16886","reference_contexts":"For example, MobileVLM series [63] , [64] use downscaled LLaMA [5] (termed as MobileLLaMA 1.4B/2.7B), enabling efficient inference on mobile processors.","reference_id":"MTIxNjk1MTgyNzAxOTgxOTQ3MDQ="} +{"index":3679,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2402.03766","reference_contexts":"For example, MobileVLM series [63] , [64] use downscaled LLaMA [5] (termed as MobileLLaMA 1.4B/2.7B), enabling efficient inference on mobile processors.","reference_id":"MTIxNjk1MTgyNzAxOTgxOTQ3MDQ="} +{"index":3680,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2401.04088","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3681,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2401.15947","reference_contexts":"Empirically, MM1 [52] and MoE-LLaVA [68] find that MoE implementation achieves better performance than the dense counterpart on almost all the benchmarks","reference_id":"NzM1MDk1NTk4MjA3ODIzNjI0Mg=="} +{"index":3682,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04160","reference_contexts":"A common and feasible solution is to leverage a group of learnable query tokens to extract information in a query-based manner [69], which first has been implemented in BLIP-2 [59] , and subsequently inherited by a variety of work [26] , [60] , [70]","reference_id":"OTg1NzM2MjA0NjMxNzYzNDA0Nw=="} +{"index":3683,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04160","reference_contexts":"In this case, the responses can be naturally judged by benchmark metrics [20] , [60] , [70] , [76], [101] , [102] , [103] , [104]","reference_id":"ODgzODA3MDA1Njc1NDgzMzc0OQ=="} +{"index":3684,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04160","reference_contexts":"Subsequent works follow this idea and prompt ChatGPT [81] or GPT-4 [35] , [70] , [101] , [104] , [105] to rate results [35] , [70] , [81] , [101] , [104] or judge which one is better [103] .","reference_id":"NjE0MTA2MzE2Njk3MTA4MTU2Mw=="} +{"index":3685,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04160","reference_contexts":"Hence, abundant works [60] , [70] , [76], [82] , [101] , [102] , [103] , [104] have utilized existing highquality datasets to construct instruction-formatted datasets","reference_id":"MTA4MzU3Nzg1Mzc2MzEwMjIxOTY="} +{"index":3686,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.16355","reference_contexts":"The training objective is typically the original auto-regressive objective used to train LLMs [20] , [37] , [71] , [101] , based on which the MLLM is encouraged to predict the next token of the response","reference_id":"MzE4NTI2NzMxMTQ5NzQ4MjkwNw=="} +{"index":3687,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.16355","reference_contexts":"In contrast, some methods simply use a MLP-based interface to bridge the modality gap [20] , [37] , [71] , [72]","reference_id":"MTgyMjI0OTM4MDI3Njk3NDAwMw=="} +{"index":3688,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.16355","reference_contexts":"Note that the instruction template can also be generalized to the case of multi-round conversations [20] , [37] , [71] , [98] .","reference_id":"OTQzNDEzNzc4MzExOTQ3OTAx"} +{"index":3689,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.14167","reference_contexts":"Following this idea, subsequent works such as MiniGPT-4 [21] , Chat-Bridge [104] , GPT4Tools [107] , and DetGPT [72] develop different datasets catering for different needs","reference_id":"NDY4OTA1NzYzNDY3NTIwNDc0"} +{"index":3690,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.14167","reference_contexts":"In contrast, some methods simply use a MLP-based interface to bridge the modality gap [20] , [37] , [71] , [72]","reference_id":"MTgyMjI0OTM4MDI3Njk3NDAwMw=="} +{"index":3691,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.14167","reference_contexts":"visual encoders and LLMs) frozen and train a learnable interface [20] , [35] , [72]","reference_id":"ODU2MjI1NjkxMTg2MDAyMTk4OQ=="} +{"index":3692,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.02469","reference_contexts":"Lynx [73] finds that models pre-trained on large-scale but noisy image-text pairs do not perform as well as models pre-trained with smaller but cleaner datasets","reference_id":"MTMzNjQ1MzA1MDMwNDQ1NTA1NTY="} +{"index":3693,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.02469","reference_contexts":"[73] compare the performance of token and feature-level fusion, and empirically reveal that the token-level fusion variant performs better in terms of VQA benchmarks","reference_id":"MTM3ODg0Mzc4ODM3NzI1NTI3NjM="} +{"index":3694,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.03079","reference_contexts":"Similarly, CogVLM [75] plugs in a visual expert module in each Transformer layer to enable dual interaction and fusion between vision and language features","reference_id":"MTgzNTIzMzE3NDI0MTk1Mjg0NjU="} +{"index":3695,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.16199","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3696,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.06594","reference_contexts":"Apart from the learnable interface, using expert models, such as an image captioning model, is also a feasible way to bridge the modality gap [77] , [78] , [79] , [80]","reference_id":"MTU4MTkyMDI0MTk4MjMxMDQ0MzY="} +{"index":3697,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.06594","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3698,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.14178","reference_contexts":"Apart from the multimodal instruction data, language-only user-assistant conversation data can also be used to improve conversational proficiencies and instruction-following abilities [81] , [98] , [101] , [103] .","reference_id":"NzY4MjgxNDE1NTIzMDc4NDkwNg=="} +{"index":3699,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.14178","reference_contexts":"Some methods [34] , [81] , [82] also unfreeze more modules (e.g","reference_id":"NjI4MjgzMTAxNjUyNDIyMTQ4"} +{"index":3700,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.14178","reference_contexts":"Subsequent works follow this idea and prompt ChatGPT [81] or GPT-4 [35] , [70] , [101] , [104] , [105] to rate results [35] , [70] , [81] , [101] , [104] or judge which one is better [103] .","reference_id":"NjE0MTA2MzE2Njk3MTA4MTU2Mw=="} +{"index":3701,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.14178","reference_contexts":"For example, mPLUG-Owl [81] collects a visually related evaluation set to judge capabilities like natural image understanding, diagram, and flowchart understanding","reference_id":"MTU5ODUwOTI0NjY2ODQwMDcyMDM="} +{"index":3702,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.11175","reference_contexts":"Some methods [34] , [81] , [82] also unfreeze more modules (e.g","reference_id":"NjI4MjgzMTAxNjUyNDIyMTQ4"} +{"index":3703,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.11175","reference_contexts":"Hence, abundant works [60] , [70] , [76], [82] , [101] , [102] , [103] , [104] have utilized existing highquality datasets to construct instruction-formatted datasets","reference_id":"MTA4MzU3Nzg1Mzc2MzEwMjIxOTY="} +{"index":3704,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.12793","reference_contexts":"ShareGPT4V-PT [83] 1.2M Nov-2023 LVIS-Instruct4V [91] 111K Nov-2023 ALLaVA [92] 709K Feb-2024","reference_id":"MjI3OTMwOTgzNzc3ODcyMDg5"} +{"index":3705,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.12793","reference_contexts":"Recently, more works [83] , [91] , [92] have explored generating high-quality fine-grained data through prompting strong MLLMs (e.g","reference_id":"OTg2OTgzNzU3OTk5MTc0NDQ5MQ=="} +{"index":3706,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.12793","reference_contexts":"Besides, ShareGPT4V [83] finds that with high-quality caption data in the pretraining stage, unlocking the vision encode promotes better alignment.","reference_id":"NTM5NjgyOTE3ODg1Mjk3MTQxNg=="} +{"index":3707,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.07574","reference_contexts":"ShareGPT4V-PT [83] 1.2M Nov-2023 LVIS-Instruct4V [91] 111K Nov-2023 ALLaVA [92] 709K Feb-2024","reference_id":"MjI3OTMwOTgzNzc3ODcyMDg5"} +{"index":3708,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.07574","reference_contexts":"Recently, more works [83] , [91] , [92] have explored generating high-quality fine-grained data through prompting strong MLLMs (e.g","reference_id":"OTg2OTgzNzU3OTk5MTc0NDQ5MQ=="} +{"index":3709,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.07574","reference_contexts":"Recently, with the release of the more powerful multimodal model GPT-4V, many works have adopted GPT-4V to generate data of higher quality, as exemplified by LVIS-Instruct4V [91] and ALLaVA [92]","reference_id":"MTc1ODU0NzAwMzQxNDIyMDkyNTk="} +{"index":3710,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2402.11684","reference_contexts":"ShareGPT4V-PT [83] 1.2M Nov-2023 LVIS-Instruct4V [91] 111K Nov-2023 ALLaVA [92] 709K Feb-2024","reference_id":"MjI3OTMwOTgzNzc3ODcyMDg5"} +{"index":3711,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2402.11684","reference_contexts":"Recently, more works [83] , [91] , [92] have explored generating high-quality fine-grained data through prompting strong MLLMs (e.g","reference_id":"OTg2OTgzNzU3OTk5MTc0NDQ5MQ=="} +{"index":3712,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2402.11684","reference_contexts":"Recently, with the release of the more powerful multimodal model GPT-4V, many works have adopted GPT-4V to generate data of higher quality, as exemplified by LVIS-Instruct4V [91] and ALLaVA [92]","reference_id":"MTc1ODU0NzAwMzQxNDIyMDkyNTk="} +{"index":3713,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2212.12017","reference_contexts":"This simple yet effective idea has sparked the success of subsequent NLP works, such as ChatGPT [2], InstructGPT [95] , FLAN [19] , [56] , and OPT-IML [96] .","reference_id":"MTYyNTE4NjkwNzkzOTIyNjAxOTU="} +{"index":3714,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2110.08207","reference_contexts":"Moreover, instruction tuning is highly related to multi-task prompting [97] .","reference_id":"MTIzMTg4OTMzNDI5NjgxNDA0NTM="} +{"index":3715,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04790","reference_contexts":"Apart from the multimodal instruction data, language-only user-assistant conversation data can also be used to improve conversational proficiencies and instruction-following abilities [81] , [98] , [101] , [103] .","reference_id":"NzY4MjgxNDE1NTIzMDc4NDkwNg=="} +{"index":3716,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04790","reference_contexts":"The example is adapted from [98]","reference_id":"Njc4MDY1MTE0NjgxMjc5MzEwNA=="} +{"index":3717,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04790","reference_contexts":"Write a response that appropriately completes the request ### Instruction: {instruction} ### Image: ### Response: {response} ### Image: ### Response: {response} ### Image: ### Response: TABLE 9: A simplified example of the template to structure an M-ICL query, adapted from [98]","reference_id":"NjUyMjAwNTkxMTcxNzQ3MTI5Mg=="} +{"index":3718,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.04790","reference_contexts":"The other works manually design some seed instructions and use these to prompt GPT to generate more [25], [82] , [98]","reference_id":"OTgzMzEyOTc0MDg4ODEzMjQ4NQ=="} +{"index":3719,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.15023","reference_contexts":"The training objective is typically the original auto-regressive objective used to train LLMs [20] , [37] , [71] , [101] , based on which the MLLM is encouraged to predict the next token of the response","reference_id":"MzE4NTI2NzMxMTQ5NzQ4MjkwNw=="} +{"index":3720,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.15023","reference_contexts":"In this case, the responses can be naturally judged by benchmark metrics [20] , [60] , [70] , [76], [101] , [102] , [103] , [104]","reference_id":"ODgzODA3MDA1Njc1NDgzMzc0OQ=="} +{"index":3721,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.15023","reference_contexts":"Subsequent works follow this idea and prompt ChatGPT [81] or GPT-4 [35] , [70] , [101] , [104] , [105] to rate results [35] , [70] , [81] , [101] , [104] or judge which one is better [103] .","reference_id":"NjE0MTA2MzE2Njk3MTA4MTU2Mw=="} +{"index":3722,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.15023","reference_contexts":"LaVIN [101] directly constructs a minibatch by randomly sampling from both language-only and multimodal data.","reference_id":"NjM1NjM2Nzk2OTc0ODk4OTQ1OQ=="} +{"index":3723,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.15023","reference_contexts":"Hence, abundant works [60] , [70] , [76], [82] , [101] , [102] , [103] , [104] have utilized existing highquality datasets to construct instruction-formatted datasets","reference_id":"MTA4MzU3Nzg1Mzc2MzEwMjIxOTY="} +{"index":3724,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.15023","reference_contexts":"Apart from the multimodal instruction data, language-only user-assistant conversation data can also be used to improve conversational proficiencies and instruction-following abilities [81] , [98] , [101] , [103] .","reference_id":"NzY4MjgxNDE1NTIzMDc4NDkwNg=="} +{"index":3725,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2212.10773","reference_contexts":"MultiInstruct [102] probes different strategies for training with a fusion of single modal and multimodal data, including mixed instruction tuning (combine both types of data and randomly shuffle) and sequential instruction tuning (text data followed by multimodal data).","reference_id":"MTc5MTYxNzkxNzMwMjQ2NjU2NTI="} +{"index":3726,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2212.10773","reference_contexts":"In this case, the responses can be naturally judged by benchmark metrics [20] , [60] , [70] , [76], [101] , [102] , [103] , [104]","reference_id":"ODgzODA3MDA1Njc1NDgzMzc0OQ=="} +{"index":3727,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2212.10773","reference_contexts":"Hence, abundant works [60] , [70] , [76], [82] , [101] , [102] , [103] , [104] have utilized existing highquality datasets to construct instruction-formatted datasets","reference_id":"MTA4MzU3Nzg1Mzc2MzEwMjIxOTY="} +{"index":3728,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2212.10773","reference_contexts":"There are also evaluation strategies designed to evaluate a specific aspect of the model [102] , as exemplified by POPE [132] for assessment of hallucination degree.","reference_id":"MTA2MDE4NzA5MDI5MTk1OTc3MDE="} +{"index":3729,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.15010","reference_contexts":"Apart from the multimodal instruction data, language-only user-assistant conversation data can also be used to improve conversational proficiencies and instruction-following abilities [81] , [98] , [101] , [103] .","reference_id":"NzY4MjgxNDE1NTIzMDc4NDkwNg=="} +{"index":3730,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.15010","reference_contexts":"In this case, the responses can be naturally judged by benchmark metrics [20] , [60] , [70] , [76], [101] , [102] , [103] , [104]","reference_id":"ODgzODA3MDA1Njc1NDgzMzc0OQ=="} +{"index":3731,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.15010","reference_contexts":"Subsequent works follow this idea and prompt ChatGPT [81] or GPT-4 [35] , [70] , [101] , [104] , [105] to rate results [35] , [70] , [81] , [101] , [104] or judge which one is better [103] .","reference_id":"NjE0MTA2MzE2Njk3MTA4MTU2Mw=="} +{"index":3732,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.15010","reference_contexts":"Hence, abundant works [60] , [70] , [76], [82] , [101] , [102] , [103] , [104] have utilized existing highquality datasets to construct instruction-formatted datasets","reference_id":"MTA4MzU3Nzg1Mzc2MzEwMjIxOTY="} +{"index":3733,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.04387","reference_contexts":"The evaluation settings are typically zero-shot [60] , [102] , [104] , [105] or finetuning [20] , [35] , [60] , [70] , [76], [101] , [103] , [105]","reference_id":"MzExNDU1MDUwNjQ5MzQ4NTE3OQ=="} +{"index":3734,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.04387","reference_contexts":"Subsequent works follow this idea and prompt ChatGPT [81] or GPT-4 [35] , [70] , [101] , [104] , [105] to rate results [35] , [70] , [81] , [101] , [104] or judge which one is better [103] .","reference_id":"NjE0MTA2MzE2Njk3MTA4MTU2Mw=="} +{"index":3735,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.04387","reference_contexts":"Specifically, some works [21] , [35] , [60] , [70] , [102] , [105] hand-craft a pool of candidate instructions and sample one of them during training","reference_id":"NjgyNjY5OTg0NTAyODExNTE4NA=="} +{"index":3736,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2212.10560","reference_contexts":"To tackle this issue, some works collect samples through self-instruction [106] , which utilizes LLMs to generate textual instruction-following data using a few hand-annotated samples","reference_id":"NjUxMTkzODYyMjA2NDI0MzIyNg=="} +{"index":3737,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.18752","reference_contexts":"Similarly, GPT4Tools [107] builds two sets for the finetuning and zeroshot performance, respectively, and evaluates the responses in terms of thought, action, arguments, and the whole.","reference_id":"ODY3NDE2NTMyNzM5ODAzMjE0MA=="} +{"index":3738,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.18752","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3739,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.18752","reference_contexts":"Following this idea, subsequent works such as MiniGPT-4 [21] , Chat-Bridge [104] , GPT4Tools [107] , and DetGPT [72] develop different datasets catering for different needs","reference_id":"NDY4OTA1NzYzNDY3NTIwNDc0"} +{"index":3740,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.18752","reference_contexts":"Some works adopt further finetuning to improve the planning abilities with respect to tool usage [107] or to improve localization capabilities [142] , [205] of the system","reference_id":"NjE4MTI2ODIzMjEyNzUwMTA3Mg=="} +{"index":3741,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.18752","reference_contexts":"Specifically, LLMs are prompted explicitly to output task planning [181] or, more directly, the modules to call [107] , [169] , [170]","reference_id":"MTAyOTkyMTA4Nzg4MDM5OTU0OA=="} +{"index":3742,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"1909.08593","reference_contexts":"RLHF [110] , [111]","reference_id":"MTczNjQ3MDUzMDQzNzc4OTYyNTM="} +{"index":3743,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.14525","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3744,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.00849","reference_contexts":"RLHF-V [114] collects fine-grained (segment-level) preference data pairs by correcting hallucinations in the model TABLE 8: A summary of datasets for alignment-tuning","reference_id":"MTcxNjA3MjE0Mzc0NjcwODMzNDM="} +{"index":3745,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.00849","reference_contexts":"LLaVA-RLHF [112] 10K I + T → T Human RLHF-V [114] 5.7K I + T → T Human VLFeedback [115] 380K I + T → T GPT-4V","reference_id":"MTY4NzM3OTUwMDc3MzYxNTIyNjM="} +{"index":3746,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2003.10286","reference_contexts":"LLaVA-Med [35] reports results on biomedical VQA [120] , [121] , [122] .","reference_id":"MTQ0NTcxMjk0MjE1NDI5NDk2MjI="} +{"index":3747,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.13394","reference_contexts":"To this end, some efforts have endeavored to develop new benchmarks specially designed for MLLMs [123] , [124] , [125] , [126] , [127] , [128] , [129]","reference_id":"MTUwMTA3MjUyNjAxNTIyNjYyNw=="} +{"index":3748,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.13394","reference_contexts":"With a similar evaluation approach, MME [123] provides a more comprehensive evaluation, covering aspects of existence, count, position and color, as exemplified in [77] .","reference_id":"OTYyMTM5NDA3NTM2Nzg1NDkyNQ=="} +{"index":3749,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.06281","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3750,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.02490","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3751,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.16502","reference_contexts":"To this end, some efforts have endeavored to develop new benchmarks specially designed for MLLMs [123] , [124] , [125] , [126] , [127] , [128] , [129]","reference_id":"MTUwMTA3MjUyNjAxNTIyNjYyNw=="} +{"index":3752,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.05424","reference_contexts":"Video-ChatGPT [130] and Video-Bench [131] focus on video domains and propose specialized benchmarks as well as evaluation tools for assessment","reference_id":"Nzc4NTQyMjg3MDg3NjYyODk2Nw=="} +{"index":3753,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.10355","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3754,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.17421","reference_contexts":"[135] perform indepth qualitative analysis on GPT-4V by crafting a series of samples across various domains and tasks, spanning from preliminary skills, such as caption and object counting, to complex tasks that require world knowledge and reasoning, such as joke understanding and indoor navigation as an embodied agent","reference_id":"MTc4Njk3MTAyNTE3NTAzOTg2MDY="} +{"index":3755,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.12436","reference_contexts":"[137] carry out a comprehensive evaluation on Gemini-Pro by comparing the model against GPT-4V","reference_id":"MTM5OTMxNTI4Mjg2NTMwMTExNjg="} +{"index":3756,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.03601","reference_contexts":"On the input side, models that support finer control from user prompts are developed progressively, evolving from image to region [28] , [138] , [139] and even pixels [29] , [140] , [141]","reference_id":"MTAzMzAxMzIyMzY3MjUwMTk1OTI="} +{"index":3757,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2310.07704","reference_contexts":"On the input side, models that support finer control from user prompts are developed progressively, evolving from image to region [28] , [138] , [139] and even pixels [29] , [140] , [141]","reference_id":"MTAzMzAxMzIyMzY3MjUwMTk1OTI="} +{"index":3758,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.00692","reference_contexts":"Some works adopt further finetuning to improve the planning abilities with respect to tool usage [107] or to improve localization capabilities [142] , [205] of the system","reference_id":"NjE4MTI2ODIzMjEyNzUwMTA3Mg=="} +{"index":3759,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.00692","reference_contexts":"LISA [142] further supports masklevel understanding and reasoning, which makes pixel-level grounding possible","reference_id":"NzgyMzg0MTM0NzQ5MjIxNjAzMQ=="} +{"index":3760,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.16911","reference_contexts":"On the one hand, researchers have explored adapting MLLMs to support the input of more multimodal content, such as 3D point cloud [41] , [143] , [144] , [145]","reference_id":"NTY5NDkxMTE1ODgyMzU1MTg4OQ=="} +{"index":3761,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2402.12226","reference_contexts":"On the other hand, MLLMs are also extended to generate responses of more modalities, such as image [32], [146] , [147] , [148] , audio [32], [147] , [149] , [150] , and video [32], [151]","reference_id":"MTAyNTEzMTYxMjAwMzMxNzYzNzI="} +{"index":3762,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.15564","reference_contexts":"On the other hand, MLLMs are also extended to generate responses of more modalities, such as image [32], [146] , [147] , [148] , audio [32], [147] , [149] , [150] , and video [32], [151]","reference_id":"MTAyNTEzMTYxMjAwMzMxNzYzNzI="} +{"index":3763,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.11000","reference_contexts":"On the other hand, MLLMs are also extended to generate responses of more modalities, such as image [32], [146] , [147] , [148] , audio [32], [147] , [149] , [150] , and video [32], [151]","reference_id":"MTAyNTEzMTYxMjAwMzMxNzYzNzI="} +{"index":3764,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.15021","reference_contexts":"Other works develop agents that interact with real-world [41] , [154] , [155] , e.g","reference_id":"MTM1NjYyMTc5NzcyNDE5MzY2NzI="} +{"index":3765,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2403.12895","reference_contexts":"document understanding [38] , [39] , [156] , [157] and medical domains [35] , [36] , [37]","reference_id":"MTE5NjYwMDg2MTQyNjA4NjQzOQ=="} +{"index":3766,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.00890","reference_contexts":"For example, LLaVA-Med [158] injects medical knowledge into vanilla LLaVA [20] and develops an assistant specialized in medical image understanding and question answering.","reference_id":"MTIwMDczNTQxNDQ1NjQ4NzE3Njk="} +{"index":3767,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2308.15126","reference_contexts":"Different from previous approaches that use matching mechanisms to detect and decide hallucinations, HaELM [161] proposes using text-only LLMs as a judge to automatically decide whether MLLMs' captions are correct against reference captions","reference_id":"MjQwMDYxNzA3OTk1ODEzMzYzMg=="} +{"index":3768,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2311.07397","reference_contexts":"Based on previous studies, AMBER [163] is an LLM-free benchmark that encompasses both discriminative tasks and generative tasks and involves three types of possible hallucinations (see §6.1).","reference_id":"MTA1ODQzOTIxNjMzOTE4MzY3MTg="} +{"index":3769,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2310.00754","reference_contexts":"The other method LURE [167] trains a specialized revisor to mask objects with high uncertainty in the descriptions and regenerates the responses again.","reference_id":"MTcyMzE1MDY4OTY5ODczNjY4MDQ="} +{"index":3770,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2301.00234","reference_contexts":"ditional supervised learning paradigms that learn implicit patterns from abundant data, the crux of ICL is to learn from analogy [168]","reference_id":"MTMyMzQ5MTcwNjQyODM0NzY0MjA="} +{"index":3771,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2301.00234","reference_contexts":"In fact, models are commonly sensitive to the arrangement of demonstrations [168] , [171] .","reference_id":"MTE4MTA4NjQyNzcwMjQxODIxODA="} +{"index":3772,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"In terms of applications in multimodality, M-ICL is mainly used in two scenarios: (1) solving various visual reasoning tasks [22] , [74] , [178] , [179] , [180] and (2) teaching LLMs to use external tools [169] , [170] , [181]","reference_id":"NDY3OTkzNzQwMzQwOTA3NjYzMw=="} +{"index":3773,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"Specifically, LLMs are prompted explicitly to output task planning [181] or, more directly, the modules to call [107] , [169] , [170]","reference_id":"MTAyOTkyMTA4Nzg4MDM5OTU0OA=="} +{"index":3774,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3775,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"Specifically, in the ICL setting, LLMs learn from a few examples along with an optional instruction and extrapolate to new questions, thereby solving complex and unseen tasks in a few-shot manner [22] , [169] , [170]","reference_id":"MTcwNTMzNTIzMjQ5MTk4NTcwOTI="} +{"index":3776,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"Equipped with rich open-world knowledge learned from large-scale pretraining, these systems can easily generalize to unseen objects or concepts with remarkable zero/few-shot performance [169] , [170] , [195] , [197] , [198] , [202]","reference_id":"NzQ0MDk0NjY3MTQ5NDQ2MDczNQ=="} +{"index":3777,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"According to the setting, the reasoning systems can be further categorized into few-shot models [22] , [169] , [170] , [181] and zero-shot models [79] , [197]","reference_id":"MTA2ODE2MTk2NTU0NDI5MDAyMTE="} +{"index":3778,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"The former configuration requires LLMs to decide on their own when to halt the reasoning chains [22] , [116] , [169] , [170] , [185] , [188] , while the latter setting stops the chains with a pre-defined length [79] , [184] , [186] , [187] .","reference_id":"MTc5NTAyOTY4NDYxNjQxNDczMDQ="} +{"index":3779,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2304.09842","reference_contexts":"In contrast, the predictingbased pattern requires extending the reasoning chains given conditions such as instructions and previous reasoning history [22] , [116] , [169] , [170] , [185] , [188]","reference_id":"NDc3NDg5Nzk4NDQ5MzIzMDEx"} +{"index":3780,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2104.08786","reference_contexts":"In fact, models are commonly sensitive to the arrangement of demonstrations [168] , [171] .","reference_id":"MTE4MTA4NjQyNzcwMjQxODIxODA="} +{"index":3781,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2306.05425","reference_contexts":"MIMIC-IT [172] combines in-context learning with instruction tuning by building an instruction dataset formatted with multimodal context","reference_id":"MjM0MDQzNTE0NjE5MzkyNTU5"} +{"index":3782,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2307.05222","reference_contexts":"Emu [173] extends the idea of Flamingo [74] by introducing extra modalities in model generation and corresponding training corpus","reference_id":"NDI2Njk3NDAwNjEzNjQ3ODQx"} +{"index":3783,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2309.07915","reference_contexts":"MMICL [176] aims to augment the capabilities in reasoning with multiple related images","reference_id":"MTU2NjYyNzc4Nzc0NTAyOTk3MDE="} +{"index":3784,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.03726","reference_contexts":"In terms of applications in multimodality, M-ICL is mainly used in two scenarios: (1) solving various visual reasoning tasks [22] , [74] , [178] , [179] , [180] and (2) teaching LLMs to use external tools [169] , [170] , [181]","reference_id":"NDY3OTkzNzQwMzQwOTA3NjYzMw=="} +{"index":3785,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.17580","reference_contexts":"In terms of applications in multimodality, M-ICL is mainly used in two scenarios: (1) solving various visual reasoning tasks [22] , [74] , [178] , [179] , [180] and (2) teaching LLMs to use external tools [169] , [170] , [181]","reference_id":"NDY3OTkzNzQwMzQwOTA3NjYzMw=="} +{"index":3786,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.17580","reference_contexts":"According to the setting, the reasoning systems can be further categorized into few-shot models [22] , [169] , [170] , [181] and zero-shot models [79] , [197]","reference_id":"MTA2ODE2MTk2NTU0NDI5MDAyMTE="} +{"index":3787,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.17580","reference_contexts":"Taking LLMs as helpers with different roles, these works build task-specific [79] , [197] , [198] or general-purpose [22] , [169] , [170] , [181] , [188] visual reasoning systems.","reference_id":"MTA2NDY4NzYzOTI5MTgwMjI2NTQ="} +{"index":3788,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.17580","reference_contexts":"Specifically, LLMs are prompted explicitly to output task planning [181] or, more directly, the modules to call [107] , [169] , [170]","reference_id":"MTAyOTkyMTA4Nzg4MDM5OTU0OA=="} +{"index":3789,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2205.11916","reference_contexts":"As the pioneer work [8] points out, CoT is \"a series of intermediate reasoning steps\", which has been proven to be effective in complex reasoning tasks [8] , [182] , [183]","reference_id":"MTU4MDgyOTM3MDg5NTc3OTIwMzE="} +{"index":3790,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2210.03493","reference_contexts":"As the pioneer work [8] points out, CoT is \"a series of intermediate reasoning steps\", which has been proven to be effective in complex reasoning tasks [8] , [182] , [183]","reference_id":"MTU4MDgyOTM3MDg5NTc3OTIwMzE="} +{"index":3791,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.00923","reference_contexts":"Inspired by the success in NLP, multiple works [184] , [185] , [186] , [187] have been proposed to extend the unimodal CoT to Multimodal CoT (M-CoT)","reference_id":"MTIyNDUwMjk1MzkwNzE3NjQ2NDQ="} +{"index":3792,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.00923","reference_contexts":"Reasoning with a single chain is a paradigm widely used in various methods [116] , [185]","reference_id":"ODE5NDMwODE4NzA5NzMwMDc3MA=="} +{"index":3793,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.00923","reference_contexts":"Multimodal-CoT [185] also uses the ScienceQA benchmark but generates the output in a two-step fashion, i.e","reference_id":"NDMzNTU5NDU5MzYyMzc5ODc3MQ=="} +{"index":3794,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.00923","reference_contexts":"In contrast, the predictingbased pattern requires extending the reasoning chains given conditions such as instructions and previous reasoning history [22] , [116] , [169] , [170] , [185] , [188]","reference_id":"NDc3NDg5Nzk4NDQ5MzIzMDEx"} +{"index":3795,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.04671","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3796,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.04671","reference_contexts":"Similarly, some works [22] , [188] prompt models with descriptions of the task and tool usage to decompose complex tasks into sub-tasks.","reference_id":"MTI3NzM3MTk3OTQyNDA0MDEwMTU="} +{"index":3797,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.04671","reference_contexts":"The former configuration requires LLMs to decide on their own when to halt the reasoning chains [22] , [116] , [169] , [170] , [185] , [188] , while the latter setting stops the chains with a pre-defined length [79] , [184] , [186] , [187] .","reference_id":"MTc5NTAyOTY4NDYxNjQxNDczMDQ="} +{"index":3798,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2303.04671","reference_contexts":"In contrast, the predictingbased pattern requires extending the reasoning chains given conditions such as instructions and previous reasoning history [22] , [116] , [169] , [170] , [185] , [188]","reference_id":"NDc3NDg5Nzk4NDQ5MzIzMDEx"} +{"index":3799,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2205.12255","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3800,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2211.10435","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3801,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2302.04761","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3802,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2112.09332","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3803,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2204.00598","reference_contexts":"Inspired by the success of tool-augmented LLMs [190] , [191] , [192] , [193] , some researches have explored the possibilities of invoking external tools [22] , [107] , [169] , [170] or vision foundation models [22] , [79] , [80] , [188] , [194] , [195] , [196] for visual reasoning tasks","reference_id":"MTQzMzkwMTk5NzQxMDQ5MjM4NTE="} +{"index":3804,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2211.11682","reference_contexts":"Specifically, LLMs are often instructed to integrate information into consistent and fluent natural language sentences [202] or generate texts according to different specific needs [79] , [197] , [198] .","reference_id":"NTE4NjM3ODEyNTYwMjIzODkwMg=="} +{"index":3805,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2211.11682","reference_contexts":"According to the setting, the reasoning systems can be further categorized into few-shot models [22] , [169] , [170] , [181] and zero-shot models [79] , [197]","reference_id":"MTA2ODE2MTk2NTU0NDI5MDAyMTE="} +{"index":3806,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2211.11682","reference_contexts":"Equipped with rich open-world knowledge learned from large-scale pretraining, these systems can easily generalize to unseen objects or concepts with remarkable zero/few-shot performance [169] , [170] , [195] , [197] , [198] , [202]","reference_id":"NzQ0MDk0NjY3MTQ5NDQ2MDczNQ=="} +{"index":3807,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2211.11682","reference_contexts":"Taking LLMs as helpers with different roles, these works build task-specific [79] , [197] , [198] or general-purpose [22] , [169] , [170] , [181] , [188] visual reasoning systems.","reference_id":"MTA2NDY4NzYzOTI5MTgwMjI2NTQ="} +{"index":3808,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2204.00598","reference_contexts":"Specifically, LLMs are often instructed to integrate information into consistent and fluent natural language sentences [202] or generate texts according to different specific needs [79] , [197] , [198] .","reference_id":"NTE4NjM3ODEyNTYwMjIzODkwMg=="} +{"index":3809,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2204.00598","reference_contexts":"Equipped with rich open-world knowledge learned from large-scale pretraining, these systems can easily generalize to unseen objects or concepts with remarkable zero/few-shot performance [169] , [170] , [195] , [197] , [198] , [202]","reference_id":"NzQ0MDk0NjY3MTQ5NDQ2MDczNQ=="} +{"index":3810,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2312.14135","reference_contexts":"Some works adopt further finetuning to improve the planning abilities with respect to tool usage [107] or to improve localization capabilities [142] , [205] of the system","reference_id":"NjE4MTI2ODIzMjEyNzUwMTA3Mg=="} +{"index":3811,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2205.10625","reference_contexts":"This is closely related to the optimization of reasoning chains (see §7.2), or more specifically, the least-to-most prompting [206] technique","reference_id":"MjE2NTczMzc3MzczNTI1NzE1OA=="} +{"index":3812,"citer_arxiv_id":"2306.13549","cited_arxiv_id":"2305.16934","reference_contexts":"Similar to LLMs, MLLMs can be vulnerable to crafted attacks [177] , [207] , [208]","reference_id":"OTU4MTQ4MDMyNzg5ODM2MjEwOQ=="} +{"index":3813,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2204.14198","reference_contexts":"Recent advancement in multimodal pretraining (Wang et al., 2022a; Alayrac et al., 2022; Bao et al., 2022; Wang et al., 2022c) has shown the potential of jointly interpreting text and images in a shared semantic space, which further leads us to ask: can the instruction tuning be leveraged to improve the generalizability of Vision-Language pretrained models on multi-modal and vision tasks?","reference_id":"MTEzNzg3Nzg0NDk3NTIzNjIyNzU="} +{"index":3814,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2204.14198","reference_contexts":"Multimodal Pretraining Multimodal pretraining (Tan and Bansal, 2019; Cho et al., 2021; Singh et al., 2022; Alayrac et al., 2022; Wang et al., 2022a; Li et al., 2022b,a) has significantly advanced the vision-language tasks","reference_id":"MTA0MjQ3NzIxMDkxOTMzNTU2MjY="} +{"index":3815,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2211.09808","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3816,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2206.07160","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3817,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2205.05638","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3818,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2107.13586","reference_contexts":"Prompt tuning (Liu et al., 2021; Li and Liang, 2021; Han et al., 2022; Wang et al., 2022b; Sanh et al., 2022) aims to learn a task-specific prompt by reformulating the downstream tasks to the format that the model was initially trained on and has shown competitive performance across various natural language processing applications","reference_id":"NDk4MzUwODc1OTQzODY2NjU0"} +{"index":3819,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2107.13586","reference_contexts":"With the advances in large-scale pre-trained language models (PLMs), recent studies have explored various efficient learning paradigms (Brown et al., 2020; Liu et al., 2021; Wei et al., 2021; Xie et al., 2021) to generalize PLMs to new tasks without task-specific tuning","reference_id":"NDI2MDc5OTIzNjIxMzQ2ODA4Mw=="} +{"index":3820,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2110.15943","reference_contexts":"As a special form of prompt tuning, in-context learning (Xie et al., 2021; Min et al., 2021) takes one or a few examples as the prompt to demonstrate the task","reference_id":"Njk0ODI3NzAzMjE0ODU4NzE1MA=="} +{"index":3821,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"1212.0402","reference_contexts":"Datasets from more diverse modalities should be considered such as audio (Panayotov et al., 2015; Gemmeke et al., 2017; You et al., 2022) and video (Soomro et al., 2012; Ionescu et al., 2014)","reference_id":"OTY5MjkwMjI2ODI2MDY4NjIyMg=="} +{"index":3822,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"1908.07490","reference_contexts":"Multimodal Pretraining Multimodal pretraining (Tan and Bansal, 2019; Cho et al., 2021; Singh et al., 2022; Alayrac et al., 2022; Wang et al., 2022a; Li et al., 2022b,a) has significantly advanced the vision-language tasks","reference_id":"MTA0MjQ3NzIxMDkxOTMzNTU2MjY="} +{"index":3823,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"1601.07140","reference_contexts":"To build MULTIINSTRUCT, we first collect 34 tasks from the existing studies in visual and multimodal learning, covering Visual Question Answering (Goyal et al., 2017; Krishna et al., 2017; Zhu et al., 2016; Hudson and Manning, 2019; Singh et al., 2019; Marino et al., 2019) , Commonsense Reasoning (Suhr et al., 2017; Liu et al., 2022a; Zellers et al., 2019; Xie et al., 2019) , Region Understanding (Krishna et al., 2017) , Image Understanding (Kafle and Kanan, 2017; Chiu et al., 2020) , Grounded Generation (Krishna et al., 2017; Yu et al., 2016; Lin et al., 2014) , Image-Text Matching (Lin et al., 2014; Goyal et al., 2017) , Grounded Matching (Krishna et al., 2017; Veit et al., 2016; Yu et al., 2016) , Visual Relationship (Krishna et al., 2017; Pham et al., 2021) , Temporal Ordering tasks that are created from WikiHow 3 , and Miscellaneous (Yao et al., 2022; Kiela et al., 2020; Das et al., 2017; Lin et al., 2014; Veit et al., 2016; Alam et al., 2022)","reference_id":"NzI3ODM0NTc1ODI3ODI2MTkwNA=="} +{"index":3824,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2202.03052","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3825,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2208.10442","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3826,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2109.01652","reference_contexts":"We follow the same instruction tuning setting as the previous study (Wei et al., 2021) and mainly evaluate the zero-shot learning capabilities of the finetuned large language models","reference_id":"MTkxNjgwODkxMzQ0ODgxNTA0Mw=="} +{"index":3827,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2109.01652","reference_contexts":"With the advances in large-scale pre-trained language models (PLMs), recent studies have explored various efficient learning paradigms (Brown et al., 2020; Liu et al., 2021; Wei et al., 2021; Xie et al., 2021) to generalize PLMs to new tasks without task-specific tuning","reference_id":"NDI2MDc5OTIzNjIxMzQ2ODA4Mw=="} +{"index":3828,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2109.01652","reference_contexts":"Among these, instruction tuning (Wei et al., 2021) has achieved significant success in zero-shot learning on natural language processing tasks","reference_id":"MTQzNzE4NTE2MDE2MDEzODE0MzU="} +{"index":3829,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2109.01652","reference_contexts":"Instruction tuning (Wei et al., 2021) is another simple yet effective strategy to improve the generalizability of large language models","reference_id":"MzIzMzc3NTk4NDUxMTkwMzU4Mg=="} +{"index":3830,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"1901.06706","reference_contexts":"In this work, we propose MULTIINSTRUCT, the first benchmark dataset for multimodal instruction tuning with 62 diverse tasks from 10 broad categories, including Visual Question Answering (Goyal et al., 2017; Suhr et al., 2017) , Commonsense Reasoning (Zellers et al., 2019; Xie et al., 2019) , Visual Relationship Understanding (Krishna et al., 2017) and so on","reference_id":"MTQzMjI4MzgwNTAxMTMzNjUwMTU="} +{"index":3831,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"1901.06706","reference_contexts":"Visual Entailment (Xie et al., 2019) requires the model to determine whether the image semantically entails the text.","reference_id":"MTUyNzc2ODc4MzU0OTAyNjQ2NzM="} +{"index":3832,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"1901.06706","reference_contexts":"To build MULTIINSTRUCT, we first collect 34 tasks from the existing studies in visual and multimodal learning, covering Visual Question Answering (Goyal et al., 2017; Krishna et al., 2017; Zhu et al., 2016; Hudson and Manning, 2019; Singh et al., 2019; Marino et al., 2019) , Commonsense Reasoning (Suhr et al., 2017; Liu et al., 2022a; Zellers et al., 2019; Xie et al., 2019) , Region Understanding (Krishna et al., 2017) , Image Understanding (Kafle and Kanan, 2017; Chiu et al., 2020) , Grounded Generation (Krishna et al., 2017; Yu et al., 2016; Lin et al., 2014) , Image-Text Matching (Lin et al., 2014; Goyal et al., 2017) , Grounded Matching (Krishna et al., 2017; Veit et al., 2016; Yu et al., 2016) , Visual Relationship (Krishna et al., 2017; Pham et al., 2021) , Temporal Ordering tasks that are created from WikiHow 3 , and Miscellaneous (Yao et al., 2022; Kiela et al., 2020; Das et al., 2017; Lin et al., 2014; Veit et al., 2016; Alam et al., 2022)","reference_id":"NzI3ODM0NTc1ODI3ODI2MTkwNA=="} +{"index":3833,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2111.02080","reference_contexts":"As a special form of prompt tuning, in-context learning (Xie et al., 2021; Min et al., 2021) takes one or a few examples as the prompt to demonstrate the task","reference_id":"Njk0ODI3NzAzMjE0ODU4NzE1MA=="} +{"index":3834,"citer_arxiv_id":"2212.10773","cited_arxiv_id":"2111.02080","reference_contexts":"With the advances in large-scale pre-trained language models (PLMs), recent studies have explored various efficient learning paradigms (Brown et al., 2020; Liu et al., 2021; Wei et al., 2021; Xie et al., 2021) to generalize PLMs to new tasks without task-specific tuning","reference_id":"NDI2MDc5OTIzNjIxMzQ2ODA4Mw=="} +{"index":3835,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"1604.06174","reference_contexts":"The standard approach to alleviate this memory pressure is to simply not store most of the activations and recompute them as necessary to calculate gradients during the backward pass [5]","reference_id":"MzcwMjA1MjkzMjExMjM4MDU1Mw=="} +{"index":3836,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"1604.06174","reference_contexts":"Activation recomputation [5] overcomes this memory limitation by storing (or \"checkpointing\") the input activations to a group of layers and recomputing other required activations using an extra forward pass during back-propagation (this is referred to in this paper as full activation recomputation)","reference_id":"MTYxNzY5NTYyMTIzODU0Njc1MTM="} +{"index":3837,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2104.04473","reference_contexts":"Compared to model parallelism, these techniques, which are based on data parallelism, are less efficient and do not scale well to a large numbers of GPUs [13] and are thus a better fit for finetuning models in resource-constrained environments","reference_id":"MTcwNTU3MTEyNDE4ODcwNTExNzI="} +{"index":3838,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2104.04473","reference_contexts":"For example, the interleaving schedule developed in Megatron-LM [13] requires storing activations for L(1 + p-1 pm ) layers where m is the number of interleaving stages","reference_id":"NjM4MzY5MTg5MjgyNjIxMzc3Mw=="} +{"index":3839,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2104.04473","reference_contexts":"Some recent approaches combine both types of model parallelism to enable training large models up to 1T parameters [13] .","reference_id":"NTczOTQ5MDk2ODgyNjg2ODQxNw=="} +{"index":3840,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2104.04473","reference_contexts":"To keep the pipeline pressurized and avoid extra idle time, the first stage must store activations for p microbatches (for more details see Figure 4 -top of [13] )","reference_id":"MTgzNDEyMzg2MzcyMDI1OTgyMTA="} +{"index":3841,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2104.04473","reference_contexts":"This is due to the overlapping that pipeline parallel schedules introduce to reduce the pipeline bubble [13] .","reference_id":"ODc0NjAwNTM1MTQxMDg0NTQzNw=="} +{"index":3842,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2104.04473","reference_contexts":"[13] and only consider the matrix multiplications (GEMMs) which are the main contributors to the number of floating-point operations","reference_id":"MTkzNTM5MjQ4NDc4NDc4OTAyOQ=="} +{"index":3843,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2104.04473","reference_contexts":"Pipeline parallelism requires storing the activations of several microbatches to reduce the pipeline bubble [13]","reference_id":"MTE4OTY0ODI3NDA1ODYxNTY2OTI="} +{"index":3844,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"1909.08053","reference_contexts":"In addition, tensor parallelism as introduced in Megatron-LM [19] helps to reduce the activation memory to some extent","reference_id":"MTcyNjM4MTE5NTY3OTczMjk3NDA="} +{"index":3845,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"1909.08053","reference_contexts":"Two forms of model parallelism are commonly used to distribute the model parameters across GPUs: parameters of each layer are distributed across many devices [18, 19, 21] , and 2) pipeline parallelism where the model is split along the layer dimension of the network [7, 10, 11]","reference_id":"MjM5NjAxNTg1NzMwODY1NDM1MA=="} +{"index":3846,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"1909.08053","reference_contexts":"By splitting A along its columns (A c 1 and A c 2 ) and B along its rows (B r 1 and B r 2 ), we avoid communications (for more details please see [19] ) and arrive at W 1 and W 2","reference_id":"MTc3MTg5MTk1MTkwNjEyOTc0NTE="} +{"index":3847,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"1909.08053","reference_contexts":"[19] and parallelize the attention as well as MLP blocks as shown in Figure 4","reference_id":"NDI4NTk0MDE2MTAwMjk2MjgwNA=="} +{"index":3848,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2105.04663","reference_contexts":"Sagemaker [8] and GSPMD [21] propose memory efficient versions of tensor parallelism which splits the activations across the devices along the hidden dimension throughout the network","reference_id":"NTE5NDUzNTQzNjYzMDk4OTUzNg=="} +{"index":3849,"citer_arxiv_id":"2205.05198","cited_arxiv_id":"2105.04663","reference_contexts":"Two forms of model parallelism are commonly used to distribute the model parameters across GPUs: parameters of each layer are distributed across many devices [18, 19, 21] , and 2) pipeline parallelism where the model is split along the layer dimension of the network [7, 10, 11]","reference_id":"MjM5NjAxNTg1NzMwODY1NDM1MA=="} +{"index":3850,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2204.14198","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3851,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2204.06745","reference_contexts":"Models are trained with self-supervision on large, general corpuses and they perform well on hundreds of tasks (Brown et al., 2020; Rae et al., 2021; Hoffmann et al., 2022; Black et al., 2022; Zhang et al., 2022; Chowdhery et al., 2022)","reference_id":"MTUwNTY1NjkyMjAwNDk5ODU1NDc="} +{"index":3852,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2005.14050","reference_contexts":"Models may mirror misinformation, stereotypes and bias in the corpus (Sheng et al., 2019; Kurita et al., 2019; Dev et al., 2019; Blodgett et al., 2020; Sheng et al., 2021)","reference_id":"NDg5MzcyNjc5Mjc4MjI1MDU3OA=="} +{"index":3853,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2112.04426","reference_contexts":"Examples of such models include RAG, RETRO and Atlas (Lewis et al., 2020b; Borgeaud et al., 2021; Izacard et al., 2022)","reference_id":"OTM3ODk1Mzk5MTQ3MTk3ODc1OA=="} +{"index":3854,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2005.14165","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3855,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2204.02311","reference_contexts":"The capability set increases with scale, and recent work has highlighted reasoning capabilities at larger scales with a suitable prompting strategy (Wei et al., 2022; Chowdhery et al., 2022; Kojima et al., 2022; Lewkowycz et al., 2022) .","reference_id":"MTM4OTYxMTg5MjgxMDQ5ODU2NTc="} +{"index":3856,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2204.02311","reference_contexts":"• No Biases -following PaLM, we do not use biases in any of the dense kernels or layer norms (Chowdhery et al., 2022)","reference_id":"MTIyMzkxMzE3MDA2MDA3MTA1NzA="} +{"index":3857,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2204.02311","reference_contexts":"For example, the PaLM corpus is 50% social media conversations, which may have limited transfer towards scientific tasks (Chowdhery et al., 2022)","reference_id":"OTkzNjk4NzkwNTUxNTA4Mzg1NQ=="} +{"index":3858,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2204.02311","reference_contexts":"Models are trained with self-supervision on large, general corpuses and they perform well on hundreds of tasks (Brown et al., 2020; Rae et al., 2021; Hoffmann et al., 2022; Black et al., 2022; Zhang et al., 2022; Chowdhery et al., 2022)","reference_id":"MTUwNTY1NjkyMjAwNDk5ODU1NDc="} +{"index":3859,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2210.11416","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3860,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1905.10044","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3861,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2110.14168","reference_contexts":"For GSM8k we use the provided training dataset and convert so the calculator steps are performed by writing a Python program, following the format (Cobbe et al., 2021)","reference_id":"MTI1MzQ1NTg5NDk2NzEyMzA4MzY="} +{"index":3862,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1908.09369","reference_contexts":"Models may mirror misinformation, stereotypes and bias in the corpus (Sheng et al., 2019; Kurita et al., 2019; Dev et al., 2019; Blodgett et al., 2020; Sheng et al., 2021)","reference_id":"NDg5MzcyNjc5Mjc4MjI1MDU3OA=="} +{"index":3863,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1811.01241","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3864,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2009.11462","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3865,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1606.08415","reference_contexts":"• GeLU Activation -we use GeLU activations for all model sizes (Hendrycks and Gimpel, 2016) .","reference_id":"Nzk1OTUzMDM0ODU4OTUwODI2Ng=="} +{"index":3866,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2009.03300","reference_contexts":"This includes scientific knowledge tasks such as MMLU (Hendrycks et al., 2020)","reference_id":"NjY3MTYxMDY3NDAxNTM3OTc4Mw=="} +{"index":3867,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2009.03300","reference_contexts":"They perform well on knowledge-intensive tasks such as general knowledge (TriviaQA) and specialist knowledge (MMLU) without an external retrieval mechanism (Brown et al., 2020; Hendrycks et al., 2020) .","reference_id":"MTUwNDQ5NDAyMTE1MTgyNzMwNjQ="} +{"index":3868,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2009.03300","reference_contexts":"We start by evaluating on the MMLU mathematics benchmarks, which we report in Table 8 (Hendrycks et al., 2020)","reference_id":"NDI1MDM3NDg1ODE3OTk5ODkwOQ=="} +{"index":3869,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2103.03874","reference_contexts":"For KhanProblems, we used the problems from AMPS and converted to a format (Hendrycks et al., 2021)","reference_id":"MTIzNjc4NzA5MDI0NDY1NTAwMw=="} +{"index":3870,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2103.03874","reference_contexts":"We also evaluate on the MATH dataset to further probe the reasoning capabilities of Galactica (Hendrycks et al., 2021)","reference_id":"OTM5MDg0MDg3ODUxMDQ1MDUzMg=="} +{"index":3871,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2205.10487","reference_contexts":"This is unexpected as existing research suggests repeated tokens can be harmful on performance (Hernandez et al., 2022)","reference_id":"MTM2MTkwOTE3ODMxMzg3MDgxMDg="} +{"index":3872,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2203.15556","reference_contexts":"The Chinchilla paper derived scaling \"laws\" taking into account number of tokens, training a 70bn model for 1.4 trillion tokens (Hoffmann et al., 2022)","reference_id":"OTg5MjAyNDIwMTI2ODQzODA5OA=="} +{"index":3873,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2203.15556","reference_contexts":"Another implication is that the focus on tokens → ∞ from Chinchilla needs to be complemented with strong data quality procedures (Hoffmann et al., 2022)","reference_id":"NTU2OTQzOTcwNTQxOTA1ODU2"} +{"index":3874,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2203.15556","reference_contexts":"Models are trained with self-supervision on large, general corpuses and they perform well on hundreds of tasks (Brown et al., 2020; Rae et al., 2021; Hoffmann et al., 2022; Black et al., 2022; Zhang et al., 2022; Chowdhery et al., 2022)","reference_id":"MTUwNTY1NjkyMjAwNDk5ODU1NDc="} +{"index":3875,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1607.01759","reference_contexts":"(Joulin et al., 2016) using a noisy set of 600 domains","reference_id":"Mjc3NDc1NjY0ODk2NzE4MDUzNw=="} +{"index":3876,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2001.08361","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3877,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2005.00700","reference_contexts":"And additionally there is the UnifiedQA approach (Khashabi et al., 2020)","reference_id":"MTM4MjU3MjA5Njk2ODM1MTk4Mg=="} +{"index":3878,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2205.11916","reference_contexts":"The capability set increases with scale, and recent work has highlighted reasoning capabilities at larger scales with a suitable prompting strategy (Wei et al., 2022; Chowdhery et al., 2022; Kojima et al., 2022; Lewkowycz et al., 2022) .","reference_id":"MTM4OTYxMTg5MjgxMDQ5ODU2NTc="} +{"index":3879,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1906.07337","reference_contexts":"Models may mirror misinformation, stereotypes and bias in the corpus (Sheng et al., 2019; Kurita et al., 2019; Dev et al., 2019; Blodgett et al., 2020; Sheng et al., 2021)","reference_id":"NDg5MzcyNjc5Mjc4MjI1MDU3OA=="} +{"index":3880,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2005.11401","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3881,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2206.14858","reference_contexts":"The capability set increases with scale, and recent work has highlighted reasoning capabilities at larger scales with a suitable prompting strategy (Wei et al., 2022; Chowdhery et al., 2022; Kojima et al., 2022; Lewkowycz et al., 2022) .","reference_id":"MTM4OTYxMTg5MjgxMDQ5ODU2NTc="} +{"index":3882,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1711.05101","reference_contexts":"We train using AdamW with β 1 = 0.9, β 2 = 0.95 and weight decay of 0.1 (Loshchilov and Hutter, 2017)","reference_id":"MjA0MzY2NjYyMjA4NDY2MDc1Nw=="} +{"index":3883,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2108.12409","reference_contexts":"We experimented with ALiBi at smaller scales but did not observe large gains, so we did not use it (Press et al., 2021)","reference_id":"MjI5MzI0MjEwMTI0MzgxNzA2OA=="} +{"index":3884,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2112.11446","reference_contexts":"They obtained state-of-the-art performance on MMLU, beating much larger models such as Gopher (Rae et al., 2021) .","reference_id":"NTM2MzY3NzU2NzI5NzAzOTg1Nw=="} +{"index":3885,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2112.11446","reference_contexts":"Models are trained with self-supervision on large, general corpuses and they perform well on hundreds of tasks (Brown et al., 2020; Rae et al., 2021; Hoffmann et al., 2022; Black et al., 2022; Zhang et al., 2022; Chowdhery et al., 2022)","reference_id":"MTUwNTY1NjkyMjAwNDk5ODU1NDc="} +{"index":3886,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2202.07206","reference_contexts":"Prior work has shown that accuracy on tasks like multiplication is proportional to term frequency (Razeghi et al., 2022)","reference_id":"ODUwNTY5MTc5NDE1ODUxNzY1MA=="} +{"index":3887,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2110.08207","reference_contexts":"Separately, research such as FLAN and T0 showed prompt tuning can boost downstream performance (Wei et al., 2021; Sanh et al., 2021; Chung et al., 2022)","reference_id":"MjM4MzEzNDA2ODE2NTExNjY0OQ=="} +{"index":3888,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1508.07909","reference_contexts":"• Vocabulary -we construct a vocabulary of 50k tokens using BPE (Sennrich et al., 2015)","reference_id":"ODg0NTcyMTQ4NzM0Mzg2OTQ3OQ=="} +{"index":3889,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1909.01326","reference_contexts":"Models may mirror misinformation, stereotypes and bias in the corpus (Sheng et al., 2019; Kurita et al., 2019; Dev et al., 2019; Blodgett et al., 2020; Sheng et al., 2021)","reference_id":"NDg5MzcyNjc5Mjc4MjI1MDU3OA=="} +{"index":3890,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2105.04054","reference_contexts":"Models may mirror misinformation, stereotypes and bias in the corpus (Sheng et al., 2019; Kurita et al., 2019; Dev et al., 2019; Blodgett et al., 2020; Sheng et al., 2021)","reference_id":"NDg5MzcyNjc5Mjc4MjI1MDU3OA=="} +{"index":3891,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2210.09261","reference_contexts":"This is seen in recent work on chain-of-thought prompting (Wei et al., 2022; Suzgun et al., 2022)","reference_id":"MTQ0NTM1MjExNTM0MTg0OTg0ODQ="} +{"index":3892,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1811.00937","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3893,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2207.10551","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3894,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2210.11399","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3895,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2201.08239","reference_contexts":"For example, prior work has looked at the possibilities of external tool augmentation, such as calculators (Thoppilan et al., 2022)","reference_id":"MTM4NDA1NzU0ODM5MjEyNDUwMTI="} +{"index":3896,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3897,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2109.01652","reference_contexts":"Separately, research such as FLAN and T0 showed prompt tuning can boost downstream performance (Wei et al., 2021; Sanh et al., 2021; Chung et al., 2022)","reference_id":"MjM4MzEzNDA2ODE2NTExNjY0OQ=="} +{"index":3898,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2201.11903","reference_contexts":"The capability set increases with scale, and recent work has highlighted reasoning capabilities at larger scales with a suitable prompting strategy (Wei et al., 2022; Chowdhery et al., 2022; Kojima et al., 2022; Lewkowycz et al., 2022) .","reference_id":"MTM4OTYxMTg5MjgxMDQ5ODU2NTc="} +{"index":3899,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2201.11903","reference_contexts":"This is seen in recent work on chain-of-thought prompting (Wei et al., 2022; Suzgun et al., 2022)","reference_id":"MTQ0NTM1MjExNTM0MTg0OTg0ODQ="} +{"index":3900,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2205.01068","reference_contexts":"Language models such as OPT use the Pushshift.io Reddit corpus as a primary data source, which likely leads the model to learn more discriminatory associations (Zhang et al., 2022)","reference_id":"MTY4MTcwNzUxNTE5NjU1NjQxNTc="} +{"index":3901,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2205.01068","reference_contexts":"We build on the open LLM training foundation they made possible with the OPT project (Zhang et al., 2022) .","reference_id":"MTA3OTAxNDY0MDUzMDA1MzkwMTY="} +{"index":3902,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"2205.01068","reference_contexts":"Models are trained with self-supervision on large, general corpuses and they perform well on hundreds of tasks (Brown et al., 2020; Rae et al., 2021; Hoffmann et al., 2022; Black et al., 2022; Zhang et al., 2022; Chowdhery et al., 2022)","reference_id":"MTUwNTY1NjkyMjAwNDk5ODU1NDc="} +{"index":3903,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"W O is the linear projection from dvh to dv dimensions (Vaswani et al., 2017) .","reference_id":"MTMxMTg2NTU4OTA5NDk2ODYyNDU="} +{"index":3904,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"The multi-head attention mechanism is defined as several scaled dot-product attention stacked together, where the input is linearly projected h times with different, learned linear projections to d k , d k , and dv dimensions, respectively (Vaswani et al., 2017) .","reference_id":"MTgxNTUwMDYzOTI3NTMxOTU1NjM="} +{"index":3905,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"The setup for dense retrieval is: (1) each reference is encoded by the model using its title and abstract, ( 2 ) a text query is encoded by the same model, (3) the references that match the query re returned","reference_id":"Njc1NzcwMTMwNjU0NDQ1MDI1Ng=="} +{"index":3906,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"Multi-Head Attention is an attention mechanism used in the Transformer architecture for neural networks (Vaswani et al., 2017) .","reference_id":"MTMwOTI5NjY1OTM5NDEwOTExMDQ="} +{"index":3907,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"Example of citation processed text from Attention Is All You Need (Vaswani et al., 2017)","reference_id":"MTY1ODAzODYxNTU2NjgxNDE5Nzg="} +{"index":3908,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key (Vaswani et al., 2017) .","reference_id":"OTQwOTc3MTExMzUwMDY2MjI1Nw=="} +{"index":3909,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1706.03762","reference_contexts":"Galactica uses a Transformer architecture in a decoder-only setup (Vaswani et al., 2017) , with the following modifications:","reference_id":"MTIzMjAwMDU3NDM4MjEwMzk3NDg="} +{"index":3910,"citer_arxiv_id":"2211.09085","cited_arxiv_id":"1810.04805","reference_contexts":"For example, in the BERT model, multi-head attention is used to learn the context-aware representation of the input sequence (Devlin et al., 2019) .","reference_id":"NTMzMTQ1NzMxMDI3MTM4NTQzMw=="} +{"index":3911,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2005.14165","reference_contexts":"Previous works have shown that large language models pre-trained on a colossal amount of data would internalize rich world knowledge that can be probed to perform various downstream tasks [39, 5]","reference_id":"MTQzOTE5NTk5MDcyNzA2MTgxNzE="} +{"index":3912,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2005.14165","reference_contexts":"We find that large GPT-3 [5] and Codex [7] models, when prompted with a single fixed example of a task description and its associated sequence of actions, can produce very plausible action plans for the task we're interested in","reference_id":"NzY3NTU3MDczMTQ2ODA5NDU0MQ=="} +{"index":3913,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2005.14165","reference_contexts":"Notably, autoregressive LLMs can even perform in-context learning, an ability to solve tasks using only contextual information without gradient updates [5]","reference_id":"Nzc3MDU0NzI4NzA2ODM5MDgxNw=="} +{"index":3914,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2005.14165","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3915,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2005.14165","reference_contexts":"Large language models (LLMs) have made impressive advances in language generation and understanding in recent years [10, 39, 40, 5]","reference_id":"MTcwNzY3NDg5MTM0OTg1Njg5NQ=="} +{"index":3916,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1708.00055","reference_contexts":"We compare two size variants of Sentence BERT and Sentence RoBERTa [10, 27, 41] trained on the STS benchmark [6] and a baseline using averaged GloVe embeddings [35]","reference_id":"Mjg4MDU2MDg4OTg2OTU0NTM5NA=="} +{"index":3917,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2107.03374","reference_contexts":"[7] to sample multiple outputs for each query","reference_id":"MTUzMzIyNjc0Njg3NDI2MzM0MTY="} +{"index":3918,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2107.03374","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3919,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2107.03374","reference_contexts":"We find that large GPT-3 [5] and Codex [7] models, when prompted with a single fixed example of a task description and its associated sequence of actions, can produce very plausible action plans for the task we're interested in","reference_id":"NzY3NTU3MDczMTQ2ODA5NDU0MQ=="} +{"index":3920,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2110.14168","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3921,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1810.04805","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3922,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1810.04805","reference_contexts":"To embed the output action phrase and environment actions, we use a BERT-style LM [10, 27] pre-trained with Sentence-BERT [41] objective, to which we refer as Translation LM 3","reference_id":"MTU5MzUxMzU5MzI5Mzc1NzIzMjQ="} +{"index":3923,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1810.04805","reference_contexts":"We compare two size variants of Sentence BERT and Sentence RoBERTa [10, 27, 41] trained on the STS benchmark [6] and a baseline using averaged GloVe embeddings [35]","reference_id":"Mjg4MDU2MDg4OTg2OTU0NTM5NA=="} +{"index":3924,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1810.04805","reference_contexts":"Large language models (LLMs) have made impressive advances in language generation and understanding in recent years [10, 39, 40, 5]","reference_id":"MTcwNzY3NDg5MTM0OTg1Njg5NQ=="} +{"index":3925,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2012.15723","reference_contexts":"This approach bears resemblance to several recent works [37, 13, 26, 43 ]","reference_id":"MTU5MDAxOTE2NjczMDU5MTg5MzE="} +{"index":3926,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2105.09938","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3927,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1904.09751","reference_contexts":"To obtain text completion results, we sample from autoregressive LLM using temperature sampling and nucleus sampling [18]","reference_id":"MTcxOTUwODczODI2NDQ5NTc4OTc="} +{"index":3928,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1712.05474","reference_contexts":"Recently, many efforts have been put into creating more realistic environments with the goal to further advances in this area [38, 47, 48, 22, 44, 1]","reference_id":"MTcwNTY3NDM3OTM1MTg2NTU4MQ=="} +{"index":3929,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2202.01771","reference_contexts":"[24] validates similar hypothesis that LMs contain rich actionable knowledge","reference_id":"Njk3Njc3MDM4NTUwMzEwMjEyMQ=="} +{"index":3930,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2101.06804","reference_contexts":"This approach bears resemblance to several recent works [37, 13, 26, 43 ]","reference_id":"MTU5MDAxOTE2NjczMDU5MTg5MzE="} +{"index":3931,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1907.11692","reference_contexts":"We compare two size variants of Sentence BERT and Sentence RoBERTa [10, 27, 41] trained on the STS benchmark [6] and a baseline using averaged GloVe embeddings [35]","reference_id":"Mjg4MDU2MDg4OTg2OTU0NTM5NA=="} +{"index":3932,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1907.11692","reference_contexts":"To embed the output action phrase and environment actions, we use a BERT-style LM [10, 27] pre-trained with Sentence-BERT [41] objective, to which we refer as Translation LM 3","reference_id":"MTU5MzUxMzU5MzI5Mzc1NzIzMjQ="} +{"index":3933,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1907.11692","reference_contexts":"Then we use an off-the-shelf Sentence-RoBERTa [27, 41] as Translation LM to create embeddings for actions and output text","reference_id":"MTIwNTE2Mzc2NDU3MzIyMjUwNTY="} +{"index":3934,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1907.11692","reference_contexts":"First, we enumerate all admissible actions and map the model's output phrases to the most semantically-similar admissible action (we use similarity measure between sentence embeddings produced by a RoBERTa model [27] in this work, but other choices are possible)","reference_id":"NjM1MDcwOTEwNTU1ODY0NDQwOA=="} +{"index":3935,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2103.05247","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3936,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2005.07648","reference_contexts":"At the same time, by leveraging the better representation power of neural architectures, a number of works have looked into creating instruction-following agents that can perform manipulation [29, 30] , navigation [11, 54, 31] , or both [49, 16, 12]","reference_id":"ODUyNDE5MjU1MjkwMDM3ODA5Nw=="} +{"index":3937,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1909.01066","reference_contexts":"\"Dante was born in PLACE \") can be extracted [36, 21, 9, 50, 42]","reference_id":"MTMxNzczNzg3MTM0ODczNjc3NjI="} +{"index":3938,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1910.10683","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3939,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1910.10683","reference_contexts":"Large language models (LLMs) have made impressive advances in language generation and understanding in recent years [10, 39, 40, 5]","reference_id":"MTcwNzY3NDg5MTM0OTg1Njg5NQ=="} +{"index":3940,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1908.10084","reference_contexts":"To embed the output action phrase and environment actions, we use a BERT-style LM [10, 27] pre-trained with Sentence-BERT [41] objective, to which we refer as Translation LM 3","reference_id":"MTU5MzUxMzU5MzI5Mzc1NzIzMjQ="} +{"index":3941,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1908.10084","reference_contexts":"We compare two size variants of Sentence BERT and Sentence RoBERTa [10, 27, 41] trained on the STS benchmark [6] and a baseline using averaged GloVe embeddings [35]","reference_id":"Mjg4MDU2MDg4OTg2OTU0NTM5NA=="} +{"index":3942,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1908.10084","reference_contexts":"GPT-3 and Codex are accessed using OpenAI API, and the remaining models are accessed through open-source packages, Hugging Face Transformers [55] and SentenceTransformers [41] , all without additional training (except for the fine-tuning baseline).","reference_id":"NjIyNTM2ODcxNDYwMjcwNjc="} +{"index":3943,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1908.10084","reference_contexts":"Then we use an off-the-shelf Sentence-RoBERTa [27, 41] as Translation LM to create embeddings for actions and output text","reference_id":"MTIwNTE2Mzc2NDU3MzIyMjUwNTY="} +{"index":3944,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2002.08910","reference_contexts":"\"Dante was born in PLACE \") can be extracted [36, 21, 9, 50, 42]","reference_id":"MTMxNzczNzg3MTM0ODczNjc3NjI="} +{"index":3945,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2002.08910","reference_contexts":"Being trained on large corpora of human-produced language, these models are thought to contain a lot of information about the world [42, 23, 3] -albeit in linguistic form.","reference_id":"NjA0NzA1Mzk5OTg3NTM2NjEyMg=="} +{"index":3946,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2110.01517","reference_contexts":"Recent works also use language as hierarchical abstractions to plan actions using imitation learning [45] and to guide exploration in reinforcement learning [32] .","reference_id":"MTU4MjkzMTkwNzI0ODAyNjM1MTU="} +{"index":3947,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2109.03034","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3948,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2010.03768","reference_contexts":"Recently, many efforts have been put into creating more realistic environments with the goal to further advances in this area [38, 47, 48, 22, 44, 1]","reference_id":"MTcwNTY3NDM3OTM1MTg2NTU4MQ=="} +{"index":3949,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"2106.13884","reference_contexts":"It has been shown by recent works that large language models (LLMs) pre-trained on large unstructured text corpus not only can perform strongly on various down-stream NLP tasks [10, 39, 40, 5] but the learned representations can also be used to model relations of entities [23] , retrieve matching visual features [19] , synthesize code from docstrings [15, 7] , solve math problems [8, 46] , and even as valuable priors when applied to diverse tasks from different modalities [28, 52]","reference_id":"MTA4NzM5Mjk2NDYxODA1ODgwMjU="} +{"index":3950,"citer_arxiv_id":"2201.07207","cited_arxiv_id":"1910.03771","reference_contexts":"GPT-3 and Codex are accessed using OpenAI API, and the remaining models are accessed through open-source packages, Hugging Face Transformers [55] and SentenceTransformers [41] , all without additional training (except for the fine-tuning baseline).","reference_id":"NjIyNTM2ODcxNDYwMjcwNjc="} +{"index":3951,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1412.3555","reference_contexts":"Two kinds of activation functions have recently gained traction: long short-term memory (LSTM) [9] and the gated recurrent unit (GRU) [3]","reference_id":"OTA4MzA3MDc2Mjg3NzIyMjY5Mw=="} +{"index":3952,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1412.6980","reference_contexts":"Adam algorithm [12] is used for optimization.","reference_id":"MTE1NjUxNjE4MTUxMDA3NzQyMTU="} +{"index":3953,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1411.2539","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":3954,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1410.1090","reference_contexts":"Recently, several approaches based on RNNs emerged, generating captions via a learned joint image-text embedding [13, 11, 36, 21]","reference_id":"MTE2MTQ3MTMxMTUxMjIyMjEwMTg="} +{"index":3955,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1410.1090","reference_contexts":"The field has tackled a diverse set of tasks such as captioning [13, 11, 36, 35, 21] , alignment [11, 15, 34] , Q&A [20, 19] , visual model learning from textual descriptions [8, 26] , and semantic visual search with natural multisentence queries [17] .","reference_id":"NjA1OTIzMjE0OTQwMzA2MzU5MA=="} +{"index":3956,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1301.3781","reference_contexts":"the skip-gram [22] architecture for learning representations of words","reference_id":"MTY2MzE1NDgxMjE2MTQ3ODY5MTg="} +{"index":3957,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1409.4842","reference_contexts":"We used the GoogLeNet architecture [32] as well as hybrid-CNN [38] for extracting frame features","reference_id":"MTU5NDc3ODUyMTU5Nzg5NjU5NjM="} +{"index":3958,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1312.6229","reference_contexts":"These approaches have also been extended to generate descriptions of short video clips [35]","reference_id":"MTI2NjgzNDUxMjYwNzU5OTA5NjU="} +{"index":3959,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1312.6229","reference_contexts":"The field has tackled a diverse set of tasks such as captioning [13, 11, 36, 35, 21] , alignment [11, 15, 34] , Q&A [20, 19] , visual model learning from textual descriptions [8, 26] , and semantic visual search with natural multisentence queries [17] .","reference_id":"NjA1OTIzMjE0OTQwMzA2MzU5MA=="} +{"index":3960,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1411.4555","reference_contexts":"Recently, several approaches based on RNNs emerged, generating captions via a learned joint image-text embedding [13, 11, 36, 21]","reference_id":"MTE2MTQ3MTMxMTUxMjIyMjEwMTg="} +{"index":3961,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1411.4555","reference_contexts":"The field has tackled a diverse set of tasks such as captioning [13, 11, 36, 35, 21] , alignment [11, 15, 34] , Q&A [20, 19] , visual model learning from textual descriptions [8, 26] , and semantic visual search with natural multisentence queries [17] .","reference_id":"NjA1OTIzMjE0OTQwMzA2MzU5MA=="} +{"index":3962,"citer_arxiv_id":"1506.06724","cited_arxiv_id":"1502.03044","reference_contexts":"[37] combines neural embeddings with soft attention in order to align the words to image regions.","reference_id":"OTAyOTgyNzY4MjI2NTQ0NDkz"} +{"index":3963,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1409.0473","reference_contexts":"For many years, machine translation was also achieved by a series of separate tasks (translating words individually, aligning words, reordering, etc), but recent work has shown that translation can be done in a much simpler way using Recurrent Neural Networks (RNNs) [3, 2, 30] and still reach state-of-the-art performance","reference_id":"OTQ5MTM5MDg4OTg3MTg0MTg3NQ=="} +{"index":3964,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1409.0473","reference_contexts":"The model is inspired by recent successes of sequence generation in machine translation [3, 2, 30] , with the difference that instead of starting with a sentence, we provide an image processed by a convolutional net","reference_id":"OTgzOTUyMDM3NjIxMjM1NjI4Mw=="} +{"index":3965,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1308.0850","reference_contexts":"To address this challenge, a particular form of recurrent nets, called LSTM, was introduced [10] and applied with great success to translation [3, 30] and sequence generation [9] .","reference_id":"MTc2OTI2MzI1OTMxOTY0NzYxMjY="} +{"index":3966,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1502.03167","reference_contexts":"In this work we combine deep convolutional nets for image classification [12] with recurrent networks for sequence modeling [10] , to create a single network that generates descriptions of images","reference_id":"MTUyODUxNjgzNDM1NzA2MDgyMDc="} +{"index":3967,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1502.03167","reference_contexts":"Our particular choice of CNN uses a novel approach to batch normalization and yields the current best performance on the ILSVRC 2014 classification competition [12]","reference_id":"MTQzMzQ1NTY0NDA3NTM4NTk0Mjk="} +{"index":3968,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1411.2539","reference_contexts":"[14] propose to construct a joint multimodal embedding space by using a powerful computer vision model and an LSTM that encodes text","reference_id":"ODgxMDQyMDE0Mjk1MDg2MzQ2Ng=="} +{"index":3969,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1411.2539","reference_contexts":"We computed these BLEU scores with the outputs that the authors of [14] kindly provided for their OxfordNet system.","reference_id":"MTgzNzk4NzU5Njg0Nzk2MTc1MDI="} +{"index":3970,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1411.2539","reference_contexts":"The Flickr datasets have been used recently [11, 21, 14] , but mostly evaluated in a retrieval framework","reference_id":"MzQwNTcwMTMyOTk5NDg4ODI0NQ=="} +{"index":3971,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1411.2539","reference_contexts":"11 TreeTalk [18] 19 BabyTalk [16] 25 Tri5Sem [11] 48 m-RNN [21] 55 58 MNLM [14] is needed towards better metrics","reference_id":"ODQyOTM4MTcwNTA4MTE1MDMwNw=="} +{"index":3972,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1411.2539","reference_contexts":"Lastly, the current literature on image description has also been using the proxy task of ranking a set of available descriptions with respect to a given image (see for instance [14] )","reference_id":"MTI5OTMyMTE1Mzg1Mjg0MzkwNzQ="} +{"index":3973,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1405.0312","reference_contexts":"test Pascal VOC 2008 [6] --1000 Flickr8k [26] 6000 1000 1000 Flickr30k [33] 28000 1000 1000 MSCOCO [20] 82783 40504 40775 SBU [24] 1M --With the exception of SBU, each image has been annotated by labelers with 5 sentences that are relatively visual and unbiased","reference_id":"MTY1MjY5NjE1MzU4NjI3OTk3Ng=="} +{"index":3974,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1410.1090","reference_contexts":"Note that for the Image Annotation task, we normalized our scores similar to what [21] used.","reference_id":"OTc5NDM5NjY2OTQxOTQ4MjYwMQ=="} +{"index":3975,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1410.1090","reference_contexts":"The Flickr datasets have been used recently [11, 21, 14] , but mostly evaluated in a retrieval framework","reference_id":"MzQwNTcwMTMyOTk5NDg4ODI0NQ=="} +{"index":3976,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1410.1090","reference_contexts":"11 TreeTalk [18] 19 BabyTalk [16] 25 Tri5Sem [11] 48 m-RNN [21] 55 58 MNLM [14] is needed towards better metrics","reference_id":"ODQyOTM4MTcwNTA4MTE1MDMwNw=="} +{"index":3977,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1410.1090","reference_contexts":"[21] uses a recurrent NN for the same prediction task","reference_id":"NDkwMzQxNzU2MzYxOTgyMzAwNQ=="} +{"index":3978,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1410.1090","reference_contexts":"A notable exception is [21] , where they did both retrieval and generation, and which yields the best performance on the Flickr datasets up to now.","reference_id":"OTkzMzYwMjgxOTkyNDM5Mjk3MA=="} +{"index":3979,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1312.6229","reference_contexts":"Over the last few years it has been convincingly shown that CNNs can produce a rich representation of the input image by embedding it to a fixed-length vector, such that this representation can be used for a variety of vision tasks [28]","reference_id":"ODk0MTg4OTA3ODE3MDY0OTcyOQ=="} +{"index":3980,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1411.5726","reference_contexts":"Despite recent efforts on better evaluation metrics [31] , our model fares strongly versus human raters","reference_id":"ODQ5MTQwNjgzMTkxOTQ1NjM2"} +{"index":3981,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1411.5726","reference_contexts":"A much more detailed discussion regarding metrics can be found in [31] , and research groups working on this topic have been reporting other metrics which are deemed more appropriate for evaluating caption","reference_id":"MTMwMzU4NzUwNTQ4Nzc1NDI5ODI="} +{"index":3982,"citer_arxiv_id":"1411.4555","cited_arxiv_id":"1409.2329","reference_contexts":"We tried dropout [34] and ensembling models, as well as exploring the size (i.e., capacity) of the model by trading off number of hidden units versus depth","reference_id":"MTM5NTgyMjU3NzM1MzU4NjY4MTI="} +{"index":3983,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1409.0473","reference_contexts":"Artificial neural networks have enjoyed a spectacularly successful decade, having made considerable advances in computer vision [1, 2] , translation [3, 4, 5] , speech recognition [6, 7] , image generation [8, 9, 10, 11, 12] , game playing [13, 14, 15] , and robotics [16, 17]","reference_id":"MTI2Mjg5MjAzODQxNDgwMzgwMzE="} +{"index":3984,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1609.08144","reference_contexts":"Artificial neural networks have enjoyed a spectacularly successful decade, having made considerable advances in computer vision [1, 2] , translation [3, 4, 5] , speech recognition [6, 7] , image generation [8, 9, 10, 11, 12] , game playing [13, 14, 15] , and robotics [16, 17]","reference_id":"MTI2Mjg5MjAzODQxNDgwMzgwMzE="} +{"index":3985,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1511.06434","reference_contexts":"Artificial neural networks have enjoyed a spectacularly successful decade, having made considerable advances in computer vision [1, 2] , translation [3, 4, 5] , speech recognition [6, 7] , image generation [8, 9, 10, 11, 12] , game playing [13, 14, 15] , and robotics [16, 17]","reference_id":"MTI2Mjg5MjAzODQxNDgwMzgwMzE="} +{"index":3986,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1710.10196","reference_contexts":"Artificial neural networks have enjoyed a spectacularly successful decade, having made considerable advances in computer vision [1, 2] , translation [3, 4, 5] , speech recognition [6, 7] , image generation [8, 9, 10, 11, 12] , game playing [13, 14, 15] , and robotics [16, 17]","reference_id":"MTI2Mjg5MjAzODQxNDgwMzgwMzE="} +{"index":3987,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1912.06680","reference_contexts":"Artificial neural networks have enjoyed a spectacularly successful decade, having made considerable advances in computer vision [1, 2] , translation [3, 4, 5] , speech recognition [6, 7] , image generation [8, 9, 10, 11, 12] , game playing [13, 14, 15] , and robotics [16, 17]","reference_id":"MTI2Mjg5MjAzODQxNDgwMzgwMzE="} +{"index":3988,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1910.07113","reference_contexts":"Artificial neural networks have enjoyed a spectacularly successful decade, having made considerable advances in computer vision [1, 2] , translation [3, 4, 5] , speech recognition [6, 7] , image generation [8, 9, 10, 11, 12] , game playing [13, 14, 15] , and robotics [16, 17]","reference_id":"MTI2Mjg5MjAzODQxNDgwMzgwMzE="} +{"index":3989,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"2005.14165","reference_contexts":"Language modeling improvements have been demonstrated from better pre-training tasks, using various objectives such as auto-regressive generation [19, 20, 21] , token masking [22] or sequence masking [36] , but resulting language models have so far felt short when applied to reasoning oriented tasks such as algebraic word problems [37, 38]","reference_id":"MTE4MTMwNTEyODU0NDM4MTYyMTc="} +{"index":3990,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"2005.14165","reference_contexts":"We use decoder-only Transformers similar to GPT-2 [20] and GPT-3 [21]","reference_id":"MTUwMjYxMjEzNjczMDM0Mjg2OA=="} +{"index":3991,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"2005.14165","reference_contexts":"Models are pre-trained on CommonCrawl using GPT-3's [21] methodology for 260B tokens","reference_id":"NTc5OTEwMTU3ODU4MDU5MTkxMw=="} +{"index":3992,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"2005.14165","reference_contexts":"Especially notable is the recent rapid progress in language understanding and generation capabilities [18, 19, 20, 21, 22] .","reference_id":"MTAzNTMzNTMyMjYyODc4NzI2MzY="} +{"index":3993,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"2005.14165","reference_contexts":"The models are trained with the BPE encoding reported in [21] , the same tokenization being used for text, code or formalized statements","reference_id":"MTA5Mzg5MjczNDMxMDc3MzA="} +{"index":3994,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1810.04805","reference_contexts":"Especially notable is the recent rapid progress in language understanding and generation capabilities [18, 19, 20, 21, 22] .","reference_id":"MTAzNTMzNTMyMjYyODc4NzI2MzY="} +{"index":3995,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1810.04805","reference_contexts":"Language modeling improvements have been demonstrated from better pre-training tasks, using various objectives such as auto-regressive generation [19, 20, 21] , token masking [22] or sequence masking [36] , but resulting language models have so far felt short when applied to reasoning oriented tasks such as algebraic word problems [37, 38]","reference_id":"MTE4MTMwNTEyODU0NDM4MTYyMTc="} +{"index":3996,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1712.01815","reference_contexts":"With the possible exception of AlphaGo [13] and AlphaZero [23] , reasoning tasks are conspicuously absent from the list above","reference_id":"MTYxNzUxNzYyNzYyNTI2Mzk1NzA="} +{"index":3997,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1806.00608","reference_contexts":"GamePad [33] and CoqGymn/ASTactic [34] introduce environments based on the Coq theorem prover","reference_id":"NDY1MDI0MDg0MzYwMDUxODk5OQ=="} +{"index":3998,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1910.10683","reference_contexts":"Language modeling improvements have been demonstrated from better pre-training tasks, using various objectives such as auto-regressive generation [19, 20, 21] , token masking [22] or sequence masking [36] , but resulting language models have so far felt short when applied to reasoning oriented tasks such as algebraic word problems [37, 38]","reference_id":"MTE4MTMwNTEyODU0NDM4MTYyMTc="} +{"index":3999,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1705.04146","reference_contexts":"Language modeling improvements have been demonstrated from better pre-training tasks, using various objectives such as auto-regressive generation [19, 20, 21] , token masking [22] or sequence masking [36] , but resulting language models have so far felt short when applied to reasoning oriented tasks such as algebraic word problems [37, 38]","reference_id":"MTE4MTMwNTEyODU0NDM4MTYyMTc="} +{"index":4000,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1905.13319","reference_contexts":"Language modeling improvements have been demonstrated from better pre-training tasks, using various objectives such as auto-regressive generation [19, 20, 21] , token masking [22] or sequence masking [36] , but resulting language models have so far felt short when applied to reasoning oriented tasks such as algebraic word problems [37, 38]","reference_id":"MTE4MTMwNTEyODU0NDM4MTYyMTc="} +{"index":4001,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1807.03819","reference_contexts":"The Universal Transformer [40] , a Transformer with tied weights, was also shown to be successful at more algorithmic tasks","reference_id":"ODc4NDAwMTgxNzQ1NjIwNzU2Ng=="} +{"index":4002,"citer_arxiv_id":"2009.03393","cited_arxiv_id":"1904.01557","reference_contexts":"[41] evaluated the Transformer architecture on a variety of mathematical problems.","reference_id":"OTQwODA0OTYzNDg4NDU4NDYzOQ=="} +{"index":4003,"citer_arxiv_id":"2004.04906","cited_arxiv_id":"2002.08909","reference_contexts":"Their approach outperforms the BM25 plus reader paradigm on multiple open-domain QA datasets in QA accuracy, and is further extended by REALM (Guu et al., 2020) , which includes tuning the passage encoder asynchronously by re-indexing the passages during training","reference_id":"MTA2NTA0NzMwNDA0MjA4MTk3ODQ="} +{"index":4004,"citer_arxiv_id":"2004.04906","cited_arxiv_id":"2002.08909","reference_contexts":"It is interesting to contrast our results to those of ORQA (Lee et al., 2019) and also the concurrently developed approach, REALM (Guu et al., 2020)","reference_id":"MTIyMjk4MzU0ODg3NjAwNTczNjc="} +{"index":4005,"citer_arxiv_id":"2004.04906","cited_arxiv_id":"1705.00652","reference_contexts":"The trick of in-batch negatives has been used in the full batch setting (Yih et al., 2011) and more recently for mini-batch (Henderson et al., 2017; Gillick et al., 2019)","reference_id":"MTI4NTczMjQxODg5OTk2NzY3MjE="} +{"index":4006,"citer_arxiv_id":"2004.04906","cited_arxiv_id":"2007.01282","reference_contexts":"Recent work (Izacard and Grave, 2020; Lewis et al., 2020b) have also shown that DPR can be combined with generation models such as BART (Lewis et al., 2020a) and T5 (Raffel et al., 2019) , achieving good performance on open-domain QA and other knowledge-intensive tasks.","reference_id":"MTgwOTUyMzc5NzAwNzUxNTUyMQ=="} +{"index":4007,"citer_arxiv_id":"2004.04906","cited_arxiv_id":"1702.08734","reference_contexts":"Inference During inference time, we apply the passage encoder E P to all the passages and index them using FAISS (Johnson et al., 2017) offline","reference_id":"MTQ1MTc2NTI1MTI4NjI5ODgwOTk="} +{"index":4008,"citer_arxiv_id":"2004.04906","cited_arxiv_id":"1910.10683","reference_contexts":"Recent work (Izacard and Grave, 2020; Lewis et al., 2020b) have also shown that DPR can be combined with generation models such as BART (Lewis et al., 2020a) and T5 (Raffel et al., 2019) , achieving good performance on open-domain QA and other knowledge-intensive tasks.","reference_id":"MTgwOTUyMzc5NzAwNzUxNTUyMQ=="} +{"index":4009,"citer_arxiv_id":"2004.04906","cited_arxiv_id":"2002.08910","reference_contexts":"Exceptions include (Seo et al., 2019) and (Roberts et al., 2020) , which retrieves and generates the answers, respectively.","reference_id":"NDU2NzE1OTkwOTUzODIxMDY2Ng=="} +{"index":4010,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2301.13826","reference_contexts":"A14 , when Cones incorporates the Attend-and-Excite (Feng et al., 2022; Chefer et al., 2023) method to address this issue, it generates better results.","reference_id":"MjYzMjU4ODUyMjYyNTg1MTA5Mw=="} +{"index":4011,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.05032","reference_contexts":"A14 , when Cones incorporates the Attend-and-Excite (Feng et al., 2022; Chefer et al., 2023) method to address this issue, it generates better results.","reference_id":"MjYzMjU4ODUyMjYyNTg1MTA5Mw=="} +{"index":4012,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.01618","reference_contexts":"They are Dreambooth (Ruiz et al., 2022) that fine-tunes all parameters in the diffusion model; Text Inversion (Gal et al., 2022 ) that adds a new token for each new concept and only updates the new token embedding during fine-tuning; and Custom Diffusion (Kumari et al., 2022) that optimizes the newly added token embedding in text encoder and a few parameters in diffusion model, namely the key and value mapping from text to latent features in the cross-attention (Yu et al., 2022; Vaswani et al., 2017) layers.","reference_id":"MzQ4Mjg3NTY0MjkwMDcxMTg1NA=="} +{"index":4013,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.01618","reference_contexts":"We evaluate Cones with two famous metrics for customized generation proposed in Textual Inversion (Gal et al., 2022)","reference_id":"Mzk5NTI2NDE5OTE1MDcxODMzNg=="} +{"index":4014,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2207.12598","reference_contexts":"The second one is known as the classifier-free approach (Ho & Salimans, 2022) , which directly collects a large amount of data pairs for joint optimization under the guarantee of conditional probabilistic derivation","reference_id":"MTQ3MzE5NTM4Mzc4MDM1NDI3Mjg="} +{"index":4015,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"Our purpose here is to locate the corresponding neurons that control the generation of the given subject and use those neurons to guide customized generation (Ruiz et al., 2022; Kumari et al., 2022)","reference_id":"MTgwMTY4OTQ2NTc5ODAwMDk1MTg="} +{"index":4016,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"Consistent with the Custom Diffusion (Kumari et al., 2022) setting, we use Stable Diffusion V1.4 as the pretrained model","reference_id":"MTYyMTc3ODE5MzM4NDQzNTM2MzQ="} +{"index":4017,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"Storing the information to construct a given subject costs around only 10% of memory compared with previous subjectdriven methods (Ruiz et al., 2022; Kumari et al., 2022) , which is extremely economical and environment-friendly for commercial usage in mobile devices","reference_id":"MzQ2NDI1MTg4OTY1NzI1NDgyMA=="} +{"index":4018,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"They are Dreambooth (Ruiz et al., 2022) that fine-tunes all parameters in the diffusion model; Text Inversion (Gal et al., 2022 ) that adds a new token for each new concept and only updates the new token embedding during fine-tuning; and Custom Diffusion (Kumari et al., 2022) that optimizes the newly added token embedding in text encoder and a few parameters in diffusion model, namely the key and value mapping from text to latent features in the cross-attention (Yu et al., 2022; Vaswani et al., 2017) layers.","reference_id":"MzQ4Mjg3NTY0MjkwMDcxMTg1NA=="} +{"index":4019,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"As mentioned in Custom Diffusion (Kumari et al., 2022) , after fine-tuning the models, the generations shift towards the target images and have less diversity compared to the pretrained model with the prompt \"photo of a V 1 * class\"","reference_id":"MTMxMTE2NDAxNTMwODQ2OTA5MzQ="} +{"index":4020,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"As mentioned in Custom Diffusion (Kumari et al., 2022) , the pretrained model encounters difficulty generating multiple subjects described in a single text prompt","reference_id":"MTE3Nzk1NjU0MTYyNzA5MTYwMDQ="} +{"index":4021,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"While previous research shows that the K-V attention layer (Kumari et al., 2022) dominates the subject generation process, we follow its setting and limit the search for concept neurons in the parameters of the K-V attention layers","reference_id":"MTIyNTE2NDY2NjYwMDM2NDU4NzQ="} +{"index":4022,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"Thanks to the sparsity and robust binary representation of concept neurons, storing concept neurons consumes only around 10% memory of Custom Diffusion (Kumari et al., 2022) and 0.05% memory of Dreembooth (Ruiz et al., 2022)","reference_id":"MjIzOTE0MDQ3ODU0NTE4MzIxMg=="} +{"index":4023,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"Previous customized generation methods (Ruiz et al., 2022; Kumari et al., 2022) demand to save the parameters of the diffusion model in full digital accuracy","reference_id":"MTM3MzM0MDAyMjA1MDU1NTMxNw=="} +{"index":4024,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"Specifically, to mimic the imaginative ability of the human brain, do generative diffusion models (Ho et al., 2020; Dhariwal & Nichol, 2021) encode different subjects separately with their neurons? This paper is about to answer this question from the perspective of subject-driven generation (Kumari et al., 2022; Ruiz et al., 2022)","reference_id":"MTU0MzY3MzIxMzA3NjU1MTU4NzQ="} +{"index":4025,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"This could be due to that the image alignment metric is easy to overfit as is pointed out in Custom Diffusion (Kumari et al., 2022)","reference_id":"MTgwNjMxNjY3MTU2MzMxNjc3NjE="} +{"index":4026,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2212.04488","reference_contexts":"All images used in the paper are downloaded from anonymous ecommerce websites or Unsplash, like the dataset of Custom Diffusion (Kumari et al., 2022)","reference_id":"MTYwOTU4MzIwOTA2NTM1NzgzMjU="} +{"index":4027,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2301.07093","reference_contexts":"However, the expressiveness of text is still limited, and more work wants to utilize additional, conditional information (e.g., reference image, grounding (Li et al., 2023) and sketch (Voynov et al., 2022) ) to guide the global control further.","reference_id":"MTU4MTY5MzM2MzM1ODYyNjU0OA=="} +{"index":4028,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2206.00927","reference_contexts":"For a fair comparison, we use 50 steps of DPM-Solver (Lu et al., 2022) sampler with a scale 7.5 for all above methods","reference_id":"MTM5Njk4NTg0ODE4MTIyMDQ5MDY="} +{"index":4029,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2112.10741","reference_contexts":"Owing to advances in language (Radford et al., 2021) and cross-modal foundation models (Radford et al., 2021) , much text-to-image work (Saharia et al., 2022; Ramesh et al., 2022; Nichol et al., 2021) with classifier-free techniques is beginning to emerge, facilitating explicit control on the corresponding semantics and style.","reference_id":"NjgxNTk3MzU1MDQ3MTgwMzk1NQ=="} +{"index":4030,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2204.06125","reference_contexts":"Owing to advances in language (Radford et al., 2021) and cross-modal foundation models (Radford et al., 2021) , much text-to-image work (Saharia et al., 2022; Ramesh et al., 2022; Nichol et al., 2021) with classifier-free techniques is beginning to emerge, facilitating explicit control on the corresponding semantics and style.","reference_id":"NjgxNTk3MzU1MDQ3MTgwMzk1NQ=="} +{"index":4031,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2204.06125","reference_contexts":"However, Cones can also be simply applied to most text-to-image diffusion models like Imagen (Saharia et al., 2022) and DALLE-2 (Ramesh et al., 2022) .","reference_id":"MTM2OTIyNDk4MDMyMjk3MzgwNjU="} +{"index":4032,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"Our purpose here is to locate the corresponding neurons that control the generation of the given subject and use those neurons to guide customized generation (Ruiz et al., 2022; Kumari et al., 2022)","reference_id":"MTgwMTY4OTQ2NTc5ODAwMDk1MTg="} +{"index":4033,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"We optimize a concept-implanting loss (Ruiz et al., 2022) on the concept neurons using float32, float16, quaternary, and binary (shutting those concept neurons directly without training) digital accuracy correspondingly","reference_id":"NDgzNDIxNzQyMTI2NDc3MTQxNA=="} +{"index":4034,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"Specifically, to mimic the imaginative ability of the human brain, do generative diffusion models (Ho et al., 2020; Dhariwal & Nichol, 2021) encode different subjects separately with their neurons? This paper is about to answer this question from the perspective of subject-driven generation (Kumari et al., 2022; Ruiz et al., 2022)","reference_id":"MTU0MzY3MzIxMzA3NjU1MTU4NzQ="} +{"index":4035,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"Storing the information to construct a given subject costs around only 10% of memory compared with previous subjectdriven methods (Ruiz et al., 2022; Kumari et al., 2022) , which is extremely economical and environment-friendly for commercial usage in mobile devices","reference_id":"MzQ2NDI1MTg4OTY1NzI1NDgyMA=="} +{"index":4036,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"To avoid over-fitting and language-drift (Ruiz et al., 2022) , we further need the prior-preserving loss","reference_id":"MzI5MTgyODU1NDgxMTY5MTg3MQ=="} +{"index":4037,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"They are Dreambooth (Ruiz et al., 2022) that fine-tunes all parameters in the diffusion model; Text Inversion (Gal et al., 2022 ) that adds a new token for each new concept and only updates the new token embedding during fine-tuning; and Custom Diffusion (Kumari et al., 2022) that optimizes the newly added token embedding in text encoder and a few parameters in diffusion model, namely the key and value mapping from text to latent features in the cross-attention (Yu et al., 2022; Vaswani et al., 2017) layers.","reference_id":"MzQ4Mjg3NTY0MjkwMDcxMTg1NA=="} +{"index":4038,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"Thanks to the sparsity and robust binary representation of concept neurons, storing concept neurons consumes only around 10% memory of Custom Diffusion (Kumari et al., 2022) and 0.05% memory of Dreembooth (Ruiz et al., 2022)","reference_id":"MjIzOTE0MDQ3ODU0NTE4MzIxMg=="} +{"index":4039,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"Previous customized generation methods (Ruiz et al., 2022; Kumari et al., 2022) demand to save the parameters of the diffusion model in full digital accuracy","reference_id":"MTM3MzM0MDAyMjA1MDU1NTMxNw=="} +{"index":4040,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2208.12242","reference_contexts":"The purpose of customized generation, as first proposed in DreamBooth (Ruiz et al., 2022) , is to implant a given subject into the diffusion model and bind it with a unique text identifier to indicate its presence; so that the model can generate various renditions of the subject vividly guided by text prompts (Lu et al., 2020; Lee et al., 2019)","reference_id":"OTY1MTIxOTY3MTg1NjExMjY1NQ=="} +{"index":4041,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2205.11487","reference_contexts":"Owing to advances in language (Radford et al., 2021) and cross-modal foundation models (Radford et al., 2021) , much text-to-image work (Saharia et al., 2022; Ramesh et al., 2022; Nichol et al., 2021) with classifier-free techniques is beginning to emerge, facilitating explicit control on the corresponding semantics and style.","reference_id":"NjgxNTk3MzU1MDQ3MTgwMzk1NQ=="} +{"index":4042,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2205.11487","reference_contexts":"However, Cones can also be simply applied to most text-to-image diffusion models like Imagen (Saharia et al., 2022) and DALLE-2 (Ramesh et al., 2022) .","reference_id":"MTM2OTIyNDk4MDMyMjk3MzgwNjU="} +{"index":4043,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2205.11487","reference_contexts":"A text-to-image diffusion model (Yu et al., 2022; Saharia et al., 2022; Rombach et al., 2022) xθ will guide this denoising procedure with a text prompt describing the image content","reference_id":"MTM5MzI1NzI4NTI3NjQyODYwOTc="} +{"index":4044,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2011.13456","reference_contexts":"Diffusion models (Ho et al., 2020; Dhariwal & Nichol, 2021; Rombach et al., 2022; Song et al., 2020) are parametric neural networks that learn image distributions by gradual denoising","reference_id":"NDkzMDAwNjQ0MjM0ODI0ODQ4NQ=="} +{"index":4045,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2211.13752","reference_contexts":"However, the expressiveness of text is still limited, and more work wants to utilize additional, conditional information (e.g., reference image, grounding (Li et al., 2023) and sketch (Voynov et al., 2022) ) to guide the global control further.","reference_id":"MTU4MTY5MzM2MzM1ODYyNjU0OA=="} +{"index":4046,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2206.10789","reference_contexts":"They are Dreambooth (Ruiz et al., 2022) that fine-tunes all parameters in the diffusion model; Text Inversion (Gal et al., 2022 ) that adds a new token for each new concept and only updates the new token embedding during fine-tuning; and Custom Diffusion (Kumari et al., 2022) that optimizes the newly added token embedding in text encoder and a few parameters in diffusion model, namely the key and value mapping from text to latent features in the cross-attention (Yu et al., 2022; Vaswani et al., 2017) layers.","reference_id":"MzQ4Mjg3NTY0MjkwMDcxMTg1NA=="} +{"index":4047,"citer_arxiv_id":"2303.05125","cited_arxiv_id":"2206.10789","reference_contexts":"A text-to-image diffusion model (Yu et al., 2022; Saharia et al., 2022; Rombach et al., 2022) xθ will guide this denoising procedure with a text prompt describing the image content","reference_id":"MTM5MzI1NzI4NTI3NjQyODYwOTc="} +{"index":4048,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1802.01548","reference_contexts":"While often computationally demanding, evolutionary search has had notable successes which include finding good memory cells for sequence models [27] , [28] and learning sophisticated architectures for largescale image classification [29] , [30] , [31]","reference_id":"NDU3Mjg1MDIyMDUwMTIyMTA3OQ=="} +{"index":4049,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1802.01548","reference_contexts":"This method uses reinforcement learning to develop new policies for data augmentation during training to improve the performance of the architecture searched by [31]","reference_id":"NzM4MTc3Mzg4NTI0MjQ3NjkwNw=="} +{"index":4050,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1804.09081","reference_contexts":"With the goal of reducing the computational burden of these methods, efficient alternatives to this approach have been proposed based on Lamarckian inheritance [32] and differentiable architecture search [33] .","reference_id":"MTQzMDQxOTM3MTc1MzExMzcwMjI="} +{"index":4051,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1806.09055","reference_contexts":"With the goal of reducing the computational burden of these methods, efficient alternatives to this approach have been proposed based on Lamarckian inheritance [32] and differentiable architecture search [33] .","reference_id":"MTQzMDQxOTM3MTc1MzExMzcwMjI="} +{"index":4052,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1704.08792","reference_contexts":"By formulating architecture search as hyperparameter optimisation, random search [34] and other more sophisticated model-based optimisation techniques [35] , [36] can also be used to tackle the problem","reference_id":"OTIxNjY4Nzk1ODcyODk2MTY2Ng=="} +{"index":4053,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1706.06905","reference_contexts":"Attention mechanisms have demonstrated their utility across many tasks including sequence learning [52] , [53] , localisation and understanding in images [9] , [54] , image captioning [55] , [56] and lip reading [57]","reference_id":"MTExNDA3NDc0MTA5NjkxODQ4NzQ="} +{"index":4054,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1704.04861","reference_contexts":"MobileNet refers to \"1.0 MobileNet-224\" in [64] and ShuffleNet refers to \"ShuffleNet 1 × (g = 3)\" in [65]","reference_id":"MzE2MTEyNjc2NzI1Njc2MDM3NQ=="} +{"index":4055,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1704.04861","reference_contexts":"Further variants that integrate SE blocks with ResNeXt [19] , Inception-ResNet [21] , MobileNet [64] and ShuffleNet [65] can be constructed by following similar schemes","reference_id":"MTAyMzcyMzM4NzE3ODM1NDYzMDU="} +{"index":4056,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1704.04861","reference_contexts":"Finally, we consider two representative architectures from the class of mobile-optimised networks, MobileNet [64] and ShuffleNet [65]","reference_id":"MTY4NzU4NTU2MzM2NTEyODYxOTk="} +{"index":4057,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1705.07485","reference_contexts":"We perform experiments with several popular baseline architectures and techniques (ResNet-110 [14] , ResNet-164 [14] , WideResNet-16-8 [67] , Shake-Shake [68] and Cutout [69] ) on the CIFAR-10 and CIFAR-100 datasets [70]","reference_id":"MTU2MTQyNTQ1Nzk0MTAzMDI4OTc="} +{"index":4058,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1708.04552","reference_contexts":"We perform experiments with several popular baseline architectures and techniques (ResNet-110 [14] , ResNet-164 [14] , WideResNet-16-8 [67] , Shake-Shake [68] and Cutout [69] ) on the CIFAR-10 and CIFAR-100 datasets [70]","reference_id":"MTU2MTQyNTQ1Nzk0MTAzMDI4OTc="} +{"index":4059,"citer_arxiv_id":"1709.01507","cited_arxiv_id":"1805.09501","reference_contexts":"The best performance using only Im-ageNet data was recently reported by [79]","reference_id":"MTg2Nzk2MTM1NzU0MDI5MzMyNw=="} +{"index":4060,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00704","reference_contexts":"This is a slightly modified version of the architecture from [1]","reference_id":"MTMyOTk5NzgwODg2NjgxNTAzNDY="} +{"index":4061,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1412.6515","reference_contexts":"[3] identifies the generator objective that corresponds to minimizing the KL divergence, but does not empirically evaluate this objective.","reference_id":"NDQxMzI3MjM4NjA0MDIxMTM4Mg=="} +{"index":4062,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1610.03483","reference_contexts":"[7] discusses many approaches similar to the one presented here, but do not present experimental results.","reference_id":"NzE2MDcwMDU5OTEzMjQ3NzA3NA=="} +{"index":4063,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"Recent work has generalized the GAN theory to target reverse KL [14] and additional f -divergences (including KL, reverse KL, and JS), allowing GANs to target a diverse set of behaviors [9]","reference_id":"MTUzODA4NjYxNTA1NjYwODAwOTc="} +{"index":4064,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"Recently, [9] proposed an extension to GANs to target divergences other than Jensen-Shannon","reference_id":"MTQ4MDEyMzM0OTkwNzM2MDQ2OTM="} +{"index":4065,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"The second approach is empirically motivated in [2, 9] as speeding up training, and theoretically motivated by the observation that p = q remains a fixed point of the learning dynamics","reference_id":"NzE2MDAyMDk0ODI4MjI3MzgzNQ=="} +{"index":4066,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"As in [9] , we parameterize the discriminator as T (x) = g f (V (x)) where g f has the same range as f D","reference_id":"NzY5MDQ4NjYwNzg3OTg1MjM5Mg=="} +{"index":4067,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"The discriminator in all models was trained to optimize the normal GAN objective, corresponding to maximizing Equation 4 with f D (u) = u log u -(u + 1) log(u + 1), and using T (x) = g f (V (x)) with V (x) ∈ R being the output of a neural network and g f (v) = -log(1 + exp(v)) being used to constrain the range of T as in [9]","reference_id":"NzgzMzY0MjQzODcyNTI0Nzk2Ng=="} +{"index":4068,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"We use q as the data distribution and p as the model distribution, which is the opposite of [9] .","reference_id":"MTMzNzg1MzY0MTA2NzUwMTIwMTk="} +{"index":4069,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"Given known data and model densities, the optimal discriminator with respect to an f -divergence, f D , was derived in [9] as:","reference_id":"OTUyNzE3NDU4MTQyNjY2Mzc2OQ=="} +{"index":4070,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.00709","reference_contexts":"In both the original GAN paper [2] and the f -GAN paper [9] , two objectives are proposed (denoted as G LB and G ALT ):","reference_id":"MTM1MTE0NTIxMjg5Nzk5NjU2NTQ="} +{"index":4071,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1610.09585","reference_contexts":"Generative adversarial networks (GANs) have become a popular method for fitting latent-variable directed generative models to complex datasets [2, 10, 13, 6]","reference_id":"MTI5NzIyNTE2NjM1NTczNjk3ODE="} +{"index":4072,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1511.06434","reference_contexts":"Many recent papers have focused on new architectures and regularization techniques for improved stability and performance [13, 11, 4] , but the objectives they optimize are fundamentally the same as the objectives in the original proposal [2] .","reference_id":"OTc2NzQ2Mzg2NTUyMDI1ODk2MQ=="} +{"index":4073,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.03498","reference_contexts":"Generative adversarial networks (GANs) have become a popular method for fitting latent-variable directed generative models to complex datasets [2, 10, 13, 6]","reference_id":"MTI5NzIyNTE2NjM1NTczNjk3ODE="} +{"index":4074,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1606.03498","reference_contexts":"Many recent papers have focused on new architectures and regularization techniques for improved stability and performance [13, 11, 4] , but the objectives they optimize are fundamentally the same as the objectives in the original proposal [2] .","reference_id":"OTc2NzQ2Mzg2NTUyMDI1ODk2MQ=="} +{"index":4075,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1610.02920","reference_contexts":"This approach requires learning a network that directly outputs the density ratio, which can be very small or very large and in practice the networks that parameterize the density ratio must be clipped [16]","reference_id":"MTY4OTQ5NzU5OTQ1MDM3ODE5NTk="} +{"index":4076,"citer_arxiv_id":"1612.02780","cited_arxiv_id":"1610.02920","reference_contexts":"In [16] , they directly estimate the density ratio by optimizing a different discriminator objective that corresponds to rewriting the discriminator in terms of the density ratio:","reference_id":"MTM2MTQ1NTA5NzIzOTQxNzY2MzA="} +{"index":4077,"citer_arxiv_id":"1412.5567","cited_arxiv_id":"1409.3215","reference_contexts":"[42] to accelerate RNNs for text.","reference_id":"ODk4MTYxMDg4NzAyOTEyMzgwOQ=="} +{"index":4078,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2111.10952","reference_contexts":"A range of prior work has shown that finetuning language models on a collection of NLP tasks can improve downstream performance on a broad range of downstream tasks (Aghajanyan et al., 2021; Aribandi et al., 2022; Wei et al., 2021; Sanh et al., 2022; Ouyang et al., 2022, inter alia)","reference_id":"MzI4NzM1NDc1NTIyMDg1OTM2OQ=="} +{"index":4079,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2111.10952","reference_contexts":"The majority of this prior work, however, requires additional data such as aggregating dozens or hundreds of NLP datasets (Raffel et al., 2019; Aghajanyan et al., 2021; Aribandi et al., 2022) , writing additional templates of instructions (Wei et al., 2021; Sanh et al., 2022) , or finetuning on human-labeled annotations (Ouyang et al., 2022)","reference_id":"NDcwMzg1Mjk4OTU3MDgzMjExNg=="} +{"index":4080,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2005.14165","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4081,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2204.02311","reference_contexts":"The key idea is to continue training an existing causal language model (Chowdhery et al., 2022) with a mixture of new objectives-specifically, the UL2 training objective mixture (Tay et al., 2022b)","reference_id":"NTYzMTcyNDQ1NjkxMjgwMTUyOA=="} +{"index":4082,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2204.02311","reference_contexts":"Large language models Scaling and improving large language models is one of the most impactful research areas in modern artificial intelligence (Chowdhery et al., 2022)","reference_id":"NzQ1NDA5MDg1OTY5MzUxMTQyOA=="} +{"index":4083,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2204.02311","reference_contexts":"This is also backed by the continued training of PaLM 62B in (Chowdhery et al., 2022) which showed that repeated data could result in small gains, albeit not as strong as fresh tokens","reference_id":"MTU4NjExMzYzNjMwMTc0MDY1NTQ="} +{"index":4084,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2204.02311","reference_contexts":"To this end, we postulate that imbuing a state-of-the-art large language model such as PaLM (Chowdhery et al., 2022) with these diverse pretraining schemes as a complement to the original language model objective, enables the model to perform significantly better","reference_id":"NzE5ODkyNDQ2Mjk3MjM3MTA1Nw=="} +{"index":4085,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2204.02311","reference_contexts":"For example, flagship large language models such as GPT-3 (Brown et al., 2020 ), Gopher (Rae et al., 2021) and PaLM (Chowdhery et al., 2022) are all trained as causal language models","reference_id":"NTQ2MjcyMjc1ODQxMjUyNzgxMQ=="} +{"index":4086,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2204.02311","reference_contexts":"There has been significant interest in scaling of language models (Rae et al., 2021; Chowdhery et al., 2022; Brown et al., 2020)","reference_id":"NDE4NTE5NDc3MzU4MDg3MjI0Mw=="} +{"index":4087,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2003.10555","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4088,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2110.14168","reference_contexts":"In addition to competitive performance across a range of well-established NLP (Wang et al., 2019) , multilingual (Clark et al., 2020a; Shi et al., 2022) , and reasoning (Cobbe et al., 2021 ) benchmarks, we also study the impact of UL2R on a suite of challenging BigBench tasks from Wei et al","reference_id":"NDA5NzA4OTY2NDY0Mjc4NDcyMg=="} +{"index":4089,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2110.14168","reference_contexts":"We use the GSM8K (Cobbe et al., 2021) , BBH (Suzgun et al., 2022) , StrategyQA (Geva et al., 2021) and CommonsenseQA (Talmor et al., 2019) benchmarks","reference_id":"MTQxOTYzMDQ5Nzk0ODgwMjI0MjE="} +{"index":4090,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1810.04805","reference_contexts":"Meanwhile, bidirectional models (e.g., BERT (Devlin et al., 2018) , T5 (Raffel et al., 2019) , ST-MoE (Zoph et al., 2022) ) have also been very popular as the goto model of choice, especially in smaller computational regimes (e.g., less than 30B parameters and often times in the ranges of hundred of millions of parameters)","reference_id":"Mjk2ODU2OTI2NzcxNzg5NDA2Nw=="} +{"index":4091,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1810.04805","reference_contexts":"While there have been many paradigms and self-supervision methods proposed to train these models (Devlin et al., 2018; Clark et al., 2020b; Yang et al., 2019; Raffel et al., 2019) , to this date most large language models (i.e., more than 100B parameters) are trained as decoder-only casual language models","reference_id":"NDI2MDUyODU0MDkwNzY2NjcxMg=="} +{"index":4092,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2202.07785","reference_contexts":"Scaling has inspired new research across multiple fronts, e.g., scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022; Tay et al., 2022a) , emergent abilities (Wei et al., 2022a; Ganguli et al., 2022) , reasoning capabilities (Wei et al., 2022b; Lewkowycz et al., 2022) , inter alia","reference_id":"ODgwMDAwNjgwMTQzNTYxMDAwMA=="} +{"index":4093,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2202.07785","reference_contexts":"Emergent Abilities New behaviors that arise due to scaling language models have been increasingly referred to as emergent abilities (Steinhardt, 2022; Ganguli et al., 2022; Wei et al., 2022a)","reference_id":"MTAyMzQwNTI4NjEyNTY3NjIwOA=="} +{"index":4094,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2009.03300","reference_contexts":"We compare PaLM and U-PaLM on the Massively Multi-Task Language Understanding (MMLU) benchmark (Hendrycks et al., 2020) .","reference_id":"MTY5Mzk5NDg4NjMwMjgyNTUxMTA="} +{"index":4095,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2205.10487","reference_contexts":"Secondly, we did not want to over-index on scaling studies that only measure impact on upstream cross entropy (Hernandez et al., 2022) which claims that repeating data in small quantities could be dis-proportionally harmful","reference_id":"NTg2ODkzNzE5ODYxODAxMTk5Mw=="} +{"index":4096,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2203.15556","reference_contexts":"Overall, the amount of savings is quite proportionate to the point of training and stage of convergence of the model and can probably be predicted by standard scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022) .","reference_id":"MTczOTE5NTc0MzQyOTc1OTUxNzg="} +{"index":4097,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2203.15556","reference_contexts":"To this end, large language models not only continue to improve as we scale in terms of data or computational budget (Hoffmann et al., 2022; Kaplan et al., 2020) but also acquire new abilities (Wei et al., 2022a)","reference_id":"MTQyMjc2ODA4NDI2NjUyNTQwNDE="} +{"index":4098,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2203.15556","reference_contexts":"To date, most large language models that form the basis of scaling law research are trained almost exclusively as left-to-right causal language models (Kaplan et al., 2020; Hoffmann et al., 2022) .","reference_id":"NjU3Nzg5NjMxMjQzMzk1ODk3OQ=="} +{"index":4099,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2203.15556","reference_contexts":"Specifically, following (Hoffmann et al., 2022) , we use BoolQ (Clark et al., 2019) , PIQA (Bisk et al., 2020) , HellaSWAG (Zellers et al., 2019) and Winogrande (Sakaguchi et al., 2019)","reference_id":"MTU1MTE2ODAwOTg3NDIwNzczMTM="} +{"index":4100,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2203.15556","reference_contexts":"Scaling has inspired new research across multiple fronts, e.g., scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022; Tay et al., 2022a) , emergent abilities (Wei et al., 2022a; Ganguli et al., 2022) , reasoning capabilities (Wei et al., 2022b; Lewkowycz et al., 2022) , inter alia","reference_id":"ODgwMDAwNjgwMTQzNTYxMDAwMA=="} +{"index":4101,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1705.03551","reference_contexts":"We evaluate zero-shot and few-shot closed book question answering (CBQA) tasks (Kwiatkowski et al., 2019; Joshi et al., 2017; Roberts et al., 2020) along with the zero-shot Lambada reading comprehension task (Paperno et al., 2016)","reference_id":"MTYwOTE2NjcyNjU2NDY0NTI1MjI="} +{"index":4102,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2001.08361","reference_contexts":"Scaling has inspired new research across multiple fronts, e.g., scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022; Tay et al., 2022a) , emergent abilities (Wei et al., 2022a; Ganguli et al., 2022) , reasoning capabilities (Wei et al., 2022b; Lewkowycz et al., 2022) , inter alia","reference_id":"ODgwMDAwNjgwMTQzNTYxMDAwMA=="} +{"index":4103,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2001.08361","reference_contexts":"To this end, large language models not only continue to improve as we scale in terms of data or computational budget (Hoffmann et al., 2022; Kaplan et al., 2020) but also acquire new abilities (Wei et al., 2022a)","reference_id":"MTQyMjc2ODA4NDI2NjUyNTQwNDE="} +{"index":4104,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2001.08361","reference_contexts":"Overall, the amount of savings is quite proportionate to the point of training and stage of convergence of the model and can probably be predicted by standard scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022) .","reference_id":"MTczOTE5NTc0MzQyOTc1OTUxNzg="} +{"index":4105,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2104.08691","reference_contexts":"The other direction is also possible; a model trained with a masked language objective can be adapted with the causal language modeling objective (Wang et al., 2022a; Lester et al., 2021)","reference_id":"MTEyMTQ1MzI5MDA3MTUzNDQxMTE="} +{"index":4106,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2206.14858","reference_contexts":"Scaling has inspired new research across multiple fronts, e.g., scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022; Tay et al., 2022a) , emergent abilities (Wei et al., 2022a; Ganguli et al., 2022) , reasoning capabilities (Wei et al., 2022b; Lewkowycz et al., 2022) , inter alia","reference_id":"ODgwMDAwNjgwMTQzNTYxMDAwMA=="} +{"index":4107,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2206.14858","reference_contexts":"The impact of large language models has been ubiquitous and pervasive, unlocking breakthroughs across many fields, e.g., reasoning (Wei et al., 2022b; Wang et al., 2022b; Zhou et al., 2022; Drozdov et al., 2022) , math (Lewkowycz et al., 2022) , dialog (Thoppilan et al., 2022) , multimodal applications (Yu et al., 2022) , retrieval (Tay et al., 2022c) inter alia.","reference_id":"MTA5MzI1MzI0NDI0MTgyODA4MDE="} +{"index":4108,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2203.02155","reference_contexts":"The majority of this prior work, however, requires additional data such as aggregating dozens or hundreds of NLP datasets (Raffel et al., 2019; Aghajanyan et al., 2021; Aribandi et al., 2022) , writing additional templates of instructions (Wei et al., 2021; Sanh et al., 2022) , or finetuning on human-labeled annotations (Ouyang et al., 2022)","reference_id":"NDcwMzg1Mjk4OTU3MDgzMjExNg=="} +{"index":4109,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2112.11446","reference_contexts":"There has been significant interest in scaling of language models (Rae et al., 2021; Chowdhery et al., 2022; Brown et al., 2020)","reference_id":"NDE4NTE5NDc3MzU4MDg3MjI0Mw=="} +{"index":4110,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2112.11446","reference_contexts":"Aside from PaLM 62B and PaLM 540B which we use for direct comparisons with U-PaLM, we also compare with Chinchilla 70B (Hoffmann et al., 2022) and Gopher 280B (Rae et al., 2021) .","reference_id":"MTY4ODg2MjYwMzM4MjQ5Nzc0NA=="} +{"index":4111,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1910.10683","reference_contexts":"The UL2 objective combines prefix language modeling and long-short span corruption (e.g., infilling) tasks (Raffel et al., 2019) that can be controlled at inference time using a mode switching prompt","reference_id":"Nzc4NDAwMDk3NTYxNzM5NTQ0NQ=="} +{"index":4112,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1910.10683","reference_contexts":"While there have been many paradigms and self-supervision methods proposed to train these models (Devlin et al., 2018; Clark et al., 2020b; Yang et al., 2019; Raffel et al., 2019) , to this date most large language models (i.e., more than 100B parameters) are trained as decoder-only casual language models","reference_id":"NDI2MDUyODU0MDkwNzY2NjcxMg=="} +{"index":4113,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1910.10683","reference_contexts":"The majority of this prior work, however, requires additional data such as aggregating dozens or hundreds of NLP datasets (Raffel et al., 2019; Aghajanyan et al., 2021; Aribandi et al., 2022) , writing additional templates of instructions (Wei et al., 2021; Sanh et al., 2022) , or finetuning on human-labeled annotations (Ouyang et al., 2022)","reference_id":"NDcwMzg1Mjk4OTU3MDgzMjExNg=="} +{"index":4114,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1910.10683","reference_contexts":"Meanwhile, bidirectional models (e.g., BERT (Devlin et al., 2018) , T5 (Raffel et al., 2019) , ST-MoE (Zoph et al., 2022) ) have also been very popular as the goto model of choice, especially in smaller computational regimes (e.g., less than 30B parameters and often times in the ranges of hundred of millions of parameters)","reference_id":"Mjk2ODU2OTI2NzcxNzg5NDA2Nw=="} +{"index":4115,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2002.08910","reference_contexts":"We evaluate zero-shot and few-shot closed book question answering (CBQA) tasks (Kwiatkowski et al., 2019; Joshi et al., 2017; Roberts et al., 2020) along with the zero-shot Lambada reading comprehension task (Paperno et al., 2016)","reference_id":"MTYwOTE2NjcyNjU2NDY0NTI1MjI="} +{"index":4116,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1907.10641","reference_contexts":"Specifically, following (Hoffmann et al., 2022) , we use BoolQ (Clark et al., 2019) , PIQA (Bisk et al., 2020) , HellaSWAG (Zellers et al., 2019) and Winogrande (Sakaguchi et al., 2019)","reference_id":"MTU1MTE2ODAwOTg3NDIwNzczMTM="} +{"index":4117,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2210.03057","reference_contexts":"In addition to competitive performance across a range of well-established NLP (Wang et al., 2019) , multilingual (Clark et al., 2020a; Shi et al., 2022) , and reasoning (Cobbe et al., 2021 ) benchmarks, we also study the impact of UL2R on a suite of challenging BigBench tasks from Wei et al","reference_id":"NDA5NzA4OTY2NDY0Mjc4NDcyMg=="} +{"index":4118,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2210.03057","reference_contexts":"Table 5 : Experiments on Multilingual GSM (MGSM) (Shi et al., 2022) and TydiQA (Clark et al., 2020a) Table 5 reports our results on MGSM and TydiQA","reference_id":"ODQ2NTg4NzUxMDAyMjk1NzQ3NA=="} +{"index":4119,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2210.03057","reference_contexts":"We use the MGSM (multilingual grade school math) benchmark proposed in (Shi et al., 2022)","reference_id":"ODE0NTU2MzQ5NzgzNjI1OTMwOQ=="} +{"index":4120,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2206.04615","reference_contexts":"This type of phenomena has been observed across dozens of tasks in the BIG-Bench benchmark (Srivastava et al., 2022)","reference_id":"MTEzODQyMTQ1MDgwOTM5MjAwNzg="} +{"index":4121,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2109.10686","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4122,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2207.10551","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4123,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2205.05131","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4124,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2202.06991","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4125,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2201.08239","reference_contexts":"The impact of large language models has been ubiquitous and pervasive, unlocking breakthroughs across many fields, e.g., reasoning (Wei et al., 2022b; Wang et al., 2022b; Zhou et al., 2022; Drozdov et al., 2022) , math (Lewkowycz et al., 2022) , dialog (Thoppilan et al., 2022) , multimodal applications (Yu et al., 2022) , retrieval (Tay et al., 2022c) inter alia.","reference_id":"MTA5MzI1MzI0NDI0MTgyODA4MDE="} +{"index":4126,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1905.00537","reference_contexts":"In addition to competitive performance across a range of well-established NLP (Wang et al., 2019) , multilingual (Clark et al., 2020a; Shi et al., 2022) , and reasoning (Cobbe et al., 2021 ) benchmarks, we also study the impact of UL2R on a suite of challenging BigBench tasks from Wei et al","reference_id":"NDA5NzA4OTY2NDY0Mjc4NDcyMg=="} +{"index":4127,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1905.00537","reference_contexts":"We conduct experiments on SuperGLUE (Wang et al., 2019) and TydiQA (Clark et al., 2020a) finetuning","reference_id":"NDA4MzQ1ODI0ODg0NTI3MDA3Ng=="} +{"index":4128,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2204.05832","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4129,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2203.11171","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4130,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2109.01652","reference_contexts":"A range of prior work has shown that finetuning language models on a collection of NLP tasks can improve downstream performance on a broad range of downstream tasks (Aghajanyan et al., 2021; Aribandi et al., 2022; Wei et al., 2021; Sanh et al., 2022; Ouyang et al., 2022, inter alia)","reference_id":"MzI4NzM1NDc1NTIyMDg1OTM2OQ=="} +{"index":4131,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2109.01652","reference_contexts":"The majority of this prior work, however, requires additional data such as aggregating dozens or hundreds of NLP datasets (Raffel et al., 2019; Aghajanyan et al., 2021; Aribandi et al., 2022) , writing additional templates of instructions (Wei et al., 2021; Sanh et al., 2022) , or finetuning on human-labeled annotations (Ouyang et al., 2022)","reference_id":"NDcwMzg1Mjk4OTU3MDgzMjExNg=="} +{"index":4132,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2206.10789","reference_contexts":"The impact of large language models has been ubiquitous and pervasive, unlocking breakthroughs across many fields, e.g., reasoning (Wei et al., 2022b; Wang et al., 2022b; Zhou et al., 2022; Drozdov et al., 2022) , math (Lewkowycz et al., 2022) , dialog (Thoppilan et al., 2022) , multimodal applications (Yu et al., 2022) , retrieval (Tay et al., 2022c) inter alia.","reference_id":"MTA5MzI1MzI0NDI0MTgyODA4MDE="} +{"index":4133,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"1905.07830","reference_contexts":"Specifically, following (Hoffmann et al., 2022) , we use BoolQ (Clark et al., 2019) , PIQA (Bisk et al., 2020) , HellaSWAG (Zellers et al., 2019) and Winogrande (Sakaguchi et al., 2019)","reference_id":"MTU1MTE2ODAwOTg3NDIwNzczMTM="} +{"index":4134,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2205.10625","reference_contexts":"The impact of large language models has been ubiquitous and pervasive, unlocking breakthroughs across many fields, e.g., reasoning (Wei et al., 2022b; Wang et al., 2022b; Zhou et al., 2022; Drozdov et al., 2022) , math (Lewkowycz et al., 2022) , dialog (Thoppilan et al., 2022) , multimodal applications (Yu et al., 2022) , retrieval (Tay et al., 2022c) inter alia.","reference_id":"MTA5MzI1MzI0NDI0MTgyODA4MDE="} +{"index":4135,"citer_arxiv_id":"2210.11399","cited_arxiv_id":"2202.08906","reference_contexts":"Meanwhile, bidirectional models (e.g., BERT (Devlin et al., 2018) , T5 (Raffel et al., 2019) , ST-MoE (Zoph et al., 2022) ) have also been very popular as the goto model of choice, especially in smaller computational regimes (e.g., less than 30B parameters and often times in the ranges of hundred of millions of parameters)","reference_id":"Mjk2ODU2OTI2NzcxNzg5NDA2Nw=="} +{"index":4136,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1310.1531","reference_contexts":"For the image part, we connect the seventh layer of a pre-trained Convolutional Neural Network [17, 3] (denoted as AlexNet)","reference_id":"OTA2NjI0NTY3MjQyNDA2NzE5NA=="} +{"index":4137,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1310.1531","reference_contexts":"The same features extracted from the seventh layer of AlexNet (also denoted as decaf features [3] ) are widely used by previous multimodal methods [16, 6, 15, 30]","reference_id":"MjExOTgyNDcyNzEzODA4MzA4OQ=="} +{"index":4138,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1406.5679","reference_contexts":"We follow the previous work [15] which used 1,000 images for testing","reference_id":"Mjg5ODg0NjcyOTE2ODU2MzY0NA=="} +{"index":4139,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1406.5679","reference_contexts":"Sentence Retrieval and Image Retrieval For Flickr8K and Flickr30K datasets, we adopted the same evaluation metrics as previous works [30, 6, 15] for both the tasks of sentences retrieval and image retrieval","reference_id":"NTcwMjgyMTk0NzUyOTQ2NjU1MQ=="} +{"index":4140,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1406.5679","reference_contexts":"Most recently, Karpathy et.al [15] showed that object level image features based on object detection results will generate better results than image features extracted at the global level.","reference_id":"MzY5Mzc5NTc5MDc1MzYwNjIwNQ=="} +{"index":4141,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1406.5679","reference_contexts":"A recent multimodal retrieval work [15] showed that using the RCNN object detection framework [7] combined with the decaf features significantly improves the performance","reference_id":"NzYyMjcxMzkyNDk3MzA1ODI3Nw=="} +{"index":4142,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1406.5679","reference_contexts":"The same features extracted from the seventh layer of AlexNet (also denoted as decaf features [3] ) are widely used by previous multimodal methods [16, 6, 15, 30]","reference_id":"MjExOTgyNDcyNzEzODA4MzA4OQ=="} +{"index":4143,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1406.5679","reference_contexts":"DeepFE-rcnn [15] uses a fragment mapping strategy to better exploit the object detection results","reference_id":"ODM1MzA1MTU5MzAwODM1MDUwNw=="} +{"index":4144,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1406.5679","reference_contexts":"Most of the sentence-image multimodal models [15, 6, 30, 16] use pre-computed word embedding vectors as the initialization of their model","reference_id":"OTk4MDM5NjgzMzUyMjYxNDQ5MA=="} +{"index":4145,"citer_arxiv_id":"1410.1090","cited_arxiv_id":"1301.3781","reference_contexts":"Secondly, the dense word embedding encodes the semantic meanings of the words [21]","reference_id":"MTIzOTAxMTE4MTM5MzcwMzY1MjY="} +{"index":4146,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1903.07785","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4147,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1810.04805","reference_contexts":"Since the introduction of Transformers (Vaswani et al., 2017) and BERT (Devlin et al., 2018) , we have a seen a steady improvement in the quality of these pre-trained models, mainly driven by increasing the size of the pre-training corpora (Radford et al., 2019; Yang et al., 2019; Lan et al., 2019)","reference_id":"MTcxNzc5NjQ1Mjc2ODcwMDUyNDA="} +{"index":4148,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1810.04805","reference_contexts":"We train a BERT-BASE architecture (Devlin et al., 2018) on each of these corpora, without next sentence prediction (NSP) as in (Lample and Conneau, 2019)","reference_id":"MjMzNjg1NTYxMjY5Nzg3ODQ3Mw=="} +{"index":4149,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1802.06893","reference_contexts":"Contrarily to (Grave et al., 2018) , we have chosen to deduplicate the data before language identification, because a lot of English boilerplate, such as cookie warnings, is present in pages of other languages","reference_id":"MTYwMjg1NzcxNDU0ODg3OTY2MjE="} +{"index":4150,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1802.06893","reference_contexts":"(2018) , we use the language classifier from fastText (Joulin et al., 2016b; Grave et al., 2018)","reference_id":"Njk2MDk0MzI0NjU2NDk0OTU4OA=="} +{"index":4151,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1802.06893","reference_contexts":"Our pipeline follows the data processing introduced in fastText (Mikolov et al., 2017; Grave et al., 2018) , that deduplicates documents and identifies their language","reference_id":"MTEwODc1ODY1MzMyNDE5NTU5OTc="} +{"index":4152,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1802.06893","reference_contexts":"We preprocess Common Crawl by following the pipeline of (Grave et al., 2018) , with the differences that we preserve the structure of documents and filter the data based on their distance to Wikipedia","reference_id":"MTc1MjE5MjY4NTM5ODMyMTY5NTk="} +{"index":4153,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1612.03651","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4154,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1607.01759","reference_contexts":"(2018) , we use the language classifier from fastText (Joulin et al., 2016b; Grave et al., 2018)","reference_id":"Njk2MDk0MzI0NjU2NDk0OTU4OA=="} +{"index":4155,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1804.10959","reference_contexts":"More precisely, for each language, we train a sentence piece tokenizer (Kudo, 2018 ) and a language model on data from en ru zh fr de ja es it pt pl nl ko id cs fa ro ar el hu da uk no fi bg he Language the targeted domain","reference_id":"MTEwMTg4ODM3MzE0MjUyMjU5NTg="} +{"index":4156,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1901.07291","reference_contexts":"This improves the quality of the resulting dataset and allows for the training of multilingual text level representations like XLM (Lample and Conneau, 2019) .","reference_id":"MTM0ODI4NzEwMTExNDAxOTU2ODQ="} +{"index":4157,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1901.07291","reference_contexts":"We train a BERT-BASE architecture (Devlin et al., 2018) on each of these corpora, without next sentence prediction (NSP) as in (Lample and Conneau, 2019)","reference_id":"MjMzNjg1NTYxMjY5Nzg3ODQ3Mw=="} +{"index":4158,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1909.11942","reference_contexts":"Since the introduction of Transformers (Vaswani et al., 2017) and BERT (Devlin et al., 2018) , we have a seen a steady improvement in the quality of these pre-trained models, mainly driven by increasing the size of the pre-training corpora (Radford et al., 2019; Yang et al., 2019; Lan et al., 2019)","reference_id":"MTcxNzc5NjQ1Mjc2ODcwMDUyNDA="} +{"index":4159,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1907.11692","reference_contexts":"In general, progress in sentence representations has been observed by increasing the size of the pre-training corpora (Yang et al., 2019; Liu et al., 2019; Raffel et al., 2019)","reference_id":"MTA3Mzg3Nzg2NDc3ODI0ODUzMzU="} +{"index":4160,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1802.05365","reference_contexts":"Pre-trained text representations have brought significant performance gains on many natural language processing tasks (Peters et al., 2018)","reference_id":"ODA3NTE1NTA3NjQxNzYwMDU1OQ=="} +{"index":4161,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1910.10683","reference_contexts":"In general, progress in sentence representations has been observed by increasing the size of the pre-training corpora (Yang et al., 2019; Liu et al., 2019; Raffel et al., 2019)","reference_id":"MTA3Mzg3Nzg2NDc3ODI0ODUzMzU="} +{"index":4162,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1906.08237","reference_contexts":"In general, progress in sentence representations has been observed by increasing the size of the pre-training corpora (Yang et al., 2019; Liu et al., 2019; Raffel et al., 2019)","reference_id":"MTA3Mzg3Nzg2NDc3ODI0ODUzMzU="} +{"index":4163,"citer_arxiv_id":"1911.00359","cited_arxiv_id":"1906.08237","reference_contexts":"Since the introduction of Transformers (Vaswani et al., 2017) and BERT (Devlin et al., 2018) , we have a seen a steady improvement in the quality of these pre-trained models, mainly driven by increasing the size of the pre-training corpora (Radford et al., 2019; Yang et al., 2019; Lan et al., 2019)","reference_id":"MTcxNzc5NjQ1Mjc2ODcwMDUyNDA="} +{"index":4164,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1504.00325","reference_contexts":"This task is closely related to image captioning (Chen et al., 2015) ; however, results from the two are not directly comparable, as region descriptions are short, incomplete sentences","reference_id":"MTA3MDQwMDcyMTY3MDcxNTM0NTI="} +{"index":4165,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1504.00325","reference_contexts":"We calculate BLEU, CIDEr, and METEOR scores (Chen et al., 2015) between the generated descriptions and their ground-truth descriptions","reference_id":"MTM0NzY4ODAyNTI0MTUzMjA3MjY="} +{"index":4166,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1504.00325","reference_contexts":"As a model trained on other datasets would generate complete sentences and would not be comparable (Chen et al., 2015) to our region descriptions, we convert all region descriptions generated by our model into complete sentences using predefined templates (Hou et al., 2002) .","reference_id":"ODM3MTY5Mjc5Njc1MDk0MTk0Mw=="} +{"index":4167,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1504.00325","reference_contexts":"Such models are usually evaluated using BLEU, CIDEr, or METEOR and other similar metrics that do not effectively measure how well these models understand the image (Chen et al., 2015)","reference_id":"MTQ5Mjk2NzM0MDk1OTA2MDI5MDA="} +{"index":4168,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1502.04390","reference_contexts":"We use a learning rate of 0.001 trained with rmsprop (Dauphin et al., 2015)","reference_id":"MTc2NzQyMTIwODU3Nzg3NzkzOTg="} +{"index":4169,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1505.05612","reference_contexts":"With this information, MS-COCO and VQA provide a fertile training and testing ground for models aimed at tasks for accurate object detection, segmentation, and summary-level image captioning (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014) as well as basic QA (Ren et al., 2015a , Antol et al., 2015 , Malinowski et al., 2015 , Gao et al., 2015 , Malinowski and Fritz, 2014)","reference_id":"Mzc1OTA0MDExMTMyODE2OTk3Nw=="} +{"index":4170,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1505.05612","reference_contexts":"Most new datasets (Yu et al., 2015 , Ren et al., 2015a , Antol et al., 2015 , Gao et al., 2015) have collected QA pairs on MS-COCO images, either generated automatically by NLP tools (Ren et al., 2015a) or written by human workers (Yu et al., 2015 , Antol et al., 2015 , Gao et al., 2015) .","reference_id":"MTA2NTAzOTE3NDg0NzY5NjQyNDA="} +{"index":4171,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1505.05612","reference_contexts":"Other evaluation methods require word ontologies (Malinowski and Fritz, 2014) , multiple choices (Antol et al., 2015 , Yu et al., 2015) , or human judges (Gao et al., 2015) .","reference_id":"MTA2MTQ3MDUyNDEwNTk3Njg5MTk="} +{"index":4172,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1505.05612","reference_contexts":"The proposed models range from SVM classifiers (Antol et al., 2015) and probabilistic inference (Malinowski and Fritz, 2014) to recurrent neural networks (Gao et al., 2015 , Malinowski et al., 2015 , Ren et al., 2015a) and convolutional networks (Ma et al., 2015)","reference_id":"NDMzODgwNTg2ODQyNTA5NDMz"} +{"index":4173,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1505.05612","reference_contexts":"While visual question answering has been studied as a standalone task (Yu et al., 2015 , Ren et al., 2015a , Antol et al., 2015 , Gao et al., 2015) , we introduce a dataset that combines all of our question answers with descriptions and scene graphs","reference_id":"Mzk0NzQ2MzkwODg0Njk2NzU0OA=="} +{"index":4174,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1505.05612","reference_contexts":"Rather than collecting unconstrained QA pairs as previous work has done (Antol et al., 2015 , Gao et al., 2015 , Malinowski and Fritz, 2014) , each question in Visual Genome starts with one of the six Ws -what, where, when, who, why, and how","reference_id":"MTU5OTU5OTA0NzE4OTk5MDY4NjA="} +{"index":4175,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1504.08083","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4176,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1412.2306","reference_contexts":"With this information, MS-COCO and VQA provide a fertile training and testing ground for models aimed at tasks for accurate object detection, segmentation, and summary-level image captioning (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014) as well as basic QA (Ren et al., 2015a , Antol et al., 2015 , Malinowski et al., 2015 , Gao et al., 2015 , Malinowski and Fritz, 2014)","reference_id":"Mzc1OTA0MDExMTMyODE2OTk3Nw=="} +{"index":4177,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1412.2306","reference_contexts":"Among those in the second category, a common theme has been to use recurrent neural networks to produce novel captions (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014)","reference_id":"NDczOTAwNTA0MjY3ODEzOTc5OA=="} +{"index":4178,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1412.2306","reference_contexts":"For both experiments, we use the NeuralTalk model (Karpathy and Fei-Fei, 2014) , since it not only provides state-of-the-art results but also is shown to be robust enough for predicting short descriptions","reference_id":"NDMwNTExMTE2NDgwOTU4MzY2OQ=="} +{"index":4179,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1412.2306","reference_contexts":"We have seen numerous image captioning papers (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014) that attempt to describe an entire image with a single caption","reference_id":"MTUwODM5NzUzOTUwMzE4Mjc1OTI="} +{"index":4180,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1412.2306","reference_contexts":"Generating sentence descriptions of images has gained popularity as a task in computer vision (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014) ; however, current state-of-the-art models fail to describe all the different events captured in an image and instead provide only a high-level summary of the image","reference_id":"NTEzODUyMTUxNjY1OTA2Mjk4Mw=="} +{"index":4181,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1412.2306","reference_contexts":"For example, a stateof-the-art model (Karpathy and Fei-Fei, 2014) provides a description of one MS-COCO image in Figure 1 as \"two men are standing next to an elephant.\" But what is missing is the further understanding of where each object is, what each person is doing, what the relationship between the person and elephant is, etc","reference_id":"OTk4MTc2OTQ5NTA2ODg3NjIyMQ=="} +{"index":4182,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1412.2306","reference_contexts":"We train one of the top 16 state-of-the-art image caption generator (Karpathy and Fei-Fei, 2014) on (1) our dataset to generate region descriptions and on (2) Flickr30K (Young et al., 2014) to generate sentence descriptions","reference_id":"NDMwMzgyMjg2NDY5MTg1NTM4MA=="} +{"index":4183,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1502.03671","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4184,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1410.1090","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4185,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1301.3781","reference_contexts":"Specifically, we use word2vec (Mikolov et al., 2013) to convert each word in a description to a 300-dimensional vector","reference_id":"MjQwNDYwNzIxMTU5OTI5MzAw"} +{"index":4186,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1312.6229","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4187,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1409.1556","reference_contexts":"For testing, we crop the ground-truth region bounding boxes of images and extract their 4096-dimensional 16-layer VGG network (Simonyan and Zisserman, 2014) features","reference_id":"NjQ1ODg3NjU5Mjc3NTUwODU3NA=="} +{"index":4188,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1409.1556","reference_contexts":"We fine-tune a 16-layer VGG network (Simonyan and Zisserman, 2014) with the same learning rates mentioned in Section 6.1.","reference_id":"MTU5NzQxNzgyOTg0MDk3ODkxMjc="} +{"index":4189,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1409.1556","reference_contexts":"\"white stripe\" (predicted \"black and white zebra\") cally, we fine-tune a 16-layer VGG network (Simonyan and Zisserman, 2014) for both of these experiments using the 50, 000 attribute and 43, 000 object-attribute pair instances respectively","reference_id":"MTA1Nzk4MzE0OTMzMTA4MzIxMDU="} +{"index":4190,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1409.4842","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4191,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1411.4555","reference_contexts":"We have seen numerous image captioning papers (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014) that attempt to describe an entire image with a single caption","reference_id":"MTUwODM5NzUzOTUwMzE4Mjc1OTI="} +{"index":4192,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1411.4555","reference_contexts":"Among those in the second category, a common theme has been to use recurrent neural networks to produce novel captions (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014)","reference_id":"NDczOTAwNTA0MjY3ODEzOTc5OA=="} +{"index":4193,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1411.4555","reference_contexts":"Generating sentence descriptions of images has gained popularity as a task in computer vision (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014) ; however, current state-of-the-art models fail to describe all the different events captured in an image and instead provide only a high-level summary of the image","reference_id":"NTEzODUyMTUxNjY1OTA2Mjk4Mw=="} +{"index":4194,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1411.4555","reference_contexts":"With this information, MS-COCO and VQA provide a fertile training and testing ground for models aimed at tasks for accurate object detection, segmentation, and summary-level image captioning (Kiros et al., 2014 , Mao et al., 2014 , Karpathy and Fei-Fei, 2014 , Vinyals et al., 2014) as well as basic QA (Ren et al., 2015a , Antol et al., 2015 , Malinowski et al., 2015 , Gao et al., 2015 , Malinowski and Fritz, 2014)","reference_id":"Mzc1OTA0MDExMTMyODE2OTk3Nw=="} +{"index":4195,"citer_arxiv_id":"1602.07332","cited_arxiv_id":"1502.03044","reference_contexts":"More recently, researchers have also used a visual attention model (Xu et al., 2015) .","reference_id":"NzM3NjIzMzM1NzAzMDUyMTQwMg=="} +{"index":4196,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1409.0473","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4197,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1508.04395","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4198,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1601.06733","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4199,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1412.3555","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4200,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1412.2007","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4201,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1310.4546","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4202,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1602.06023","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4203,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1509.00685","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4204,"citer_arxiv_id":"1602.06023","cited_arxiv_id":"1212.5701","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4205,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2001.09977","reference_contexts":"Open-domain dialogue, which involves chat about any topic, rather than a specific goal-directed topic, is commonly studied by training large language models (Adiwardana et al., 2020; Zhang et al., 2019; Roller et al., 2020)","reference_id":"MjQ4Mjk3NDIxOTY4NTE2NzQw"} +{"index":4206,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2001.09977","reference_contexts":"This is the standard approach in state-of-the-art chatbots such as Meena (Adiwardana et al., 2020) or BlenderBot (Roller et al., 2020)","reference_id":"NTE3Njc3MDUzMzU2MTI4MzYyMQ=="} +{"index":4207,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2001.08435","reference_contexts":"Datasets such as pushshift.io Reddit (Baumgartner et al., 2020) , PersonaChat (Zhang et al., 2018) or Empathetic Dialogues (Rashkin et al., 2019) (see Huang et al","reference_id":"MTQ3MTk5ODg2MDk3MjcyMTE2MjY="} +{"index":4208,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1612.04426","reference_contexts":"Retrieval-augmentation is also studied in the area of language modeling, where it is used for pre-training (Guu et al., 2020) , and as a memory (Yogatama et al., 2021) , especially using k-nearest neighbor-based cache models (Khandelwal et al., 2021 (Khandelwal et al., , 2020;; Grave et al., 2016; Merity et al., 2016) .","reference_id":"ODY0MTI1MzIwMjk0MTA2ODAyMQ=="} +{"index":4209,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2002.08909","reference_contexts":"Retrieval-augmentation is also studied in the area of language modeling, where it is used for pre-training (Guu et al., 2020) , and as a memory (Yogatama et al., 2021) , especially using k-nearest neighbor-based cache models (Khandelwal et al., 2021 (Khandelwal et al., , 2020;; Grave et al., 2016; Merity et al., 2016) .","reference_id":"ODY0MTI1MzIwMjk0MTA2ODAyMQ=="} +{"index":4210,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2007.01282","reference_contexts":"A related, but perhaps simpler, method is that of FiD (Izacard and Grave, 2020)","reference_id":"NDcyMzYyMTQ1ODk5MTYwNzcyMg=="} +{"index":4211,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2007.01282","reference_contexts":"More recently, the RAG (Retrieval-Augmented Generation) (Lewis et al., 2020) and FiD (Fusion-in-Decoder) (Izacard and Grave, 2020) models developed these ideas further, using a neural retriever as well, with superior results","reference_id":"MTAxMzMzNTgyNjg4MTk1MDE1OTU="} +{"index":4212,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2007.01282","reference_contexts":"The queries are then used to retrieve relevant knowledge that is prepended to the conversational history, which is encoded using the Fusion-in-Decoder method (Izacard and Grave, 2020)","reference_id":"ODQ3Mzc5NjU5MTY5MDU3ODY3Ng=="} +{"index":4213,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4214,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2102.01951","reference_contexts":"Such static language modeling fails to take into account the dynamic state of the world, where new information is coming in by the day -or even by the minuteas the knowledge in static models is gleaned from the point in time when the dataset was collected, and then frozen into the model that is trained; see (Lazaridou et al., 2021) for criticisms of this approach","reference_id":"MTgwNDE3MDE3Mjg2MTE5NTk5Mzc="} +{"index":4215,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1910.13461","reference_contexts":"We consider fine-tuning various existing pre-trained models: T5 (Raffel et al., 2019) , BART-Large (Lewis et al., 2019) and BlenderBot variants (Roller et al., 2020)","reference_id":"MjE0MTM0ODkzMTIzNTg0MDc1Mw=="} +{"index":4216,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1609.07843","reference_contexts":"Retrieval-augmentation is also studied in the area of language modeling, where it is used for pre-training (Guu et al., 2020) , and as a memory (Yogatama et al., 2021) , especially using k-nearest neighbor-based cache models (Khandelwal et al., 2021 (Khandelwal et al., , 2020;; Grave et al., 2016; Merity et al., 2016) .","reference_id":"ODY0MTI1MzIwMjk0MTA2ODAyMQ=="} +{"index":4217,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1910.10683","reference_contexts":"We consider fine-tuning various existing pre-trained models: T5 (Raffel et al., 2019) , BART-Large (Lewis et al., 2019) and BlenderBot variants (Roller et al., 2020)","reference_id":"MjE0MTM0ODkzMTIzNTg0MDc1Mw=="} +{"index":4218,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2004.13637","reference_contexts":"We consider fine-tuning various existing pre-trained models: T5 (Raffel et al., 2019) , BART-Large (Lewis et al., 2019) and BlenderBot variants (Roller et al., 2020)","reference_id":"MjE0MTM0ODkzMTIzNTg0MDc1Mw=="} +{"index":4219,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2004.13637","reference_contexts":"Open-domain dialogue, which involves chat about any topic, rather than a specific goal-directed topic, is commonly studied by training large language models (Adiwardana et al., 2020; Zhang et al., 2019; Roller et al., 2020)","reference_id":"MjQ4Mjk3NDIxOTY4NTE2NzQw"} +{"index":4220,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2004.13637","reference_contexts":"We note that this technique is similar to the one used in retrieve and refine architectures (Roller et al., 2020) .","reference_id":"MjI0MjMwMDEyNzc0NjkxOTM2MQ=="} +{"index":4221,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2004.13637","reference_contexts":"This is the standard approach in state-of-the-art chatbots such as Meena (Adiwardana et al., 2020) or BlenderBot (Roller et al., 2020)","reference_id":"NTE3Njc3MDUzMzU2MTI4MzYyMQ=="} +{"index":4222,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2004.13637","reference_contexts":"However, our fine-tuning task is built with crowdworkers with specific instructions to not use toxic language, a procedure which is shown to yield safer language models (Roller et al., 2020) .","reference_id":"MjAxNDgzODQzMDM1NzkzMzkzNw=="} +{"index":4223,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2104.07567","reference_contexts":"Large language models, even though they store an impressive amount of knowledge within their weights, are known to hallucinate facts when generating dialogue (Shuster et al., 2021) ; moreover, those facts are frozen in time at the point of model training","reference_id":"OTQ0MDA1MTQxNTM5OTI0MjIwNw=="} +{"index":4224,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2104.07567","reference_contexts":"It has been observed before that large language models, when augmented with retrieval, have trouble with choosing between copying knowledge remembered within their weights and knowledge provided in retrieved documents (Shuster et al., 2021) .","reference_id":"MTI3ODk1NDk1NDYxNTQwNTEyMzY="} +{"index":4225,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2104.07567","reference_contexts":"Further, static language models are known to hallucinate, that is they generate plausible look-ing statements that are factually incorrect, which can be interpreted as a form of lossy compression when employing training to encode that knowledge within the weights of a neural network; see (Shuster et al., 2021) for an in-depth study.","reference_id":"ODMyMjc5MzExMzY5NTYzMjE0OA=="} +{"index":4226,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2104.07567","reference_contexts":"FiD-RAG, proposed in (Shuster et al., 2021) combines the two methods","reference_id":"MTQ1ODcyODc3MDM1NTQyMzI2MDU="} +{"index":4227,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2104.07567","reference_contexts":"We see that the domains are rather diverse, coming from all kinds of topics, and in particular that the Wikipedia domain is actually fairly small (8.56% of queries), which is interesting because most other studies have used Wikipedia only as their knowledge resource (Chen et al., 2017; Lewis et al., 2020; Dinan et al., 2019; Shuster et al., 2021) .","reference_id":"NzY0OTI5NzY5MTg3NDg4ODc3Nw=="} +{"index":4228,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1911.00359","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4229,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1410.3916","reference_contexts":"Earlier works such as Memory Networks (Weston et al., 2014) and DrQA (Chen et al., 2017) utilized TFIDF-based retrieval over documents to provide additional input to neural models for the task of question answering, following the well studied area of non-neural methods that use retrieval for QA (Voorhees, 2001)","reference_id":"MTM5NDc0MjQxNjE4NzgzMDM2Nzc="} +{"index":4230,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"2010.07079","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4231,"citer_arxiv_id":"2107.07566","cited_arxiv_id":"1911.00536","reference_contexts":"Open-domain dialogue, which involves chat about any topic, rather than a specific goal-directed topic, is commonly studied by training large language models (Adiwardana et al., 2020; Zhang et al., 2019; Roller et al., 2020)","reference_id":"MjQ4Mjk3NDIxOTY4NTE2NzQw"} +{"index":4232,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2308.12966","reference_contexts":"(2) we build from scratch a data processing pipeline to add data, filter data, maintain data, update data and apply data to optimize our text-to-image model","reference_id":"MTAyODAwNzA3MjgzMzYzMTUwNg=="} +{"index":4233,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2308.12966","reference_contexts":"Re-captioning with Structural Captions Existing MLLMs, e.g., BLIP-2 [17] and Qwen-VL [2] tend to generate over-simplified captions that resemble MS-COCO captions [19] or highly redundant captions that are not related to the images","reference_id":"Mjk2NjMyMzQ1Nzc0NTYzNTU1NA=="} +{"index":4234,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2211.01324","reference_contexts":"Experiment Setting Following the setting in prior research [29, 3] , we evaluate different variants of the models using the zero-shot Frechet Inception Distance (FID) [14] on MS COCO [19] 256 × 256 validation dataset by generating 30,000 images from the prompts in the validation set","reference_id":"MzkyMzc5OTcxMjcxODEyOTUwMA=="} +{"index":4235,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2211.01324","reference_contexts":"(3) we refine the raw captions in the image-text data pairs with Multimodal Large Language Model (MLLM)","reference_id":"MTg0NDI0Nzc3MTUyODMxMTQ0MTQ="} +{"index":4236,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2211.01324","reference_contexts":"In contrast, eDiff-I [3] and Swinv2-Imagen [18] fuse the two encoders, CLIP and T5, to further improve their text understanding capabilities","reference_id":"MTE0NTI3NDc4NTk4Mzg1MDIyNzE="} +{"index":4237,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2403.03206","reference_contexts":"Diffusion-based text-to-image generative models, such as DALL-E [5] , Stable Diffusion [24, 10] and Pixart [7] , have shown the ability to generate images with unprecedented quality","reference_id":"MTU0MTY5MjY4NzU4MDc3OTM3NjQ="} +{"index":4238,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2310.01218","reference_contexts":"Various works have made efforts to equip text-to-image models with the multi-turn ability using MLLMs, such as Next-GPT [35] , SEED-LLaMA [11] , RPG [38] , and DALLE-3 [5]","reference_id":"OTQzNTYyMjk3MDExMzA0ODk2Mg=="} +{"index":4239,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2402.17245","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4240,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2310.03744","reference_contexts":"Our model structure is similar to LLAVA-1.6 [20]","reference_id":"NDEyNzYxOTkzMDE0NTA2NjEz"} +{"index":4241,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2310.04378","reference_contexts":"Algorithmic Acceleration Recently, various methods have been proposed to reduce the inference steps of diffusionbased text-to-image models [22, 31, 21, 37, 40]","reference_id":"MjI3NTY1ODM5MzI5OTE5MjUwMA=="} +{"index":4242,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2311.17042","reference_contexts":"Algorithmic Acceleration Recently, various methods have been proposed to reduce the inference steps of diffusionbased text-to-image models [22, 31, 21, 37, 40]","reference_id":"MjI3NTY1ODM5MzI5OTE5MjUwMA=="} +{"index":4243,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2309.05519","reference_contexts":"Various works have made efforts to equip text-to-image models with the multi-turn ability using MLLMs, such as Next-GPT [35] , SEED-LLaMA [11] , RPG [38] , and DALLE-3 [5]","reference_id":"OTQzNTYyMjk3MDExMzA0ODk2Mg=="} +{"index":4244,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2311.09257","reference_contexts":"Algorithmic Acceleration Recently, various methods have been proposed to reduce the inference steps of diffusionbased text-to-image models [22, 31, 21, 37, 40]","reference_id":"MjI3NTY1ODM5MzI5OTE5MjUwMA=="} +{"index":4245,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2401.11708","reference_contexts":"Various works have made efforts to equip text-to-image models with the multi-turn ability using MLLMs, such as Next-GPT [35] , SEED-LLaMA [11] , RPG [38] , and DALLE-3 [5]","reference_id":"OTQzNTYyMjk3MDExMzA0ODk2Mg=="} +{"index":4246,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2311.18828","reference_contexts":"Algorithmic Acceleration Recently, various methods have been proposed to reduce the inference steps of diffusionbased text-to-image models [22, 31, 21, 37, 40]","reference_id":"MjI3NTY1ODM5MzI5OTE5MjUwMA=="} +{"index":4247,"citer_arxiv_id":"2405.08748","cited_arxiv_id":"2310.20550","reference_contexts":"Re-captioning with Raw Captions: Capsfusion [41] proposed to fuse raw captions with generated descriptive captions using ChatGPT","reference_id":"MTExNTgzNDE5OTYwNDQ0NDg5MjM="} +{"index":4248,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2204.14198","reference_contexts":"Flamingo [3] uses a large private interleaved image-text dataset, using NFNet-F6 as a visual encoder while scaling up the text encoder from 1.4B to 70B parameters","reference_id":"NDc3Mjc2MDE3OTg4NjI4MTUyMg=="} +{"index":4249,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2106.08254","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4250,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2006.07159","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4251,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2209.06794","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4252,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2208.07339","reference_contexts":"1 We hypothesize that bfloat16 fixed the issue due to larger models typically showing larger activation values as observed by [16] , making bfloat16 more suitable with its wider dynamic range (8 exponent bits).","reference_id":"MTUwOTA4ODE3NTkzNDg2MTY1NjA="} +{"index":4253,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2211.07636","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4254,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2006.16241","reference_contexts":"Right: average performance on five ImageNet robustness datasets (ImageNet-V2 [60] , ImageNet-R [22] , ImageNet-Sketch [74], ObjectNet [5] , and ImageNet-A [24] )","reference_id":"MTcyOTkyOTI1NzA4Mjk1NjM0NzI="} +{"index":4255,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2006.16241","reference_contexts":"We evaluate the models on ImageNet [15] , ImageNet distribution shift datasets [22, 23, 24, 74, 5] , and the visual task adaptation benchmark (VTAB) [83]","reference_id":"MTMxMTc1NjE2MjQzMDA4MTEwMjc="} +{"index":4256,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2006.16241","reference_contexts":"Fitting power-law (E = βC α ) on the Pareto frontier of the available models, we measure scaling coefficients α openCLIP = -0.11 and α CLIP = -0.16 for zero-shot top-1 ImageNet and α openCLIP = -0.13 and α CLIP = -0.24 for ImageNet robustness datasets performance [22, 23, 24, 74]","reference_id":"MTExODcyNjUzMDUxMDI4MDkwMjg="} +{"index":4257,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2006.16241","reference_contexts":"We compare the results with OpenAI CLIP models fine-tuned with the same settings, evaluating the models using top-1 accuracy on ImageNet and the ImageNet distribution shift datasets [22, 23, 24, 74, 5]","reference_id":"MTQ4MzkwMzAzNjYzMTk1MTk4Mzc="} +{"index":4258,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1903.12261","reference_contexts":"We evaluate the models on ImageNet [15] , ImageNet distribution shift datasets [22, 23, 24, 74, 5] , and the visual task adaptation benchmark (VTAB) [83]","reference_id":"MTMxMTc1NjE2MjQzMDA4MTEwMjc="} +{"index":4259,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1903.12261","reference_contexts":"Fitting power-law (E = βC α ) on the Pareto frontier of the available models, we measure scaling coefficients α openCLIP = -0.11 and α CLIP = -0.16 for zero-shot top-1 ImageNet and α openCLIP = -0.13 and α CLIP = -0.24 for ImageNet robustness datasets performance [22, 23, 24, 74]","reference_id":"MTExODcyNjUzMDUxMDI4MDkwMjg="} +{"index":4260,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1903.12261","reference_contexts":"We compare the results with OpenAI CLIP models fine-tuned with the same settings, evaluating the models using top-1 accuracy on ImageNet and the ImageNet distribution shift datasets [22, 23, 24, 74, 5]","reference_id":"MTQ4MzkwMzAzNjYzMTk1MTk4Mzc="} +{"index":4261,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1907.07174","reference_contexts":"Right: average performance on five ImageNet robustness datasets (ImageNet-V2 [60] , ImageNet-R [22] , ImageNet-Sketch [74], ObjectNet [5] , and ImageNet-A [24] )","reference_id":"MTcyOTkyOTI1NzA4Mjk1NjM0NzI="} +{"index":4262,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1907.07174","reference_contexts":"We evaluate the models on ImageNet [15] , ImageNet distribution shift datasets [22, 23, 24, 74, 5] , and the visual task adaptation benchmark (VTAB) [83]","reference_id":"MTMxMTc1NjE2MjQzMDA4MTEwMjc="} +{"index":4263,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1907.07174","reference_contexts":"Fitting power-law (E = βC α ) on the Pareto frontier of the available models, we measure scaling coefficients α openCLIP = -0.11 and α CLIP = -0.16 for zero-shot top-1 ImageNet and α openCLIP = -0.13 and α CLIP = -0.24 for ImageNet robustness datasets performance [22, 23, 24, 74]","reference_id":"MTExODcyNjUzMDUxMDI4MDkwMjg="} +{"index":4264,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1907.07174","reference_contexts":"We compare the results with OpenAI CLIP models fine-tuned with the same settings, evaluating the models using top-1 accuracy on ImageNet and the ImageNet distribution shift datasets [22, 23, 24, 74, 5]","reference_id":"MTQ4MzkwMzAzNjYzMTk1MTk4Mzc="} +{"index":4265,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2010.14701","reference_contexts":"So far, the literature on empirical scaling laws has focused on language-only [35, 72, 28] or visiononly models [82, 25, 61]","reference_id":"MTMzMzg5MDIzMTU2NjE2NzEzMjY="} +{"index":4266,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1712.00409","reference_contexts":"Strong empirical evidence that increasing model or data scale is beneficial was initially studied in the context of deep learning and computer vision [69, 26]","reference_id":"MTI0MzM4MDU4NDQ0ODE4MDQyMzE="} +{"index":4267,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1712.00409","reference_contexts":"For instance, in [26] , the power law relation between scale and model performance was highlighted","reference_id":"NzQyNDYzNTg5ODA3NTI0NDU5Ng=="} +{"index":4268,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2210.02303","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4269,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2203.15556","reference_contexts":"In particular, large models have led to substantial advances in speech [56] , language [17, 57, 8, 28] , vision [38, 82] , and multi-modal language-vision settings [55, 33, 54, 59, 62]","reference_id":"NzgwMzYyMjY5NzkxMzk3NTEyOQ=="} +{"index":4270,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2203.15556","reference_contexts":"Moreover, neural scaling laws of the power law form were derived for language models, connecting model, data, and training compute scale to performance [35, 72, 28]","reference_id":"NTg1MDk3NzI1MTY2MzIwNjc0Mg=="} +{"index":4271,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2203.15556","reference_contexts":"Obtaining more data points for smaller and intermediate scales can provide enough sampling density to better understand the optimal configuration of model size, dataset size and number of samples seen given a fixed compute, similar to works such as [28, 39]","reference_id":"MTU4MDg0ODYyNjIzNTIzNDcwMTA="} +{"index":4272,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2203.15556","reference_contexts":"This also allowed accurate prediction of model performance at larger scales, and researchers were able to determine the scale parameters for achieving optimal performance given a fixed amount of compute [28, 39]","reference_id":"NDc4Mjg2NTQ2NzkzODI1NDgx"} +{"index":4273,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2203.15556","reference_contexts":"So far, the literature on empirical scaling laws has focused on language-only [35, 72, 28] or visiononly models [82, 25, 61]","reference_id":"MTMzMzg5MDIzMTU2NjE2NzEzMjY="} +{"index":4274,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2203.15556","reference_contexts":"This allows us to assess performance of models pre-trained with different training durations and avoid suboptimal training when using the same schedule for runs of different length [28]","reference_id":"MTI4NjU3Nzc4OTgxMzIyNDg5Mzc="} +{"index":4275,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2203.15556","reference_contexts":"scales, and can be used to determine pre-training regimes that obtain optimal model performance for a fixed amount of compute [35, 28] .","reference_id":"MzU3OTA3ODY0MzAwNTY5NjA5NA=="} +{"index":4276,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"Moreover, neural scaling laws of the power law form were derived for language models, connecting model, data, and training compute scale to performance [35, 72, 28]","reference_id":"NTg1MDk3NzI1MTY2MzIwNjc0Mg=="} +{"index":4277,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"We also observe bottleneck behaviors [35, 82] that occur when fixing one scaling dimension while increasing others","reference_id":"NjQ2NTk3NTMzMzYzMjQwMjg2Nw=="} +{"index":4278,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"Specifically, researchers have postulated scaling laws in the form of power law relationships between model performance and model compute, or data scale [35, 72, 61, 82]","reference_id":"NTIxODYxMjA2NzcxOTc3NjcyMA=="} +{"index":4279,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"In line with previous studies [35, 72, 61, 82] , our work observes scaling laws of power law form across various downstream tasks","reference_id":"NzAxOTc0NDQ5ODI5MDIzMTMzNQ=="} +{"index":4280,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"Scaling law studies were also conducted for transfer and out-of-distribution performance [35, 72, 82]","reference_id":"ODM2MDQxMzc3NDcyODE0ODU4NQ=="} +{"index":4281,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"So far, the literature on empirical scaling laws has focused on language-only [35, 72, 28] or visiononly models [82, 25, 61]","reference_id":"MTMzMzg5MDIzMTU2NjE2NzEzMjY="} +{"index":4282,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"We find-in accord with previous works on uni-modal learning [35, 82] -a power law relation between scale (model, data and the number of samples seen) and downstream performance in a broad range of settings, including zero-shot classification, retrieval, few-and full-shot linear probing and fine-tuning","reference_id":"OTg4NTIyMTU4ODU1ODU4NTg2NQ=="} +{"index":4283,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2001.08361","reference_contexts":"scales, and can be used to determine pre-training regimes that obtain optimal model performance for a fixed amount of compute [35, 28] .","reference_id":"MzU3OTA3ODY0MzAwNTY5NjA5NA=="} +{"index":4284,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4285,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1805.08974","reference_contexts":"While this technique underperforms end-to-end fine-tuning, it is often preferred because it requires far less compute [40, 55]","reference_id":"MTY5MjQ5MTkxMjEzMjk5OTc0MjU="} +{"index":4286,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2006.15704","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4287,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1807.03748","reference_contexts":"The InfoNCE loss [52] used by CLIP can be thought of as a method to maximize the mutual information between text and image representations","reference_id":"NDMwNzczMTM5NDIxNjg4OTU1MA=="} +{"index":4288,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2111.10050","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4289,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2204.06125","reference_contexts":"In particular, large models have led to substantial advances in speech [56] , language [17, 57, 8, 28] , vision [38, 82] , and multi-modal language-vision settings [55, 33, 54, 59, 62]","reference_id":"NzgwMzYyMjY5NzkxMzk3NTEyOQ=="} +{"index":4290,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1902.10811","reference_contexts":"Right: average performance on five ImageNet robustness datasets (ImageNet-V2 [60] , ImageNet-R [22] , ImageNet-Sketch [74], ObjectNet [5] , and ImageNet-A [24] )","reference_id":"MTcyOTkyOTI1NzA4Mjk1NjM0NzI="} +{"index":4291,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2111.02114","reference_contexts":"Both LAION-400M [65] and LAION-5B [64] are open, public image-text datasets that were composed by obtaining links from Common Crawl [1]","reference_id":"Njg2NTQ0MDIyMDI5Nzg0MTQ1Mg=="} +{"index":4292,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2111.02114","reference_contexts":"Having open datasets [65, 64] and open source tools [32] at hand, such experiments can be conducted and reproduced in a common effort by the broader research community.","reference_id":"MTI5NDI4ODQ3ODg5NjI4OTQxMzg="} +{"index":4293,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2111.02114","reference_contexts":"The resulting datasets, LAION-400M [65] and LAION-5B [64] are publicly available, enabling training language-vision models at larger scale [62, 49, 27]","reference_id":"MTAyMzg2NjYxMjkyMTE1NTc1NDU="} +{"index":4294,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2111.02114","reference_contexts":"We use the LAION-400M [65] and LAION-5B [64] datasets which are open, public image-text datasets validated by the pre-training of state-of-the art multi-modal models such as CLIP [55] and Stable Diffusion [62]","reference_id":"OTI4ODg1OTc1ODU0MjExMzE0MA=="} +{"index":4295,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2207.10551","reference_contexts":"Interestingly, upstream performance does not always correlate with downstream performance [72, 71]","reference_id":"MTczMTQ3MjgzMTU2NDY2NzQ2MzE="} +{"index":4296,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1905.13549","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4297,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"2211.05778","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4298,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1910.04867","reference_contexts":"Additionally, we train linear classifiers on the visual task adaptation benchmark (VTAB) [83] (Figure 3 )","reference_id":"OTA5MjAxODgyMDg1MDU3MDkyMA=="} +{"index":4299,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1910.04867","reference_contexts":"This table shows zero-shot performance at 224 pixel resolution, displaying accuracy on ImageNet [15] , average accuracy on 35 VTAB+ datasets [64, 83] , and image retrieval recall at 5 on MS-COCO image retrieval [46] .","reference_id":"MTM3NDg5Mzg3NjY3MTU3MzQ4MjI="} +{"index":4300,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1910.04867","reference_contexts":"Table 5 : Scaling model and data size leads to lower error linear classifers on ImageNet [15] , CIFAR100 [43] , and the visual task adaptation benchmark (VTAB) [83]","reference_id":"MTcyNzA1ODMxMjU3MjYwNzYxMTY="} +{"index":4301,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1910.04867","reference_contexts":"[83]","reference_id":"MTU5NzI2OTQxODA2MTc1NzgyMDg="} +{"index":4302,"citer_arxiv_id":"2212.07143","cited_arxiv_id":"1910.04867","reference_contexts":"We evaluate the models on ImageNet [15] , ImageNet distribution shift datasets [22, 23, 24, 74, 5] , and the visual task adaptation benchmark (VTAB) [83]","reference_id":"MTMxMTc1NjE2MjQzMDA4MTEwMjc="} +{"index":4303,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2204.14198","reference_contexts":"Following previous work [1, 54] , we select supporting demonstration exemplars for in-context learning based on retrieval results","reference_id":"MTAzMTQzOTE5MjQ3OTM5NzM3Nzk="} +{"index":4304,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2304.03373","reference_contexts":"For example, ReCo [56], GLIGEN [27] , and Layout-Guidance [6] can generate images based on bounding box inputs and regional captions","reference_id":"MTIzNzQ4NzcxNjExNjUwOTM1NDY="} +{"index":4305,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2304.03373","reference_contexts":"We feed the same generated layouts from LayoutGPT to Layout-Guidance [6] and compute image-level metrics","reference_id":"MTM3ODc5NDE1NDc2ODM1MTUxNTU="} +{"index":4306,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2302.12192","reference_contexts":"For instance, text-to-image generation (T2I) models suffer from generating objects with specified counts, positions, and attributes [10, 24]","reference_id":"NzIwNzQ4MjI1NzYyODcwMTc2MQ=="} +{"index":4307,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2302.12192","reference_contexts":"Compositional Image Generation Recent studies have shown that text-to-image generation (T2I) models suffer from compositional issues such as missing objects, incorrect spatial relations, and incorrect attributes [24, 2]","reference_id":"MTE5MDQzMzEzNzgxMTU5OTA5NDI="} +{"index":4308,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2304.09842","reference_contexts":"With the strong generalization ability of contemporary LLMs, recent work attempts to adapt the power of LLMs on multimodal tasks [31, 55]","reference_id":"NjkzMTMzMjE4NDc4NzQxNTE2MQ=="} +{"index":4309,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2302.13153","reference_contexts":"Visual layout is an essential symbolic representation that has been widely studied as it reflects the compositions of a visual space [33, 53, 45, 34]","reference_id":"NDI1MzQ2MDEwNjA0NTY0Mzc4MQ=="} +{"index":4310,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2303.08774","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4311,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2303.04671","reference_contexts":"Visual ChatGPT [51] proposes a prompt manager that supports the execution of various image generation models","reference_id":"ODI4NTUyNTgyODg1ODIyMjM3MA=="} +{"index":4312,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2303.04671","reference_contexts":"[51] .","reference_id":"OTU2MzM5ODkxOTgzMzI5NDc0OQ=="} +{"index":4313,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2303.11381","reference_contexts":"With the strong generalization ability of contemporary LLMs, recent work attempts to adapt the power of LLMs on multimodal tasks [31, 55]","reference_id":"NjkzMTMzMjE4NDc4NzQxNTE2MQ=="} +{"index":4314,"citer_arxiv_id":"2305.15393","cited_arxiv_id":"2302.00923","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4315,"citer_arxiv_id":"2102.01293","cited_arxiv_id":"1810.04805","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4316,"citer_arxiv_id":"2102.01293","cited_arxiv_id":"1912.06680","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4317,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1311.0701","reference_contexts":"On every music dataset the performance by this model is significantly better than those achieved by all the other models as well as the best results reported with recurrent neural networks in (Bayer et al., 2013)","reference_id":"MTE3MzkxNjg4NzE5OTgzMzc4MTU="} +{"index":4318,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1311.0701","reference_contexts":"worse than the result obtained by RNNs trained with the technique of fast dropout (FD) which are 3.09 and 8.01, respectively (Bayer et al., 2013) .","reference_id":"MTI1MzE3NzUyMzA3MzA0MTQ5ODc="} +{"index":4319,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1308.0850","reference_contexts":"The results by both the DOT(S)-RNN and the sRNN for word-level modeling surpassed the previous best performance achieved by an RNN with 1000 long short-term memory (LSTM) units (Graves, 2013) as well as that by a shallow RNN with a larger hidden state (Mikolov et al., 2011) , even when both of them used dynamic evaluation 3","reference_id":"MzcyMzEzNjEzNzI0NTMyMTc4"} +{"index":4320,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1308.0850","reference_contexts":"RNNs have been successfully used for various task such as language modeling (see, e.g., Graves, 2013; Pascanu et al., 2013a; Mikolov, 2012; Sutskever et al., 2011) , learning word embeddings (see, e.g., Mikolov et al., 2013a) , online handwritten recognition (Graves et al., 2009) and speech recognition (Graves et al., 2013) .","reference_id":"Nzg2MTcwODg1ODk3NjgwNzA0MA=="} +{"index":4321,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1308.0850","reference_contexts":"For character-level modeling the state-of-the-art results were obtained using an optimization method Hessian-free with a specific type of RNN architecture called mRNN (Mikolov et al., 2012a) or a regularization technique called adaptive weight noise (Graves, 2013)","reference_id":"MTI0NjAxODIyNzAzMzY4MzE3MTQ="} +{"index":4322,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1308.0850","reference_contexts":"An RNN may be extended deeper in yet another way by stacking multiple recurrent hidden layers on top of each other (Schmidhuber, 1992; El Hihi and Bengio, 1996; Jaeger, 2007; Graves, 2013) .","reference_id":"MzE2NzYwMzE0NDUzMzQ2MzYxNw=="} +{"index":4323,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1308.0850","reference_contexts":"Each hidden state at each level may also be made to depend on the input as well (Graves, 2013)","reference_id":"MTIzOTE4NjU3ODU4NDk1MDkwOTk="} +{"index":4324,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1308.0850","reference_contexts":"Reported by Graves (2013) using the dynamic evaluation and weight noise.3 Dynamic evaluation refers to an approach where the parameters of a model are updated as the validation/test data is predicted.","reference_id":"MTc3MjU1MDEyMjc4NjA1MTg0NDQ="} +{"index":4325,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1308.0850","reference_contexts":"We empirically evaluated the proposed designs against the conventional RNN which has only a single hidden layer and against another approach of building a deep RNN (stacked RNN, Graves, 2013) , on the task of polyphonic music prediction and language modeling.","reference_id":"NDU1NzY0MTY0Mzk0NDMwNDAwMw=="} +{"index":4326,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1207.0580","reference_contexts":"Furthermore, we used the method of dropout (Hinton et al., 2012b) instead of weight noise during training","reference_id":"OTgzMzA1OTc5NzYzNTg3NTg0Mw=="} +{"index":4327,"citer_arxiv_id":"1312.6026","cited_arxiv_id":"1301.3584","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4328,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1312.6114","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4329,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1412.6583","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4330,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1511.05644","reference_contexts":"The quality of our unsupervised disentangled representation matches previous works that made use of supervised label information [5] [6] [7] [8] [9]","reference_id":"MTE3Nzg4OTQ1ODYyMTQxMzQ3NDg="} +{"index":4331,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1511.05644","reference_contexts":"Similarly, VAEs [5] and Adversarial Autoencoders [9] were shown to learn representations in which class label is separated from other variations.","reference_id":"Nzc4MTA0Mzk3NDU5Mzc5MTI1NA=="} +{"index":4332,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1210.5474","reference_contexts":"To the best of our knowledge, the only other unsupervised method that learns disentangled representations is hossRBM [13] , a higher-order extension of the spike-and-slab restricted Boltzmann machine that can disentangle emotion from identity on the Toronto Face Dataset [25]","reference_id":"MTU4MTUxMDM4MjMyNTQ1NjE2Mjk="} +{"index":4333,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1210.5474","reference_contexts":"Early methods were based on stacked (often denoising) autoencoders or restricted Boltzmann machines [10] [11] [12] [13]","reference_id":"NjAyNTY5MzM1NjczODEyMjM2Mw=="} +{"index":4334,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1301.3781","reference_contexts":"A lot of promising recent work originates from the Skip-gram model [14] , which inspired the skip-thought vectors [15] and several techniques for unsupervised feature learning of images [16] .","reference_id":"NjAwNTAwMjg2OTc2NTUzMzQw"} +{"index":4335,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1511.06434","reference_contexts":"Since GAN is known to be difficult to train, we design our experiments based on existing techniques introduced by DC-GAN [19] , which are enough to stabilize InfoGAN training and we did not have to introduce new trick","reference_id":"NzQ2MzI5MjM2NjkzMTI4OTkxOQ=="} +{"index":4336,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1511.06434","reference_contexts":"[19] to learn an image representation that supports basic linear algebra on code space","reference_id":"MTU5ODY4MDAxNTIxMTE4NTE1MDQ="} +{"index":4337,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1511.06390","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4338,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4339,"citer_arxiv_id":"1606.03657","cited_arxiv_id":"1502.03167","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4340,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1611.02779","reference_contexts":"[3] for the POMDP setting.","reference_id":"NzgxMTIxMTU3MjUyMDEyMjcwNw=="} +{"index":4341,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1703.03400","reference_contexts":"• We point out that first-order MAML [4] is simpler to implement than was widely recognized prior to this article.","reference_id":"NzYzNTY0NTYxMzcxNDA1Njc3OQ=="} +{"index":4342,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1703.03400","reference_contexts":"[4] use transduction for evaluating MAML.","reference_id":"MTY0NTAzNDM5OTM0MDA2MzI1OTY="} +{"index":4343,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1703.03400","reference_contexts":"[4] proposed an algorithm called MAML, which directly optimizes performance with respect to this initialization-differentiating through the fine-tuning process","reference_id":"OTgwNzAzMDc1Mzc0NTQ1Mzk5MQ=="} +{"index":4344,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1703.03400","reference_contexts":"We consider the optimization problem of MAML [4] : find an initial set of parameters, φ, such that for a randomly sampled task τ with corresponding loss L τ , the learner will have low loss after k updates","reference_id":"NTQ1NjA4NTgzMzQ4MjQ1ODM4Ng=="} +{"index":4345,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1703.03400","reference_contexts":"[4]","reference_id":"Njc3MzY0NTIwMzc2ODcxODQ0NA=="} +{"index":4346,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1703.03400","reference_contexts":"MAML results are from [4]","reference_id":"Mzg2ODI1MDM3NDkzNTUwNTQwOA=="} +{"index":4347,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1703.03400","reference_contexts":"[4] ), since the training-time optimization resembles the test-time optimization (where the test set doesn't overlap with the training set)","reference_id":"MTg3MjI3NzU2NDU2MDU1NzE1Nw=="} +{"index":4348,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1502.03167","reference_contexts":"In our transductive experiments, information was shared between the test samples via batch normalization [9]","reference_id":"MTM0MzE2NzM2MTcyMjk0ODAwNzM="} +{"index":4349,"citer_arxiv_id":"1803.02999","cited_arxiv_id":"1710.10304","reference_contexts":"• Evaluating Reptile on the task of few-shot density modeling [14] .","reference_id":"MTQ1MzA0NjM0MDk4NjEwMjg5NTU="} +{"index":4350,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1606.04474","reference_contexts":"Another line of work (Hochreiter et al., 2001; Younger et al., 2001; Andrychowicz et al., 2016; Li & Malik, 2016) studies meta-learning over the optimization process","reference_id":"MTI0NzE3MzQ1ODMyNDQ4MjY2MzQ="} +{"index":4351,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1409.1259","reference_contexts":"To alleviate the difficulty of training RNNs due to vanishing and exploding gradients (Bengio et al., 1994) , we use Gated Recurrent Units (GRUs) (Cho et al., 2014) which have been demonstrated to have good empirical performance (Chung et al., 2014; Józefowicz et al., 2015)","reference_id":"MTAxMzAxNjM2OTM0ODA2MTIzMzI="} +{"index":4352,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1412.3555","reference_contexts":"To alleviate the difficulty of training RNNs due to vanishing and exploding gradients (Bengio et al., 1994) , we use Gated Recurrent Units (GRUs) (Cho et al., 2014) which have been demonstrated to have good empirical performance (Chung et al., 2014; Józefowicz et al., 2015)","reference_id":"MTAxMzAxNjM2OTM0ODA2MTIzMzI="} +{"index":4353,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1605.02097","reference_contexts":"Next, we evaluate RL 2 on a vision-based navigation task implemented using the ViZDoom environment (Kempka et al., 2016) , showing that RL 2 can also scale to high-dimensional problems.","reference_id":"MTgyMTY3NzY3NjU0NjQ3NjMxMDM="} +{"index":4354,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1509.02971","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4355,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1605.09128","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4356,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1511.06295","reference_contexts":"There has been a growing interest in using deep neural networks for multi-task learning and transfer learning (Parisotto et al., 2015; Rusu et al., 2015; 2016a; Devin et al., 2016; Rusu et al., 2016b) .","reference_id":"MTM0NTk5MTg5NzU1OTk3NTQ5Ng=="} +{"index":4357,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1606.04671","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4358,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1605.06065","reference_contexts":"Our work draws inspiration from a particular line of work (Younger et al., 2001; Santoro et al., 2016; Vinyals et al., 2016) , which formulates meta-learning as an optimization problem, and can thus be optimized end-to-end via gradient descent","reference_id":"MTQ3MTEzMDg4MTUwODIxMzIxMzU="} +{"index":4359,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1502.02251","reference_contexts":"However, these methods tend to make assumptions about the environment (e.g., instrumentation for access to the state at learning time), or become computationally intractable in high-dimensional settings (Wahlström et al., 2015) .","reference_id":"MTI5ODQ1ODU0MTQxMzE1MjEzMQ=="} +{"index":4360,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1603.04467","reference_contexts":"All experiments are implemented using TensorFlow (Abadi et al., 2016) and rllab (Duan et al., 2016)","reference_id":"OTgyODM2MzgzNTQ0Nzc1ODMwNA=="} +{"index":4361,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1602.07868","reference_contexts":"For the neural network architecture, We use rectified linear units throughout the experiments as the hidden activation, and we apply weight normalization without data-dependent initialization (Salimans & Kingma, 2016) to all weight matrices","reference_id":"NzIxODA3NDIwMDM1MzUwNDgzMA=="} +{"index":4362,"citer_arxiv_id":"1611.02779","cited_arxiv_id":"1312.6120","reference_contexts":"The hidden-to-hidden weight matrix uses an orthogonal initialization (Saxe et al., 2013) , and all other weight matrices use Xavier initialization (Glorot & Bengio, 2010)","reference_id":"MTQxMzc4Nzk4MDU2NjIwODQ0"} +{"index":4363,"citer_arxiv_id":"1610.05492","cited_arxiv_id":"1606.00372","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4364,"citer_arxiv_id":"1610.05492","cited_arxiv_id":"1510.00149","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4365,"citer_arxiv_id":"1610.05492","cited_arxiv_id":"1610.02527","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4366,"citer_arxiv_id":"1610.05492","cited_arxiv_id":"1412.6806","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4367,"citer_arxiv_id":"1506.02075","cited_arxiv_id":"1410.3916","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4368,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"1508.05326","reference_contexts":"We test Visual Entailment performance on SNLI-VE (Xie et al., 2018) , a task built on top of SNLI (Bowman et al., 2015)","reference_id":"MTY4OTgzNTMxNjg1OTY3MTQ1NDM="} +{"index":4369,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2102.06171","reference_contexts":"Frozen affixes an NFResnet (Brock et al., 2021) vision encoder to a pretrained autoregressive LM and, keeping the LM weights frozen, trains the vision encoder along with a visual prefix that linearly maps the output of the vision encoder to the dimensionality of the LM's token embeddings","reference_id":"MTc5NjAxOTc4NjY5ODE3MzY0NDY="} +{"index":4370,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2102.08981","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4371,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2102.08981","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4372,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"1504.00325","reference_contexts":"We evaluate on two datasets -CoCo Captions (Chen et al., 2015) and NoCaps (Agrawal et al., 2019) , measuring performance using the BLEU@4 and CIDEr metrics","reference_id":"MTgwOTk3Mjc3ODUxMzUxMjMzNjQ="} +{"index":4373,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"1504.00325","reference_contexts":"Following research showing that LMs become strong zero-shot learners after being finetuned on collections of structured, task-based datasets (Wei et al., 2021; Sanh et al., 2021) , we also include the training splits of the following downstream tasks: VQA (Antol et al., 2015) , GQA (Hudson and Manning, 2019) , OKVQA (Marino et al., 2019) , VizWiz (Gurari et al., 2018) , Hateful Memes (Kiela et al., 2020) , CoCo Captions (Chen et al., 2015)","reference_id":"MTAxMDk3NTg0MTgxNDY2MDExMTk="} +{"index":4374,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2104.08696","reference_contexts":"This supports preliminary research indicating that the feed-forward blocks are important in storing implicit knowledge in pretrained transformers (Dai et al., 2021) .","reference_id":"MTI5MTg3OTAxMjU3MjY2Mzk0OTU="} +{"index":4375,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2010.11929","reference_contexts":"SimVLM consists of an encoder-decoder transformer with a combined ResNet (He et al., 2015) and ViT (Dosovitskiy et al., 2020) backbone as encoder, and a transformer decoder for language modeling","reference_id":"NTk0NzA3NzU1OTc5OTM5NTcwNw=="} +{"index":4376,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2101.00027","reference_contexts":"This failure mode, as well as recent research suggesting that increased training dataset diversity improves downstream generalization capabilities (Zhang et al., 2021; Radford et al., 2021; Brown et al., 2020; Gao et al., 2021) , prompted us to construct another large-scale pretraining dataset from various publicly available image-text datasets, including a heavily filtered subset of LAION (Schuhmann et al., 2021) , Wikipedia Image-Text (Srinivasan et al., 2021), CC3M (Changpinyo et al., 2021b) , Visual Genome (Krishna et al., 2016) , Localized Narratives (Pont-Tuset et al., 2020) .","reference_id":"MTY1ODE4NTc5NjU2NDg1MjQ4NzI="} +{"index":4377,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"1512.03385","reference_contexts":"SimVLM consists of an encoder-decoder transformer with a combined ResNet (He et al., 2015) and ViT (Dosovitskiy et al., 2020) backbone as encoder, and a transformer decoder for language modeling","reference_id":"NTk0NzA3NzU1OTc5OTM5NTcwNw=="} +{"index":4378,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2106.09685","reference_contexts":"Our work builds on recent advances in parameter efficient finetuning of LMs (Houlsby et al., 2019; Li and Liang, 2021; Lester et al., 2021; He et al., 2021; Hu et al., 2021) , specifically with adapter layers (Houlsby et al., 2019; He et al., 2021; Pfeiffer et al., 2020) , which are small modules inserted in between the elements of a transformer layer which are finetuned instead of the model weights as a form of parameter efficient fine-tuning.","reference_id":"MzU2NTY3NzIzMjA3NTI3MTcxMg=="} +{"index":4379,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"1902.09506","reference_contexts":"GQA (Hudson and Manning, 2019)","reference_id":"MTA5OTQ2MjI0NDcwNDIxMDY4ODQ="} +{"index":4380,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"1902.09506","reference_contexts":"Following research showing that LMs become strong zero-shot learners after being finetuned on collections of structured, task-based datasets (Wei et al., 2021; Sanh et al., 2021) , we also include the training splits of the following downstream tasks: VQA (Antol et al., 2015) , GQA (Hudson and Manning, 2019) , OKVQA (Marino et al., 2019) , VizWiz (Gurari et al., 2018) , Hateful Memes (Kiela et al., 2020) , CoCo Captions (Chen et al., 2015)","reference_id":"MTAxMDk3NTg0MTgxNDY2MDExMTk="} +{"index":4381,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2005.04790","reference_contexts":"Following research showing that LMs become strong zero-shot learners after being finetuned on collections of structured, task-based datasets (Wei et al., 2021; Sanh et al., 2021) , we also include the training splits of the following downstream tasks: VQA (Antol et al., 2015) , GQA (Hudson and Manning, 2019) , OKVQA (Marino et al., 2019) , VizWiz (Gurari et al., 2018) , Hateful Memes (Kiela et al., 2020) , CoCo Captions (Chen et al., 2015)","reference_id":"MTAxMDk3NTg0MTgxNDY2MDExMTk="} +{"index":4382,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"1910.02054","reference_contexts":"During pretraining for the ablations and all subsequent models, we update the parameters θ by minimizing the loss (3) per mini-batch using the Adam optimizer in combination with ZeRO (Rajbhandari et al., 2019) to parallelize gradients and optimizer states across devices","reference_id":"MTI0NzQyNDc1MTYxNTA0ODc0NzU="} +{"index":4383,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2103.01913","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4384,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2104.09864","reference_contexts":"Firstly, GPT-J replaces learned positional embeddings with rotary positional embeddings (Su et al., 2021) , a form of relative positional embedding","reference_id":"NTA0NTcyNTA3OTYwNjI0MDM4MQ=="} +{"index":4385,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2106.13884","reference_contexts":"As noted in (Tsimpoukelli et al., 2021) , relative positional embeddings enable the transformer to generalize to inputs with more than one image, or a different imagetext ordering compared to the training distribution, which is key to the VL model's ability to perform in-context learning with multiple image examples","reference_id":"MTU3OTI4ODM0ODg3MzcyNTE5NDg="} +{"index":4386,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2106.13884","reference_contexts":"A recent line of work (Tsimpoukelli et al., 2021; Wang et al., 2021; Sollami and Jain, 2021) explores VL modeling using autoregressive decoder models trained with a language modeling objective","reference_id":"MTU1NDM1Mjg0NTcyNzE4NTE3Mzc="} +{"index":4387,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2106.13884","reference_contexts":"Frozen (Tsimpoukelli et al., 2021) shows that a pretrained autoregressive language model can, without any finetuning to the LM weights themselves, be harnessed to train a visual prefix which enables images to be used as its input","reference_id":"OTA0NjEzNjU3NzI3MjMzOTQyMg=="} +{"index":4388,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2106.13884","reference_contexts":"Our general approach is an image conditioned variant of soft-prompting or prefix tuning (Lester et al., 2021; Shin et al., 2020; Qin and Eisner, 2021) for language transformers and extends the Frozen method (Tsimpoukelli et al., 2021)","reference_id":"ODAzODYzNTg5NjAwMDU0ODk1NA=="} +{"index":4389,"citer_arxiv_id":"2112.05253","cited_arxiv_id":"2106.13884","reference_contexts":"Two predecessors to our method are Frozen (Tsimpoukelli et al., 2021) and SimVLM (Wang et al., 2021) , two autoregressive decoder models trained with a next token prediction language modeling objective","reference_id":"MzY2MDkyMjQ3NDc3NDIwMjM3Nw=="} +{"index":4390,"citer_arxiv_id":"1411.5654","cited_arxiv_id":"1408.5093","reference_contexts":"For the visual features v we used the 4096D 7th Layer output of BVLC reference Net [14] after ReLUs","reference_id":"MzMwODk0MDk2NDA2NjQxOTgzMQ=="} +{"index":4391,"citer_arxiv_id":"1411.5654","cited_arxiv_id":"1408.5093","reference_contexts":"We used the open source RNN code of [24] and the Caffe framework [14] to implement our model","reference_id":"MTI4NjU0Nzg0MTg4NTMxMzQ0OTE="} +{"index":4392,"citer_arxiv_id":"1411.5654","cited_arxiv_id":"1406.5679","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4393,"citer_arxiv_id":"1411.5654","cited_arxiv_id":"1410.1090","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4394,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2005.14165","reference_contexts":"Training such models, however, is challenging for two reasons: (a) it is no longer possible to fit the parameters of these models in the main memory of even the largest GPU (NVIDIA recently released 80GB-A100 cards), and (b) even if we are able to fit the model in a single GPU (e.g., by swapping parameters between host and device memory [38] ), the high number of compute operations required can result in unrealistically long training times (e.g., training GPT-3 with 175 billion parameters [11] would require approximately 288 years with a single V100 NVIDIA GPU)","reference_id":"NzQxOTU2MjA0MDYxMDY1MjIyOQ=="} +{"index":4395,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2005.14165","reference_contexts":"We use standard model architectures such as GPT-3 [11] when appropriate.","reference_id":"ODU3ODcxNzk4ODY5OTkyMDkwMA=="} +{"index":4396,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2005.14165","reference_contexts":"Figure 12 shows the per-GPU-throughput for interleaved and non-interleaved schedules on the GPT-3 [11] model with 175 billion parameters (96 layers, 96 attention heads, hidden size of 12288)","reference_id":"Mjk2NTExMjc4MTc1MTg1OTQ0Mw=="} +{"index":4397,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2005.14165","reference_contexts":"Recent work [11, 40] has shown large language models to be effective zero-or few-shot learners, with high accuracy on many NLP tasks and datasets","reference_id":"MTA2ODE5NjAzMDMzOTU2NTI5NjQ="} +{"index":4398,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2005.14165","reference_contexts":"In our experiments, we demonstrate close to linear scaling to 3072 A100 GPUs, with an achieved end-to-end training throughput of 163 teraFLOP/s per GPU (including communication, data processing, and optimization), and an aggregate throughput of 502 petaFLOP/s, on a GPT model [11] with a trillion parameters using mixed precision","reference_id":"MjEwMDAxODc0NTg4ODEyOTM0Mw=="} +{"index":4399,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1604.06174","reference_contexts":"Activation recomputation [12, 18, 20, 21] is an optional technique that trades off an increase in the number of compute operations performed for additional memory footprint, by running the forward pass a second time just before the backward pass (and stashing only the input activations for a given pipeline stage, as opposed to the entire set of intermediate activations, which is much larger)","reference_id":"MTI5MjEzOTcxNDY3MDc2OTE5NTU="} +{"index":4400,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1810.04805","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4401,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2101.03961","reference_contexts":"Switch Transformers [15] used Mesh-Tensorflow to train a sparsely activated expert-based model with 1.6 trillion parameters, with improved pre-training speed over the T5-11B model [35] .","reference_id":"MTUxNzQ0MjgzODQzODIxNTU3MTM="} +{"index":4402,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1706.02677","reference_contexts":"[17] and You et al","reference_id":"ODc5Mjg1NTc1ODkxMjM4MzM5MQ=="} +{"index":4403,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2006.15704","reference_contexts":"With data parallelism [25, 43] , each worker has a copy of the full model, the input dataset is sharded, and workers aggregate their gradients periodically to ensure that all workers see a consistent version of the weights","reference_id":"NzY1MTIzNzc4MDUwMTE5NjkxMg=="} +{"index":4404,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1907.11692","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4405,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1910.10683","reference_contexts":"Switch Transformers [15] used Mesh-Tensorflow to train a sparsely activated expert-based model with 1.6 trillion parameters, with improved pre-training speed over the T5-11B model [35] .","reference_id":"MTUxNzQ0MjgzODQzODIxNTU3MTM="} +{"index":4406,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1910.02054","reference_contexts":"ZeRO [36, 37] extends this idea: weight parameters and gradients are sharded across data-parallel workers as well, and workers fetch relevant state from their \"owning\" workers before performing computations","reference_id":"MTA0MjE3ODcyNzY3MDE2MjY2NA=="} +{"index":4407,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1910.02054","reference_contexts":"We compare PTD-P to ZeRO-3 [36, 37] in Table 2 and Figure 10 for the standard GPT-3 model architecture, as well as the 530-billionparameter model from Table 1","reference_id":"MTgwMjEwOTE4MTc3NjMzODcyNw=="} +{"index":4408,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1910.02054","reference_contexts":"Other techniques such as activation partitioning [36] can also be used in conjunction with tensor model parallelsim to reduce the memory footprint due to activations further.","reference_id":"ODgwMjI5NTkzNTc0OTIyMzc2NQ=="} +{"index":4409,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1910.02054","reference_contexts":"We also compared to ZeRO [36] , and found that our approach outperforms ZeRO-3 by 70% for models with 175 and 530 billion parameters due to less cross-node communication","reference_id":"MTU4Njc3Nzk1ODQ5MzcyMDk4MjE="} +{"index":4410,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2104.07857","reference_contexts":"ZeRO [36, 37] extends this idea: weight parameters and gradients are sharded across data-parallel workers as well, and workers fetch relevant state from their \"owning\" workers before performing computations","reference_id":"MTA0MjE3ODcyNzY3MDE2MjY2NA=="} +{"index":4411,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2104.07857","reference_contexts":"We compare PTD-P to ZeRO-3 [36, 37] in Table 2 and Figure 10 for the standard GPT-3 model architecture, as well as the 530-billionparameter model from Table 1","reference_id":"MTgwMjEwOTE4MTc3NjMzODcyNw=="} +{"index":4412,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2104.07857","reference_contexts":"ZeRO-Infinity [37] uses NVMe to efficiently swap parameters, enabling the training of very large models on a small number of GPUs","reference_id":"MTIxNjU1MTE0OTEyNzg1NTA1NjQ="} +{"index":4413,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2101.06840","reference_contexts":"Training such models, however, is challenging for two reasons: (a) it is no longer possible to fit the parameters of these models in the main memory of even the largest GPU (NVIDIA recently released 80GB-A100 cards), and (b) even if we are able to fit the model in a single GPU (e.g., by swapping parameters between host and device memory [38] ), the high number of compute operations required can result in unrealistically long training times (e.g., training GPT-3 with 175 billion parameters [11] would require approximately 288 years with a single V100 NVIDIA GPU)","reference_id":"NzQxOTU2MjA0MDYxMDY1MjIyOQ=="} +{"index":4414,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1909.08053","reference_contexts":"This result indicates that neither tensor model parallelism (used by Megatron [40] ) nor pipeline model parallelism (used by PipeDream [30] and others) in isolation can match the performance of using both techniques in conjunction.","reference_id":"MjY1Njk5NzI3NTQwODI1NDM3OA=="} +{"index":4415,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1909.08053","reference_contexts":"Recent work [11, 40] has shown large language models to be effective zero-or few-shot learners, with high accuracy on many NLP tasks and datasets","reference_id":"MTA2ODE5NjAzMDMzOTU2NTI5NjQ="} +{"index":4416,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1909.08053","reference_contexts":"We believe this is the fastest training throughput achieved for this size of model: past systems [29, 40] cannot train such large models since they do not combine pipeline and tensor parallelism","reference_id":"NTQyMDA1OTkwNTc4MjMzNjM3Mw=="} +{"index":4417,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1909.08053","reference_contexts":"Figure 5 : Blocks of transformer model partitioned with tensor model parallelism (figures borrowed from Megatron [40] )","reference_id":"MTEyNTI1NTY0Mjk1MjAwNTE0OTA="} +{"index":4418,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1909.08053","reference_contexts":"In this paper, we use the particular partitioning strategy used by Megatron [40] for transformer layers, the bedrock of language models","reference_id":"MTQwNTU2NDgzMDU3NTg2ODc3Mzc="} +{"index":4419,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1706.03762","reference_contexts":"Further details of the transformer layer can be found in Vaswani et al [42] .","reference_id":"MTAxMTUyNzYyMDMyNjE2ODE5ODA="} +{"index":4420,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"2004.13336","reference_contexts":"As part of performance optimizations for MLPerf 0.6 [28] , sharded data parallelism [24, 44] , where optimizer state is sharded over data-parallel workers, was introduced","reference_id":"MTAxODE1NzAyODI2ODIwODYyMTI="} +{"index":4421,"citer_arxiv_id":"2104.04473","cited_arxiv_id":"1906.08237","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4422,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2001.07676","reference_contexts":"It is not obvious a priori which prompts will generalize well for a given LM, and there is high variance in how well different prompts generalize [3, 6] , even between prompts with minor differences [e.g., one comma; 5]","reference_id":"NTA3NDU2ODU1MDA3Njg3NzEwNA=="} +{"index":4423,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2001.07676","reference_contexts":"Other work claims to use no validation set for hyperparameter selection [3, 11, 14] but does not describe how they design other aspects of their learning algorithm (e.g., training objectives)","reference_id":"MjUwMDY3OTE1ODYzMzg2NTM2NQ=="} +{"index":4424,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2001.07676","reference_contexts":"Several papers claim to circumvent the need to estimate EL(A, R) by choosing hyperparameters based on an educated guess [3, 9, 14]","reference_id":"MTMzNjE4MzU4MDEzNDE3Njk0NjM="} +{"index":4425,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2001.07676","reference_contexts":"Recent work on LMs performs few-shot learning by providing training examples as input in the form of a natural language \"prompt\" [2, 3, 9]","reference_id":"MTI0MTM3MTY4MzIyMzQ4OTk0Mg=="} +{"index":4426,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2001.07676","reference_contexts":"Major progress in language model (LM) pretraining has led to the idea that LMs can learn a new task using a small number of examples only, i.e., few-shot learning [1] [2] [3]","reference_id":"MTI0NTgwMzA4MDE4MjQzMDc0Mjc="} +{"index":4427,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2001.07676","reference_contexts":"Following [3] , we also vary the textual label names used to get the logits for a given output class","reference_id":"NTY3MzMyNjg5MTY3MDA3NTc4NQ=="} +{"index":4428,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2001.07676","reference_contexts":"However, the few-shot performance of LMs is very sensitive to the textual task description [\"prompt\"; 3-6], order of training examples [6] [7] [8] , decoding strategy [9, 10] , and other hyperparameters [3, 5, 9, 11, 12] , as well as the learning algorithm itself [3, 12]","reference_id":"MTU5OTQxMDcwMDk3NjEyNDUwNTc="} +{"index":4429,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2012.15723","reference_contexts":"Other work chooses the learning algorithm and hyperparameters using one or multiple other datasets before evaluating on the target dataset [5, 11]","reference_id":"NDgyOTQyMzQ3MDYwNjA5MjA4Mw=="} +{"index":4430,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2012.15723","reference_contexts":"However, the few-shot performance of LMs is very sensitive to the textual task description [\"prompt\"; 3-6], order of training examples [6] [7] [8] , decoding strategy [9, 10] , and other hyperparameters [3, 5, 9, 11, 12] , as well as the learning algorithm itself [3, 12]","reference_id":"MTU5OTQxMDcwMDk3NjEyNDUwNTc="} +{"index":4431,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2101.06804","reference_contexts":"However, the few-shot performance of LMs is very sensitive to the textual task description [\"prompt\"; 3-6], order of training examples [6] [7] [8] , decoding strategy [9, 10] , and other hyperparameters [3, 5, 9, 11, 12] , as well as the learning algorithm itself [3, 12]","reference_id":"MTU5OTQxMDcwMDk3NjEyNDUwNTc="} +{"index":4432,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2009.07118","reference_contexts":"We use FewGLUE [9] , the 32-example subset of SuperGLUE used in prior work on few-shot learning","reference_id":"ODQyNDUxODYyNDM2MzY0NjcxNw=="} +{"index":4433,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2009.07118","reference_contexts":"For each dataset, the first prompt is the one from GPT-3 [2] and the others are from [9] , modified to be compatible with left-to-right LMs when necessary.","reference_id":"MTgzNzY3Mzg1MTg3NTU1NDI4Mw=="} +{"index":4434,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2009.07118","reference_contexts":"Several papers claim to circumvent the need to estimate EL(A, R) by choosing hyperparameters based on an educated guess [3, 9, 14]","reference_id":"MTMzNjE4MzU4MDEzNDE3Njk0NjM="} +{"index":4435,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2009.07118","reference_contexts":"Recent work on LMs performs few-shot learning by providing training examples as input in the form of a natural language \"prompt\" [2, 3, 9]","reference_id":"MTI0MTM3MTY4MzIyMzQ4OTk0Mg=="} +{"index":4436,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2009.07118","reference_contexts":"However, the few-shot performance of LMs is very sensitive to the textual task description [\"prompt\"; 3-6], order of training examples [6] [7] [8] , decoding strategy [9, 10] , and other hyperparameters [3, 5, 9, 11, 12] , as well as the learning algorithm itself [3, 12]","reference_id":"MTU5OTQxMDcwMDk3NjEyNDUwNTc="} +{"index":4437,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2009.07118","reference_contexts":"As our set of prompts, we use the manually-written prompts from [2] and [9] -3 prompts for RTE/CB and 4 prompts for WiC","reference_id":"NTM2NDkzNTEwNDYxMzQyMDY1OQ=="} +{"index":4438,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"1910.01108","reference_contexts":"We test the 5-shot accuracy of 9 popular LMs of various sizes: GPT-3 [175B, 13B, 6.7B, 2.7B parameter models; 2], GPT-2 [1.5B, 782M, 345M, 117M models; 2], and DistilGPT-2 [16] , a distilled, 82M parameter version of GPT-2 117M","reference_id":"MzI0NDUxMzIwMTE0NDQzMzgwOQ=="} +{"index":4439,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"1910.01108","reference_contexts":"For prompt selection, our observation holds for 9 LMs ranging over 3 orders of magnitude in size [1, 2, 16 ] on 3 classification tasks and 41 tasks in the LAMA benchmark [17]","reference_id":"MTQ4NTAzMjU4MTAwNTIwNjM5MDY="} +{"index":4440,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"1606.06565","reference_contexts":"where Reliability is important for many practical or commercial applications where worst-case performance is important, such as image recognition [35, 36] , dialogue systems [37, 38] , and robotics [39, 40] .","reference_id":"MTU1NzI3NjE1Mjg5MDM1NDA2"} +{"index":4441,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2104.05240","reference_contexts":"In what follows, we test on LAMA [17] , a benchmark for retrieving facts with LMs, for which prior work has developed many strategies for designing prompts [4, [54] [55] [56]","reference_id":"MzUwNzk5NjQ3OTM2NDMxNjY3Mg=="} +{"index":4442,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2104.08835","reference_contexts":"The meta-learning setting assumes access to data from many other tasks in order to perform learning and model selection [20, 21, 64, 65]","reference_id":"MTc4ODg3ODIzODA3NjU5MTkxMzY="} +{"index":4443,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"1811.01088","reference_contexts":"Transfer learning and multitask learning assume access to data that is directly related to the task with limited data [66] [67] [68] [69]","reference_id":"MTEwOTA5ODAwMzgyNjkyMjMwMjQ="} +{"index":4444,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"1811.01088","reference_contexts":"This result may be due to the fact that we retrain the model using the CV-chosen hyperparameters, but finetuning pretrained LMs often has high variance in performance [68, 81]","reference_id":"MTE0MDk4MjM5MDgyMDYyNDI1OTY="} +{"index":4445,"citer_arxiv_id":"2105.11447","cited_arxiv_id":"2002.06305","reference_contexts":"This result may be due to the fact that we retrain the model using the CV-chosen hyperparameters, but finetuning pretrained LMs often has high variance in performance [68, 81]","reference_id":"MTE0MDk4MjM5MDgyMDYyNDI1OTY="} +{"index":4446,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1310.1531","reference_contexts":"Deep convolutional neural network has achieved great successes in image classification (Krizhevsky et al., 2012; Jia, 2013; Donahue et al., 2013; Simonyan et al., 2013; Sermanet et al., 2013; Zeiler & Fergus, 2013) , object detection (Girshick et al., 2014) , and image retrieval (Razavian et al., 2014; Gong et al., 2014)","reference_id":"MzMxNjg4OTM5Njc5NjU4NTAwOA=="} +{"index":4447,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1310.1531","reference_contexts":"Almost all of the recent successful recognition systems (Jia, 2013; Donahue et al., 2013; Simonyan et al., 2013; Sermanet et al., 2013; Zeiler & Fergus, 2013; Gong et al., 2014) are built on top of this architecture","reference_id":"Mzk3MzcxNDczMzIwMzgyMTU1NQ=="} +{"index":4448,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1403.6382","reference_contexts":"Deep convolutional neural network has achieved great successes in image classification (Krizhevsky et al., 2012; Jia, 2013; Donahue et al., 2013; Simonyan et al., 2013; Sermanet et al., 2013; Zeiler & Fergus, 2013) , object detection (Girshick et al., 2014) , and image retrieval (Razavian et al., 2014; Gong et al., 2014)","reference_id":"MzMxNjg4OTM5Njc5NjU4NTAwOA=="} +{"index":4449,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1312.6229","reference_contexts":"Deep convolutional neural network has achieved great successes in image classification (Krizhevsky et al., 2012; Jia, 2013; Donahue et al., 2013; Simonyan et al., 2013; Sermanet et al., 2013; Zeiler & Fergus, 2013) , object detection (Girshick et al., 2014) , and image retrieval (Razavian et al., 2014; Gong et al., 2014)","reference_id":"MzMxNjg4OTM5Njc5NjU4NTAwOA=="} +{"index":4450,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1312.6229","reference_contexts":"For example, a typical CNN (Jia, 2013; Sermanet et al., 2013; Zeiler & Fergus, 2013) that works well for object recognition contains eight layers (five convolutional layers and three dense connected layers) and a huge number of parameters (e.g","reference_id":"MTYxOTgwMzMwMTkyMzA2NjE4NzM="} +{"index":4451,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1312.6229","reference_contexts":"Almost all of the recent successful recognition systems (Jia, 2013; Donahue et al., 2013; Simonyan et al., 2013; Sermanet et al., 2013; Zeiler & Fergus, 2013; Gong et al., 2014) are built on top of this architecture","reference_id":"Mzk3MzcxNDczMzIwMzgyMTU1NQ=="} +{"index":4452,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1409.1556","reference_contexts":"Deep convolutional neural networks (Krizhevsky et al., 2012; LeCun et al., 1990; Szegedy et al., 2014; Simonyan & Zisserman, 2014) has recently achieved significant progress and have become the gold standard for object recognition, image classification, and retrieval","reference_id":"NTYyMzI3MDQ0MjU1MzI3NzgwMw=="} +{"index":4453,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1409.4842","reference_contexts":"Deep convolutional neural networks (Krizhevsky et al., 2012; LeCun et al., 1990; Szegedy et al., 2014; Simonyan & Zisserman, 2014) has recently achieved significant progress and have become the gold standard for object recognition, image classification, and retrieval","reference_id":"NTYyMzI3MDQ0MjU1MzI3NzgwMw=="} +{"index":4454,"citer_arxiv_id":"1412.6115","cited_arxiv_id":"1311.2901","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4455,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2106.15831","reference_contexts":"When pre-trained models are adapted to a specific distribution through standard fine-tuning, effective robustness deteriorates at convergence [3]","reference_id":"MTU4NjYwMzQyNzQzODI2Mzc0OTk="} +{"index":4456,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2106.15831","reference_contexts":"[3] explored several fine-tuning approaches but found that none yielded models with improved robustness at high accuracy","reference_id":"NDAxNzg3Mzk2MDMyMTE4MDE1Ng=="} +{"index":4457,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2106.15831","reference_contexts":"models has recently also been raised as an open problem by several authors [3, 9, 82, 77]","reference_id":"NzM3NTU3MzE4ODIyODMxNzA3MA=="} +{"index":4458,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2108.07258","reference_contexts":"models has recently also been raised as an open problem by several authors [3, 9, 82, 77]","reference_id":"NzM3NTU3MzE4ODIyODMxNzA3MA=="} +{"index":4459,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2005.14165","reference_contexts":"As shown in Figures 11, 12 , evaluations along the trajectory underperform solutions generated by applying WiSE-FT.","reference_id":"NDk3MzEzMTgwMjY5MjE1NjU1MA=="} +{"index":4460,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2005.14165","reference_contexts":"[12] extensively discuss the broader impact of large zero-shot models and identify potential causes of harm including model biases and potential malicious uses such as surveillance systems","reference_id":"MTc2NzkxMjkxOTgzOTM3MDQz"} +{"index":4461,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2005.14165","reference_contexts":"Pre-training on large amounts of data is a powerful technique for building high-performing machine learning systems [90, 21, 50, 107, 81, 12]","reference_id":"MTYyNzI3Mzg4MDMxODYyNzA2NzQ="} +{"index":4462,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1311.3618","reference_contexts":"[52] : CIFAR-10, CIFAR-100 [54] , Describable Textures [14] , Food-101 [10] , SUN397 [103] , and Stanford Cars [53] .","reference_id":"MTA5NDI1Njg5MDAwODczMDMzNjA="} +{"index":4463,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1311.3618","reference_contexts":"Accuracy gain over the zero-shot model [17, 54, 14, 10, 103, 53]","reference_id":"MTgyNTEwMzQwNDQyNTE5MjM3NzU="} +{"index":4464,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1311.3618","reference_contexts":"[52] : CIFAR-10, CIFAR-100 [54] , Describable Textures [14] , Food-101 [10] , SUN397 [103] , and Stanford Cars [53]","reference_id":"MTM5ODU1MjkwOTYxMTUyMDMyOA=="} +{"index":4465,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.02918","reference_contexts":"special data augmentation [19, 22, 28, 36] or adversarially robust models [65, 15, 85, 87] .","reference_id":"NTA0NjM3OTA0NDcxNDcwNTQ3"} +{"index":4466,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2011.03395","reference_contexts":"[16] show that accuracy under certain distribution shifts cannot be reliably inferred from accuracy on the reference distribution","reference_id":"MTE4NzAzMzE3ODIzMDU4MDc5Mzc="} +{"index":4467,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2006.06666","reference_contexts":"One increasingly popular class of vision models are those pre-trained with auxiliary language supervision, which can be used for zero-shot inference [18, 86, 111, 82, 45, 77, 109]","reference_id":"NDE5NDQ5NDY2MzE2NDA1MzMzOA=="} +{"index":4468,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1708.04552","reference_contexts":"special data augmentation [19, 22, 28, 36] or adversarially robust models [65, 15, 85, 87] .","reference_id":"NTA0NjM3OTA0NDcxNDcwNTQ3"} +{"index":4469,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"[21]","reference_id":"NTE4OTA3OTAwMDE5OTg1MDM2NA=="} +{"index":4470,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"Finally, we demonstrate that our findings are more broadly applicable by exploring WiSE-FT for BASIC [77] , ALIGN [45] , and a ViT-H/14 [21] model pre-trained on JFT-300M [93] .","reference_id":"Mjk1NDUyNDI0MTExNDAyMjI2"} +{"index":4471,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"Next, we consider the application of WiSE-FT to a ViT-H/14 model [21] pre-trained on JFT-300M [93] , where the zero-shot classifier is constructed by manually identifying a class correspondence (details provided in Section C.7.2)","reference_id":"MTM5OTkzMzY1NTE5NDg3MTUwMTI="} +{"index":4472,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"α) 87.1 (+0.9) 99.5 (+0.8) 93.4 (+1.2) 93.6 (+2.0) 85.2 (+3.3) 83.3 (+2.6) 96.2 (+1.8) ImageNet (top-1, %) , a ViT-H/14 [21] model pre-trained on JFT-300M [93] and ALIGN [45] .","reference_id":"MTUyMjAyMjA3MDkxMDMyMDk0OTA="} +{"index":4473,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"Pre-training on large amounts of data is a powerful technique for building high-performing machine learning systems [90, 21, 50, 107, 81, 12]","reference_id":"MTYyNzI3Mzg4MDMxODYyNzA2NzQ="} +{"index":4474,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"Section C.7 explores the application of WiSE-FT for additional models such as ALIGN [45] , a ViT-H/14 model pre-trained on JFT [21] and BASIC [77]","reference_id":"MzI1Njg3ODcxODQ2MTk1MDc4OA=="} +{"index":4475,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"We primarily explore CLIP models [82] , although we also investigate other zero-shot models including ALIGN [45] , BASIC [77] and a ViT model pre-trained on JFT [21]","reference_id":"MzM3MDk5NDEyMzQzNDkxNjg5NA=="} +{"index":4476,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.11929","reference_contexts":"Figure 23 : WiSE-FT applied to ViT-H/14 [21] pre-trained on JFT","reference_id":"MTAyODU3MDkwODU0MDY4Mzk0MTQ="} +{"index":4477,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1707.08945","reference_contexts":"Previous work has studied model behavior under synthetic [35, 99, 65, 29, 23, 2] and natural distribution shift [37, 49, 100, 4, 38]","reference_id":"MTY2NTQzNTUwMzY4MDQ3ODAxNDM="} +{"index":4478,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1811.12231","reference_contexts":"special data augmentation [19, 22, 28, 36] or adversarially robust models [65, 15, 85, 87] .","reference_id":"NTA0NjM3OTA0NDcxNDcwNTQ3"} +{"index":4479,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1808.08750","reference_contexts":"Previous work has studied model behavior under synthetic [35, 99, 65, 29, 23, 2] and natural distribution shift [37, 49, 100, 4, 38]","reference_id":"MTY2NTQzNTUwMzY4MDQ3ODAxNDM="} +{"index":4480,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1808.08750","reference_contexts":"[35, 8, 7, 29, 2] ; and (ii) natural, where samples are not perturbed after acquisition and changes in data distributions arise through naturally occurring variations in lighting, geographic location, crowdsourcing process, image styles, etc","reference_id":"MTEyNjUyMTgwNTA0MjYzMzg3ODE="} +{"index":4481,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.01434","reference_contexts":"In certain cases, ensemble accuracy is correlated with diversity among the constituents [57, 30]","reference_id":"NTc5NzgzMjQxNDI2OTE2NDg3Mw=="} +{"index":4482,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1412.6544","reference_contexts":"Weight-space ensembles linearly interpolate between the weights of different models [25, 64, 32, 95]","reference_id":"MTE1MjA3MzIxOTg2ODUzNDYxMjU="} +{"index":4483,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1512.03385","reference_contexts":"• ObjectNet [4] , a test set of objects in various scenes with 113 classes overlapping with ImageNet [38] , a test set of natural images misclassified by a ResNet-50 [34] for 200 ImageNet classes.","reference_id":"NDcwMTgxNTI5NDAwODQ4NDM0OQ=="} +{"index":4484,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1903.12261","reference_contexts":"Previous work has studied model behavior under synthetic [35, 99, 65, 29, 23, 2] and natural distribution shift [37, 49, 100, 4, 38]","reference_id":"MTY2NTQzNTUwMzY4MDQ3ODAxNDM="} +{"index":4485,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1903.12261","reference_contexts":"[35, 8, 7, 29, 2] ; and (ii) natural, where samples are not perturbed after acquisition and changes in data distributions arise through naturally occurring variations in lighting, geographic location, crowdsourcing process, image styles, etc","reference_id":"MTEyNjUyMTgwNTA0MjYzMzg3ODE="} +{"index":4486,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2006.16241","reference_contexts":"Previous work has studied model behavior under synthetic [35, 99, 65, 29, 23, 2] and natural distribution shift [37, 49, 100, 4, 38]","reference_id":"MTY2NTQzNTUwMzY4MDQ3ODAxNDM="} +{"index":4487,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2006.16241","reference_contexts":"[97, 83, 37, 38, 49]","reference_id":"NDI3NzEzODkzMDQzOTczMjc1Mg=="} +{"index":4488,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2006.16241","reference_contexts":"Specifically, we present our key results for five natural distribution shifts derived from ImageNet (i.e., S tr ref is ImageNet): • ImageNet-V2 (IN-V2) [83] , a reproduction of the ImageNet test set with distribution shift • ImageNet-R (IN-R) [37] , renditions (e.g., sculptures, paintings) for 200 ImageNet classes","reference_id":"MTYwMDkxODE1NTI4NDM1MTI2MTc="} +{"index":4489,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2006.16241","reference_contexts":"Motivated by the above concerns, we propose a robust way of fine-tuning zero-shot models that addresses the aforementioned trade-off and achieves the best of both worlds: increased performance under distribution shift Figure 2 : Samples of the class lemon, from the reference distribution ImageNet [17] and the derived distribution shifts considered in our main experiments: ImageNet-V2 [83] , ImageNet-R [37] , ImageNet Sketch [100] , ObjectNet [4] , and ImageNet-A [38] .","reference_id":"MTgzNTI1NDgyOTM0MjQyOTQ3ODM="} +{"index":4490,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1907.07174","reference_contexts":"Previous work has studied model behavior under synthetic [35, 99, 65, 29, 23, 2] and natural distribution shift [37, 49, 100, 4, 38]","reference_id":"MTY2NTQzNTUwMzY4MDQ3ODAxNDM="} +{"index":4491,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1907.07174","reference_contexts":"[97, 83, 37, 38, 49]","reference_id":"NDI3NzEzODkzMDQzOTczMjc1Mg=="} +{"index":4492,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1907.07174","reference_contexts":"• ObjectNet [4] , a test set of objects in various scenes with 113 classes overlapping with ImageNet [38] , a test set of natural images misclassified by a ResNet-50 [34] for 200 ImageNet classes.","reference_id":"NDcwMTgxNTI5NDAwODQ4NDM0OQ=="} +{"index":4493,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1907.07174","reference_contexts":"Motivated by the above concerns, we propose a robust way of fine-tuning zero-shot models that addresses the aforementioned trade-off and achieves the best of both worlds: increased performance under distribution shift Figure 2 : Samples of the class lemon, from the reference distribution ImageNet [17] and the derived distribution shifts considered in our main experiments: ImageNet-V2 [83] , ImageNet-R [37] , ImageNet Sketch [100] , ObjectNet [4] , and ImageNet-A [38] .","reference_id":"MTgzNTI1NDgyOTM0MjQyOTQ3ODM="} +{"index":4494,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1503.02531","reference_contexts":"Network distillation [41] trains one network to match the outputs of another","reference_id":"MjA1ODc5Njk1NzE1OTAxMTExMQ=="} +{"index":4495,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1803.05407","reference_contexts":"We are not the first to observe that when interpolating between models, the accuracy of models along the path may exceed that of either endpoint [43, 73, 102]","reference_id":"MTE5ODE3NzU4MTMxMzM1MjYxNTI="} +{"index":4496,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1803.05407","reference_contexts":"However, as similarly observed by previous work where part of the optimization trajectory is shared [43, 25, 73] , we find that the zero-shot and fine-tuned models are connected by a linear path in weight-space along which accuracy remains high (explored further in Section 5.2).","reference_id":"MTQ1NzA3MTcxNDE4MTU3NzUwNw=="} +{"index":4497,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1803.05407","reference_contexts":"Moreover, previous work has suggested that interpolation in weight space may improve performance when models share part of their optimization trajectory [43, 73] .","reference_id":"OTQyNDUyMTkwOTA3MDQ2ODM1NQ=="} +{"index":4498,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1803.05407","reference_contexts":"[43] when part of the training trajectory is shared, and by Neyshabur et al","reference_id":"MTU4MTE2OTc0MTc1NDQ2MTE2NTY="} +{"index":4499,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1803.05407","reference_contexts":"[43] average checkpoints saved throughout training for improved performance","reference_id":"MTI4MDAyOTI1OTYxNDQxNTAwNjM="} +{"index":4500,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1806.07572","reference_contexts":"In certain regimes a weight-space ensemble approximates the corresponding output-space ensemble-for instance, when training is well approximated by a linear expansion, referred to as the NTK regime [44]","reference_id":"MTMwNjE5MjYxNzgwNjkyODE2NzM="} +{"index":4501,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"While these negative results highlight the difficulty of learning robust models, large pre-trained models such as CLIP [82] , ALIGN [45] and BASIC [77] have recently demonstrated unprecedented robustness to these challenging distribution shifts","reference_id":"MTEwNjk4ODgzOTExMTg1MzE0NjI="} +{"index":4502,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"Figure 22 : WiSE-FT applied to ALIGN [45]","reference_id":"Nzg0ODAwNjgyNTY3NjYyMzkyNQ=="} +{"index":4503,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"[45] , which is similar to CLIP but is pre-trained with a different dataset, finding similar trends.","reference_id":"MTAyNTIwNDQ3Mzk2OTU2ODk1NTM="} +{"index":4504,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"Finally, we demonstrate that our findings are more broadly applicable by exploring WiSE-FT for BASIC [77] , ALIGN [45] , and a ViT-H/14 [21] model pre-trained on JFT-300M [93] .","reference_id":"Mjk1NDUyNDI0MTExNDAyMjI2"} +{"index":4505,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"α) 87.1 (+0.9) 99.5 (+0.8) 93.4 (+1.2) 93.6 (+2.0) 85.2 (+3.3) 83.3 (+2.6) 96.2 (+1.8) ImageNet (top-1, %) , a ViT-H/14 [21] model pre-trained on JFT-300M [93] and ALIGN [45] .","reference_id":"MTUyMjAyMjA3MDkxMDMyMDk0OTA="} +{"index":4506,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"One increasingly popular class of vision models are those pre-trained with auxiliary language supervision, which can be used for zero-shot inference [18, 86, 111, 82, 45, 77, 109]","reference_id":"NDE5NDQ5NDY2MzE2NDA1MzMzOA=="} +{"index":4507,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"We primarily explore CLIP models [82] , although we also investigate other zero-shot models including ALIGN [45] , BASIC [77] and a ViT model pre-trained on JFT [21]","reference_id":"MzM3MDk5NDEyMzQzNDkxNjg5NA=="} +{"index":4508,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"Section C.7 explores the application of WiSE-FT for additional models such as ALIGN [45] , a ViT-H/14 model pre-trained on JFT [21] and BASIC [77]","reference_id":"MzI1Njg3ODcxODQ2MTk1MDc4OA=="} +{"index":4509,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2102.05918","reference_contexts":"In addition to CLIP, we show WiSE-FT to be effective for an additional zero-shot model, ALIGN [45]","reference_id":"NTc1NDEwMDk4MTA1NzkwMDIyNw=="} +{"index":4510,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1412.6980","reference_contexts":"The first variant is a debiased EMA, where debiasing is done as in Kingma and Ba [47] (Algorithm 1)","reference_id":"ODM4MDQyNjY1ODk3NTkyNTAwMA=="} +{"index":4511,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1612.00796","reference_contexts":"[48] ; Zenke et al","reference_id":"NDAyMTQ3OTc0MTk2MTU4NTc3Ng=="} +{"index":4512,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1912.11370","reference_contexts":"Pre-training on large amounts of data is a powerful technique for building high-performing machine learning systems [90, 21, 50, 107, 81, 12]","reference_id":"MTYyNzI3Mzg4MDMxODYyNzA2NzQ="} +{"index":4513,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1805.08974","reference_contexts":"[52] : CIFAR-10, CIFAR-100 [54] , Describable Textures [14] , Food-101 [10] , SUN397 [103] , and Stanford Cars [53] .","reference_id":"MTA5NDI1Njg5MDAwODczMDMzNjA="} +{"index":4514,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1805.08974","reference_contexts":"[52] : CIFAR-10, CIFAR-100 [54] , Describable Textures [14] , Food-101 [10] , SUN397 [103] , and Stanford Cars [53]","reference_id":"MTM5ODU1MjkwOTYxMTUyMDMyOA=="} +{"index":4515,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1608.03983","reference_contexts":"Linear classifiers are trained with mini-batch optimization, using the AdamW optimizer [61, 76] with a cosine-annealing learning rate schedule [60]","reference_id":"MTY2Njg0ODQ0NjEzMTI3NjAxNA=="} +{"index":4516,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1608.03983","reference_contexts":"Unless otherwise mentioned, we use the default PyTorch AdamW hyperparameters β 1 = 0.9, β 2 = 0.999, = 10 -8 , weight decay of 0.1 and a cosine-annealing learning rate schedule [60] with 500 warm-up steps","reference_id":"OTI2MjQxOTY4NTk0MTc1ODMwMQ=="} +{"index":4517,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1706.06083","reference_contexts":"special data augmentation [19, 22, 28, 36] or adversarially robust models [65, 15, 85, 87] .","reference_id":"NTA0NjM3OTA0NDcxNDcwNTQ3"} +{"index":4518,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1706.06083","reference_contexts":"Previous work has studied model behavior under synthetic [35, 99, 65, 29, 23, 2] and natural distribution shift [37, 49, 100, 4, 38]","reference_id":"MTY2NTQzNTUwMzY4MDQ3ODAxNDM="} +{"index":4519,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.09832","reference_contexts":"Finally, Matena and Raffel [66] introduce a method of combining models in weight-space that goes beyond linear interpolation with a single mixing-coefficient as employed in WiSE-FT","reference_id":"MTU4MzgwMzc1MjgzMDQzNzExMDE="} +{"index":4520,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2004.14444","reference_contexts":"In contrast, accuracy on the reference distribution is often a reliable predictor for accuracy under distribution shift [106, 69, 97, 94, 70]","reference_id":"MTY5MzYxNzQ3MTI3NjEyMTAwNTE="} +{"index":4521,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1803.02999","reference_contexts":"Indeed, averaging the weights along the training trajectory is a central method in optimization [84, 78, 74]","reference_id":"MTMyOTYxNjI0NjQ3MzM0NjAxMjc="} +{"index":4522,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1912.01703","reference_contexts":"Linear classifiers are trained with mini-batch optimization, using the AdamW optimizer [61, 76] with a cosine-annealing learning rate schedule [60]","reference_id":"MTY2Njg0ODQ0NjEzMTI3NjAxNA=="} +{"index":4523,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1912.01703","reference_contexts":"When fine-tuning end-to-end, unless otherwise mentioned, we use the AdamW optimizer [61, 76] and choose the largest batch size such that the model fits into 8 GPUs (512 for ViT-B/16)","reference_id":"NTM2MTcwOTUwNzMxNzg3NDk1MQ=="} +{"index":4524,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"While these negative results highlight the difficulty of learning robust models, large pre-trained models such as CLIP [82] , ALIGN [45] and BASIC [77] have recently demonstrated unprecedented robustness to these challenging distribution shifts","reference_id":"MTEwNjk4ODgzOTExMTg1MzE0NjI="} +{"index":4525,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"Finally, we demonstrate that our findings are more broadly applicable by exploring WiSE-FT for BASIC [77] , ALIGN [45] , and a ViT-H/14 [21] model pre-trained on JFT-300M [93] .","reference_id":"Mjk1NDUyNDI0MTExNDAyMjI2"} +{"index":4526,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"One increasingly popular class of vision models are those pre-trained with auxiliary language supervision, which can be used for zero-shot inference [18, 86, 111, 82, 45, 77, 109]","reference_id":"NDE5NDQ5NDY2MzE2NDA1MzMzOA=="} +{"index":4527,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"We primarily explore CLIP models [82] , although we also investigate other zero-shot models including ALIGN [45] , BASIC [77] and a ViT model pre-trained on JFT [21]","reference_id":"MzM3MDk5NDEyMzQzNDkxNjg5NA=="} +{"index":4528,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"Section C.7 explores the application of WiSE-FT for additional models such as ALIGN [45] , a ViT-H/14 model pre-trained on JFT [21] and BASIC [77]","reference_id":"MzI1Njg3ODcxODQ2MTk1MDc4OA=="} +{"index":4529,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"First, we interpolate between the weights of the zero-shot and fine-tuned BASIC-L model [77] , finding that α=0.5 improves average accuracy on five distribution shifts derived from ImageNet by over 7 pp while improving ImageNet accuracy by 0.4 pp relative to the fine-tuned BASIC-L model (a per-dataset breakdown is provided in Figure 24 and Table 12 of the Appendix)","reference_id":"MTM3OTU4MTEyMjE5NzIwNjcwMzE="} +{"index":4530,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"models has recently also been raised as an open problem by several authors [3, 9, 82, 77]","reference_id":"NzM3NTU3MzE4ODIyODMxNzA3MA=="} +{"index":4531,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"We apply WiSE-FT to BASIC [77] , fine-tuning both the image and text encoder with a contrastive loss on half of the ImageNet training data, as in Pham et al","reference_id":"NDA3NzEwMjgyNTI5OTA5OTMyOQ=="} +{"index":4532,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.10050","reference_contexts":"[77] , fine-tuning comes at the cost of robustness: across several natural distribution shifts, the accuracy of their fine-tuned models is lower than that of the original zero-shot model","reference_id":"NDM1NTkyNzI3NTY0NTc5NTEwMA=="} +{"index":4533,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"While these negative results highlight the difficulty of learning robust models, large pre-trained models such as CLIP [82] , ALIGN [45] and BASIC [77] have recently demonstrated unprecedented robustness to these challenging distribution shifts","reference_id":"MTEwNjk4ODgzOTExMTg1MzE0NjI="} +{"index":4534,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"[82] in using minimal data augmentation","reference_id":"MzU5MjYyMDc0NzAyNjAxMTExNg=="} +{"index":4535,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"One increasingly popular class of vision models are those pre-trained with auxiliary language supervision, which can be used for zero-shot inference [18, 86, 111, 82, 45, 77, 109]","reference_id":"NDE5NDQ5NDY2MzE2NDA1MzMzOA=="} +{"index":4536,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"We primarily explore CLIP models [82] , although we also investigate other zero-shot models including ALIGN [45] , BASIC [77] and a ViT model pre-trained on JFT [21]","reference_id":"MzM3MDk5NDEyMzQzNDkxNjg5NA=="} +{"index":4537,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"[82] at ImageNet scale with similar accuracy","reference_id":"MTY2MjI0NjcxNTE4OTg3ODkzNzU="} +{"index":4538,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"[82] and Pham et al","reference_id":"MTc4NjcyMTY0NDI3MjA5MDg2MDY="} +{"index":4539,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"(2) i =\"a picture of a {c i }\" (referred to as prompt ensembling [82] ).","reference_id":"MTc4MTA5MjUxNTk2MzE1NjQzMTk="} +{"index":4540,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"models has recently also been raised as an open problem by several authors [3, 9, 82, 77]","reference_id":"NzM3NTU3MzE4ODIyODMxNzA3MA=="} +{"index":4541,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"[82] and Brown et al","reference_id":"MTIxNjcyMDc2MzM2OTA4MTk2Nzk="} +{"index":4542,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"[82] , while maintaining or improving ImageNet accuracy","reference_id":"MTcxNDI0NzgwNjczMDkyNzgzNjI="} +{"index":4543,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"Additionally, unless otherwise mentioned we use the same data augmentations as [82] , randomly cropping a square from resized images with the largest dimension being 336 pixels for ViT-L/14@336px and 224 for the remaining models.","reference_id":"NjAyOTk1MTM2Mzk2NDg4OTg2NA=="} +{"index":4544,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2103.00020","reference_contexts":"However, this adaptation to class shift improves performance by 2.3% [82]","reference_id":"ODg3MDQ2ODA2Mjc2MTU1MzQzNw=="} +{"index":4545,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"Concretely, we consider: (i) ImageNet-Vid-Robust and YTBB-Robust, datasets with distribution shift induced by temporal perturbations in videos [88] ; (ii) CIFAR-10.1 [83] and CIFAR-10.2 [62] , reproductions of the popular image classification dataset CIFAR-10 [54] with a distribution shift; (iii) WILDS-FMoW, a satellite image recognition task where the test set has a geographic and temporal distribution shift [49, 13] ; (iv) WILDS-iWildCam, a wildlife recognition task where the test set has a geographic distribution shift [49, 6] .","reference_id":"MjA1MDQyODA3MzI5NzQ1NTI4MA=="} +{"index":4546,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"Beyond the five distribution shifts derived from ImageNet, WiSE-FT consistently improves robustness on a diverse set of further distributions shifts including geographic shifts in satellite imagery and wildlife recognition (WILDS-FMoW [49, 13] , WILDS-iWildCam [49, 6] ), reproductions of the popular image classification dataset CIFAR-10 [54] with a distribution shift (CIFAR-10.1 [83] and CIFAR-10.2 [62] ), and datasets with distribution shift induced by temporal perturbations in videos (ImageNet-Vid-Robust and YTBB-Robust [89] )","reference_id":"NDYwNDk5MjAwNDExODIyNDI4NQ=="} +{"index":4547,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"[97, 83, 37, 38, 49]","reference_id":"NDI3NzEzODkzMDQzOTczMjc1Mg=="} +{"index":4548,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"Motivated by the above concerns, we propose a robust way of fine-tuning zero-shot models that addresses the aforementioned trade-off and achieves the best of both worlds: increased performance under distribution shift Figure 2 : Samples of the class lemon, from the reference distribution ImageNet [17] and the derived distribution shifts considered in our main experiments: ImageNet-V2 [83] , ImageNet-R [37] , ImageNet Sketch [100] , ObjectNet [4] , and ImageNet-A [38] .","reference_id":"MTgzNTI1NDgyOTM0MjQyOTQ3ODM="} +{"index":4549,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"Figure 9 : WiSE-FT accuracy under distribution shift relative to standard fine-tuning on ImageNet-Vid-Robust, YTBB-Robust [88] , CIFAR-10.1 [83] , CIFAR-10.2 [62] , WILDS-FMoW [49, 13] , and WILDS-iWildCam [49, 6] .","reference_id":"NjU3NzMyNDM5Njk2NzgwNTI5NQ=="} +{"index":4550,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"A useful tool for studying (effective) robustness are scatter plots that illustrate model performance under distribution shift [83, 97]","reference_id":"NTMwMzA2OTY0OTYxNzMxMjEzMg=="} +{"index":4551,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"Table 6 : WiSE-FT improves results on ImageNet-Vid-Robust, YTBB-Robust [88] , CIFAR-10.1 [83] , CIFAR-10.2 [62] , WILDS-FMoW [49, 13] , and WILDS-iWildCam [49, 6]","reference_id":"ODAzNDQ2ODczNzU0MDI0MzMxNQ=="} +{"index":4552,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1902.10811","reference_contexts":"Specifically, we present our key results for five natural distribution shifts derived from ImageNet (i.e., S tr ref is ImageNet): • ImageNet-V2 (IN-V2) [83] , a reproduction of the ImageNet test set with distribution shift • ImageNet-R (IN-R) [37] , renditions (e.g., sculptures, paintings) for 200 ImageNet classes","reference_id":"MTYwMDkxODE1NTI4NDM1MTI2MTc="} +{"index":4553,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1906.04584","reference_contexts":"special data augmentation [19, 22, 28, 36] or adversarially robust models [65, 15, 85, 87] .","reference_id":"NTA0NjM3OTA0NDcxNDcwNTQ3"} +{"index":4554,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2008.01392","reference_contexts":"One increasingly popular class of vision models are those pre-trained with auxiliary language supervision, which can be used for zero-shot inference [18, 86, 111, 82, 45, 77, 109]","reference_id":"NDE5NDQ5NDY2MzE2NDA1MzMzOA=="} +{"index":4555,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1906.02168","reference_contexts":"Concretely, we consider: (i) ImageNet-Vid-Robust and YTBB-Robust, datasets with distribution shift induced by temporal perturbations in videos [88] ; (ii) CIFAR-10.1 [83] and CIFAR-10.2 [62] , reproductions of the popular image classification dataset CIFAR-10 [54] with a distribution shift; (iii) WILDS-FMoW, a satellite image recognition task where the test set has a geographic and temporal distribution shift [49, 13] ; (iv) WILDS-iWildCam, a wildlife recognition task where the test set has a geographic distribution shift [49, 6] .","reference_id":"MjA1MDQyODA3MzI5NzQ1NTI4MA=="} +{"index":4556,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1906.02168","reference_contexts":"Table 6 : WiSE-FT improves results on ImageNet-Vid-Robust, YTBB-Robust [88] , CIFAR-10.1 [83] , CIFAR-10.2 [62] , WILDS-FMoW [49, 13] , and WILDS-iWildCam [49, 6]","reference_id":"ODAzNDQ2ODczNzU0MDI0MzMxNQ=="} +{"index":4557,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1906.02168","reference_contexts":"Figure 9 : WiSE-FT accuracy under distribution shift relative to standard fine-tuning on ImageNet-Vid-Robust, YTBB-Robust [88] , CIFAR-10.1 [83] , CIFAR-10.2 [62] , WILDS-FMoW [49, 13] , and WILDS-iWildCam [49, 6] .","reference_id":"NjU3NzMyNDM5Njk2NzgwNTI5NQ=="} +{"index":4558,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1403.6382","reference_contexts":"Pre-training on large amounts of data is a powerful technique for building high-performing machine learning systems [90, 21, 50, 107, 81, 12]","reference_id":"MTYyNzI3Mzg4MDMxODYyNzA2NzQ="} +{"index":4559,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1912.04838","reference_contexts":"In contrast, accuracy on the reference distribution is often a reliable predictor for accuracy under distribution shift [106, 69, 97, 94, 70]","reference_id":"MTY5MzYxNzQ3MTI3NjEyMTAwNTE="} +{"index":4560,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1512.00567","reference_contexts":"In our experiments, this frontier is reached only through methods that average model weights, either using WiSE-FT or with a more sophisticated averaging scheme: keeping an exponential moving average of all model iterates (EMA, [95] )","reference_id":"NTMzMzUxNzA3NTA0MTUwNzk0Mg=="} +{"index":4561,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1512.00567","reference_contexts":"Weight-space ensembles linearly interpolate between the weights of different models [25, 64, 32, 95]","reference_id":"MTE1MjA3MzIxOTg2ODUzNDYxMjU="} +{"index":4562,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"Effective robustness [97] is accuracy beyond this baseline, defined formally as ρ(f","reference_id":"MzgzMTA0MDM0NjM5MTcwMzg5NQ=="} +{"index":4563,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"To assist in contextualizing the results, the scatter plots we display also show a wide range of machine learning models from a comprehensive testbed of evaluations [97, 70] , including: models trained on S tr D (standard training); models trained on additional data and fine-tuned using S tr D (trained with more data); and models trained using various existing robustness interventions, e.g","reference_id":"NTU4NTEwNjk4MzcwNTAzMTAzOQ=="} +{"index":4564,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"Interventions used for synthetic shifts do not typically provide robustness to many natural distribution shifts [97]","reference_id":"MTEyOTkyNjY5MTczNjQ1MzE5MDI="} +{"index":4565,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"[97, 83, 37, 38, 49]","reference_id":"NDI3NzEzODkzMDQzOTczMjc1Mg=="} +{"index":4566,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"Over the past few years, researchers have proposed a variety of distribution shifts on which current algorithmic approaches to enhance robustness yield little to no gains [97, 70]","reference_id":"ODcxMjQ3NzU3MzgzOTY4NDI2NQ=="} +{"index":4567,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"In contrast, accuracy on the reference distribution is often a reliable predictor for accuracy under distribution shift [106, 69, 97, 94, 70]","reference_id":"MTY5MzYxNzQ3MTI3NjEyMTAwNTE="} +{"index":4568,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"[97] categorized distribution shifts into two broad categories: (i) synthetic, e.g., ℓ ∞ -adversarial examples or artificial changes in image contrast, brightness, etc","reference_id":"OTQwNjUzMzI1MTU3OTAzNTE0NQ=="} +{"index":4569,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"[97] demonstrated that no current algorithmic robustness interventions provide consistent gains across the distribution shifts where zero-shot models excel.","reference_id":"MTY1NjU5OTc4NDY5NjQ4MjQwMjU="} +{"index":4570,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"A useful tool for studying (effective) robustness are scatter plots that illustrate model performance under distribution shift [83, 97]","reference_id":"NTMwMzA2OTY0OTYxNzMxMjEzMg=="} +{"index":4571,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"For the distribution shifts we study, accuracy on the reference distribution is a reliable predictor of accuracy under distribution shift [97, 70]","reference_id":"OTgxMTU4OTE2MzA4NjU3NDkyNg=="} +{"index":4572,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"Empirically, when applying logit (or probit) axis scaling, models trained on the reference distribution approximately lie on a linear trend [97, 70]","reference_id":"NDQ3MjcxNzk5MzcyODcwNTgyNg=="} +{"index":4573,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"[97]","reference_id":"MTA3OTY3OTE2MDMzMzkwMjYzOTE="} +{"index":4574,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2007.00644","reference_contexts":"[97] observed that no algorithmic robustness intervention consistently achieves substantial effective robustness across the distribution shifts in Figure 2 -the first method to do so was zero-shot CLIP","reference_id":"MTU1MTc1MTkyNjM5ODk5MzI2Mzg="} +{"index":4575,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1905.13549","reference_contexts":"• ImageNet Sketch (IN-Sketch) [100] , which contains sketches instead of natural images","reference_id":"NDE3OTkzMTU1NDM2MTU0ODA1MQ=="} +{"index":4576,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1905.13549","reference_contexts":"Previous work has studied model behavior under synthetic [35, 99, 65, 29, 23, 2] and natural distribution shift [37, 49, 100, 4, 38]","reference_id":"MTY2NTQzNTUwMzY4MDQ3ODAxNDM="} +{"index":4577,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1905.13549","reference_contexts":"Motivated by the above concerns, we propose a robust way of fine-tuning zero-shot models that addresses the aforementioned trade-off and achieves the best of both worlds: increased performance under distribution shift Figure 2 : Samples of the class lemon, from the reference distribution ImageNet [17] and the derived distribution shifts considered in our main experiments: ImageNet-V2 [83] , ImageNet-R [37] , ImageNet Sketch [100] , ObjectNet [4] , and ImageNet-A [38] .","reference_id":"MTgzNTI1NDgyOTM0MjQyOTQ3ODM="} +{"index":4578,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1911.04252","reference_contexts":"Concretely, we ensemble zero-shot CLIP with two Noisy Student EfficientNet models [104, 96] : (i) EfficientNet-B6 (Figure 25 , left), with performance on the reference distribution comparable to the end-to-end fine-tuned CLIP model; and (ii) EfficientNet-L2 (Figure 25 , right), the strongest model available on PyTorch ImageNet Models [101]","reference_id":"MTc4NjA4ODQ1MjU2NTkwOTA3NTc="} +{"index":4579,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1905.00546","reference_contexts":"Pre-training on large amounts of data is a powerful technique for building high-performing machine learning systems [90, 21, 50, 107, 81, 12]","reference_id":"MTYyNzI3Mzg4MDMxODYyNzA2NzQ="} +{"index":4580,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2111.07991","reference_contexts":"One increasingly popular class of vision models are those pre-trained with auxiliary language supervision, which can be used for zero-shot inference [18, 86, 111, 82, 45, 77, 109]","reference_id":"NDE5NDQ5NDY2MzE2NDA1MzMzOA=="} +{"index":4581,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1907.08610","reference_contexts":"[110] propose optimizing with a set of fast and slow weights, where every k steps, these two sets of weights are averaged and a new trajectory begins","reference_id":"ODcxMDM4NzAzMjI3MDMzMDU1MA=="} +{"index":4582,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2010.00747","reference_contexts":"One increasingly popular class of vision models are those pre-trained with auxiliary language supervision, which can be used for zero-shot inference [18, 86, 111, 82, 45, 77, 109]","reference_id":"NDE5NDQ5NDY2MzE2NDA1MzMzOA=="} +{"index":4583,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2109.01134","reference_contexts":"Additional comparisons are also presented in Appendix C.3, including distillation, additional regularization, and CoOp [112]","reference_id":"NDg0NTQ4NDI5MjkyNTAyMzc0"} +{"index":4584,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2109.01134","reference_contexts":"CoOp [112] 71 Hyperparameter: Fix learning rate, vary number of epochs","reference_id":"MzYyNDE2MDM4NTI5NzAxNDA0MQ=="} +{"index":4585,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2109.01134","reference_contexts":"Finally, Figure 14 and Table 7 demonstrate that WiSE-FT achieves better accuracy than the recently proposed CoOp method [112] on ImageNet and four derived distribution shifts","reference_id":"Nzk0NDY2ODA2Mjk5NDA3NTA1"} +{"index":4586,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"2109.01134","reference_contexts":"Section C.3 compares WiSE-FT with additional alternatives including distillation and CoOp [112]","reference_id":"MzU0MTYyNjk5MzI5ODE2NzQxOA=="} +{"index":4587,"citer_arxiv_id":"2109.01903","cited_arxiv_id":"1909.11764","reference_contexts":"In natural language processing, previous work proposed stable fine-tuning methods that incur computational overhead [46, 113] , alleviating problems such as representational collapse [1]","reference_id":"MTgwMTQ1ODkyODIxNzY3Nzg2Njg="} +{"index":4588,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2103.14030","reference_contexts":"Modern computer vision neural networks are heavily parametrized: they routinely have tens or hundreds of millions of parameters [1, 2, 3, 4]","reference_id":"MTAzOTA2MjQ2ODMyMjI5OTQ4MTY="} +{"index":4589,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2103.14030","reference_contexts":"[3] .","reference_id":"MTAwNjg1MTg1NTA5MTQ3MDU2OTQ="} +{"index":4590,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1412.7062","reference_contexts":"This simple approach has led to impressive results, which are state-of-the-art in many tasks such as detection [8, 9] , segmentation [10] and action recognition [11]","reference_id":"MTI0OTIwNzYzODIyMTM2OTM1MDI="} +{"index":4591,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2006.09882","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4592,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2104.14294","reference_contexts":"In this preliminary analysis, we consider the recent method BEiT and SplitMask, our variant as detailed in Section 4, as representatives of denoising autoencoders, and DINO [18] of a joint embedding method, respectively.","reference_id":"MTgzODU2NTY4ODU4NzQyNDU0NzA="} +{"index":4593,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2104.14294","reference_contexts":"As it was already observed in previous work [18, 19, 68] , we also see in many cases that self-supervised training outperforms supervised pre-training on ImageNet","reference_id":"MTc4ODg1MDgxMTk3NjcwNTc0MDc="} +{"index":4594,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2104.14294","reference_contexts":"It was later shown that when using a momentum encoder [16] , simpler loss functions that did not directly discriminate against other images could be used [18, 20, 41, 42, 43]","reference_id":"Mzk0NzU2NjkzODUwNzY5NDQ3NA=="} +{"index":4595,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2104.14294","reference_contexts":"Most notably, the contrastive and joint embedding approaches [16, 17, 18, 19, 20] can serve as effective pre-training strategies","reference_id":"OTY0NDE5MzIwMTM2ODc2MDI0NQ=="} +{"index":4596,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2002.05709","reference_contexts":"As the performance of these methods depends on the number of negatives, it either requires large batches or memory banks to work well [16, 19, 40]","reference_id":"Njc1MDM3ODYyMjYzNDk0MTgxNQ=="} +{"index":4597,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2002.05709","reference_contexts":"As it was already observed in previous work [18, 19, 68] , we also see in many cases that self-supervised training outperforms supervised pre-training on ImageNet","reference_id":"MTc4ODg1MDgxMTk3NjcwNTc0MDc="} +{"index":4598,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2002.05709","reference_contexts":"Following previous work [19] , we symmetrize the contrastive loss, and apply it similarly on the representation x b from the subset B","reference_id":"MTIyNDM0ODE0MzY5NzQzNDgyNTg="} +{"index":4599,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2002.05709","reference_contexts":"Most notably, the contrastive and joint embedding approaches [16, 17, 18, 19, 20] can serve as effective pre-training strategies","reference_id":"OTY0NDE5MzIwMTM2ODc2MDI0NQ=="} +{"index":4600,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2006.07733","reference_contexts":"It was later shown that when using a momentum encoder [16] , simpler loss functions that did not directly discriminate against other images could be used [18, 20, 41, 42, 43]","reference_id":"Mzk0NzU2NjkzODUwNzY5NDQ3NA=="} +{"index":4601,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2006.07733","reference_contexts":"Most notably, the contrastive and joint embedding approaches [16, 17, 18, 19, 20] can serve as effective pre-training strategies","reference_id":"OTY0NDE5MzIwMTM2ODc2MDI0NQ=="} +{"index":4602,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2007.13916","reference_contexts":"Some of the most effective transformations, like cropping, rely on the images being object centric [21]","reference_id":"MTYwNDg1NzI2NTc5NDA4NjIwOA=="} +{"index":4603,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2103.01988","reference_contexts":"YFCC or IG) data using instance discrimination, this usually requires order of magnitude more data than Ima-geNet [22, 51]","reference_id":"MTA1NTU1MjAzOTE3MjA3MTgxMjg="} +{"index":4604,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2103.01988","reference_contexts":"When applied on uncurated data, these methods degrade significantly and require larger datasets to obtain similar performance [22]","reference_id":"OTk5NDA0MTY3ODIwNjgzMjE4Nw=="} +{"index":4605,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1810.04805","reference_contexts":"In particular, the (masked) language modeling loss has been applied to transformer networks, leading to the BERT model [23] , which is now the foundation of most NLP models","reference_id":"MTA5MTQ3NDU5MDI0MTM2NDU5NzE="} +{"index":4606,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1810.04805","reference_contexts":"In natural language processing, denoising autoencoders have been applied by masking or randomly replacing some tokens of the input, and reconstructing the original sequence, leading to the BERT model [23]","reference_id":"MTQzNDA0ODQ1NDU1NjkxNDQ2MDI="} +{"index":4607,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2106.08254","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4608,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1910.10683","reference_contexts":"On the other hand, it was shown that for many natural language processing tasks, increasing the size of the pre-training dataset did not lead to strong improvement when using denoising autoencoders [33]","reference_id":"MTc2MDI1MjgwMTI0OTI2NTMxOTI="} +{"index":4609,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1910.10683","reference_contexts":"Similar methods have been proposed to pre-train sequence-to-sequence models, by considering additional kind of noises such as word shuffling or deleting [33, 34] .","reference_id":"MzkyMDUxMDM4OTYxNjExNDk3Mw=="} +{"index":4610,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1910.13461","reference_contexts":"Similar methods have been proposed to pre-train sequence-to-sequence models, by considering additional kind of noises such as word shuffling or deleting [33, 34] .","reference_id":"MzkyMDUxMDM4OTYxNjExNDk3Mw=="} +{"index":4611,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2102.12092","reference_contexts":"predict the discretized label of the missing patches corresponding to their visual words as defined by a pre-trained discrete VAE [37] .","reference_id":"MTQ5Njc3OTQxNjc5MjU4Mzk0MzE="} +{"index":4612,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2105.04906","reference_contexts":"It was later shown that when using a momentum encoder [16] , simpler loss functions that did not directly discriminate against other images could be used [18, 20, 41, 42, 43]","reference_id":"Mzk0NzU2NjkzODUwNzY5NDQ3NA=="} +{"index":4613,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1911.05371","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4614,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2006.09882","reference_contexts":"Finally, a related line of work is to use clustering techniques to pre-train deep neural networks [44, 45, 46, 47, 48] .","reference_id":"MTU3Njg5MDE0MjY5NTMwOTg1NzA="} +{"index":4615,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2012.12877","reference_contexts":"Initially, only vision transformers pre-trained on very large collections obtained good performance, but smaller models trained on ImageNet with heavy augmentation can also yield competitive tradeoffs [50] .","reference_id":"MTUxOTU0MjM1ODg0MTI2NDM5MDI="} +{"index":4616,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1904.13132","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4617,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2111.06377","reference_contexts":"[58] propose MAE","reference_id":"MTg2MDk1NzEwNzQ2NDEwNjM1"} +{"index":4618,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2111.06377","reference_contexts":"[58] ), as well as using random greyscale, solarization, Gaussian blur and color jittering as additional forms of data augmentation.","reference_id":"MTc2ODg4NDU4ODg4OTEyNjAxMjk="} +{"index":4619,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1807.03748","reference_contexts":"We then apply the InfoNCE loss [59] over these representations:","reference_id":"OTc1MDE2NjIwMzY1MjMyOTE5OA=="} +{"index":4620,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"1811.08883","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4621,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2104.02057","reference_contexts":"As it was already observed in previous work [18, 19, 68] , we also see in many cases that self-supervised training outperforms supervised pre-training on ImageNet","reference_id":"MTc4ODg1MDgxMTk3NjcwNTc0MDc="} +{"index":4622,"citer_arxiv_id":"2112.10740","cited_arxiv_id":"2104.02057","reference_contexts":"[68] .","reference_id":"MTgxNDk3NjExODI3Mzk1Nzg5MTI="} +{"index":4623,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.00546","reference_contexts":"[2] outperform the EfficientNets from Noisy Student","reference_id":"MTc2MDkyMDU3MzgyMTQ2NjE1MzM="} +{"index":4624,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.00546","reference_contexts":"The sws models are from [2]","reference_id":"MTM4ODMyNDM4NzQxMjM3NTE0MzU="} +{"index":4625,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.00546","reference_contexts":"This shows that the EfficientNet Noisy student [8] tends to overfit and does not generalize as well as the (prior) semisupervised work [2] or other works of the literature","reference_id":"MTU5OTYzODE5MzIyNTg4MzQwMw=="} +{"index":4626,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.00546","reference_contexts":"[2] .","reference_id":"MTQzNjY2MzIyNzM1MzcxODU4NTM="} +{"index":4627,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.00546","reference_contexts":"We verify this hypothesis and quantify its extent by comparing the relative accuracy of this approach with another semi-supervised approach [2] both on ImageNet and ImageNet-V2 [17] .","reference_id":"MzU3NTg3MTE1NzA4MzU3ODI3Mw=="} +{"index":4628,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1805.09501","reference_contexts":"Over the years, training and testing pre-processing procedures have evolved, but so far they have been optimized separately [3]","reference_id":"Mjc2MzUwNTcyOTkyNTU0NzA0"} +{"index":4629,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.11946","reference_contexts":"We apply this method to the recent EfficientNet [4] architecture, which offers an excellent compromise between number of parameters and accuracy","reference_id":"MTE3MjAwOTc5NTkyMzc0ODY3Ng=="} +{"index":4630,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.11946","reference_contexts":"We focus on the EfficientNet [4] architectures","reference_id":"MTI5OTU1ODEwMzQ1MjY4NzE4NDQ="} +{"index":4631,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1905.11946","reference_contexts":"In this note, we focus on the EfficientNet architecture [4] due to its good accuracy/cost trade-off and its popularity.","reference_id":"ODQ1ODQ4NzY0NzQ4NzA4ODY2Mw=="} +{"index":4632,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"2006.07159","reference_contexts":"[5]","reference_id":"NDIzMTYxMjY5NjMwNjkxNjgyNQ=="} +{"index":4633,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"2006.07159","reference_contexts":"• We report the best accuracy (top1: 88.5%) with external data on ImageNet, and with ImageNet with Reallabels [5] ;","reference_id":"MTc2MzQ5NDUwNTA5NjQ2MDA1MjU="} +{"index":4634,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1811.06965","reference_contexts":"Recent research in image classification tends towards larger networks and higher resolution images [6, 7, 8]","reference_id":"MTMzNDA5ODg3MzIwNTMxNzY4NTc="} +{"index":4635,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"Since we use pre-trained EfficientNet for our initialization, our results are comparable to those from the Noisy Student [8] , which uses the same degree of overfitting, but not directly with other semi-supervised approaches like that of Yalniz et al","reference_id":"MTAwMzU5ODYwMzQwMTk4Njk5ODY="} +{"index":4636,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"This shows that the EfficientNet Noisy student [8] tends to overfit and does not generalize as well as the (prior) semisupervised work [2] or other works of the literature","reference_id":"MTU5OTYzODE5MzIyNTg4MzQwMw=="} +{"index":4637,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"Additionally, the ImageNet validation images were used to filter the images from the unlabelled set [8]","reference_id":"MTMzNzU1MDQwMTIyNzk5MTgzMzQ="} +{"index":4638,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"Recent research in image classification tends towards larger networks and higher resolution images [6, 7, 8]","reference_id":"MTMzNDA5ODg3MzIwNTMxNzY4NTc="} +{"index":4639,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"In the literature, wo versions provide the best performance: Efficient-Net trained with adversarial examples [9] , and Efficient-Net trained with Noisy student [8] pre-trained in a weaklysupervised fashion on 300 million unlabeled images","reference_id":"OTg0NTI4MDcxMDU1MzA2MDYwOQ=="} +{"index":4640,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"As mentioned earlier, several choices in the Noisy Student [8] method are prone to overfitting","reference_id":"MTE2NTM2ODgzNjQ4MTAwMDY0MzU="} +{"index":4641,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"For example, FixResNeXt-101 32x4d [1] has the same performance as EfficientNet-B3 [8] on ImageNet but on ImageNet-V2 FixResNeXt-101 32x4d [1] is better (+0.7% Top-1 accuracy).","reference_id":"MTM4MzE5ODg3ODc2OTAwNjEwOTE="} +{"index":4642,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1911.04252","reference_contexts":"For instance, the state-of-the-art in the ImageNet ILSVRC 2012 benchmark is currently held by the EfficientNet-L2 [8] architecture with 480M parameters using 800×800 images for training","reference_id":"Nzk3NDI0NjMxNzEwODkwODkxMQ=="} +{"index":4643,"citer_arxiv_id":"2003.08237","cited_arxiv_id":"1909.13719","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4644,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1905.13736","reference_contexts":"Our finding is consistent with arguments that using unlabeled data can improve adversarial robustness [11, 77, 57, 97]","reference_id":"MTcwMzgwOTI4ODQ4MDMzODIzOTA="} +{"index":4645,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1909.13719","reference_contexts":"To noise the student, we use input noise such as Ran-dAugment data augmentation [18] and model noise such as dropout [76] and stochastic depth [37] during training.","reference_id":"ODU5NDQ2MTAxNjc0NDQzMTE0Mw=="} +{"index":4646,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1909.13719","reference_contexts":"For input noise, we use data augmentation with Ran-dAugment [18]","reference_id":"MTYxODM5Nzk1MDQyODg2MjkyNjE="} +{"index":4647,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1909.13719","reference_contexts":"We use stochastic depth [37], dropout [76], and RandAugment [18] to noise the student","reference_id":"NjYyNDcxMTEyNzk1NzM0NzU5NQ=="} +{"index":4648,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1907.07174","reference_contexts":"ResNet-101 [32] 4.7% -ResNeXt-101 [32] (32x4d) 5.9% -ResNet-152 [32] 6.1% -ResNeXt-101 [32] (64x4d) 7.3% -DPN-98 [32] 9.4% -ResNeXt-101+SE [32] (32x4d) 14.2% -ResNeXt-101 WSL [55, 59] 61.0% -EfficientNet-L2 49.6% 78.6% Noisy Student Training (L2) 83.7% 95.2% Table 4: Robustness results on ImageNet-C","reference_id":"MTYxMjM5MzI5NTc3NzM2NTc4Mjk="} +{"index":4649,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1907.07174","reference_contexts":"ImageNet-A test set [32] consists of difficult images that cause significant drops in accuracy to state-of-the-art models","reference_id":"NTM2MTUxMjI0ODU3Nzg5MzI0NQ=="} +{"index":4650,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1907.07174","reference_contexts":"Not only our method improves standard ImageNet accuracy, it also improves classification robustness on much harder test sets by large margins: ImageNet-A [32] top-1 accuracy from 61.0% to 83.7%, ImageNet-C [31] mean corruption error (mCE) from 45.7 to 28.3 and ImageNet-P [31] mean flip rate (mFR) from 27.8 to 12.2","reference_id":"MTY1NDAyOTM0MjE4NjQ4NTY1NDU="} +{"index":4651,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1503.02531","reference_contexts":"First, we run an EfficientNet-B0 trained on Ima-geNet [83] over the JFT dataset [33, 15] to predict a label for each image","reference_id":"MTMzMDE0NjkzNzE1MzY0MTE0NTc="} +{"index":4652,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1503.02531","reference_contexts":"The algorithm is an improved version of self-training, a method in semi-supervised learning (e.g., [71, 96] ), and distillation [33]","reference_id":"NzIxNjIwNTM4OTMxMjY4NjEyNg=="} +{"index":4653,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1503.02531","reference_contexts":"Our work is also related to methods in Knowledge Distillation [10, 3, 33, 21, 6] via the use of soft targets","reference_id":"NjEzMjg1Nzk2NzU5Mjk2NzM1Nw=="} +{"index":4654,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1503.02531","reference_contexts":"We obtain unlabeled images from the JFT dataset [33, 15] , which has around 300M images","reference_id":"MTczNTU0MjQ2NTM5Mjk2OTUwNTc="} +{"index":4655,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1503.02531","reference_contexts":"This makes our method different from Knowledge Distillation [33] where 1) noise is often not used and 2) a smaller student model is often used to be faster than the teacher","reference_id":"NzQxMzczNzU2NzgwOTk3MzU3MQ=="} +{"index":4656,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1312.5402","reference_contexts":"Among them, augmentation Contrast and Brightness are also used in ResNeXt-101 WSL [55] and in vision models that uses the Inception preprocessing [34, 80]","reference_id":"NDE1MjczMTM2Nzg5ODUyNDA1MQ=="} +{"index":4657,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1609.02907","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4658,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1912.11370","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4659,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1602.05473","reference_contexts":"Finally, frameworks in semi-supervised learning also include graph-based methods [102, 89, 94, 42] , methods that make use of latent variables as target variables [41, 53, 95] and methods based on low-density separation [26, 70, 19] , which might provide complementary benefits to our method","reference_id":"MTE3MTUxODIyNjA5ODU1ODgyNjk="} +{"index":4660,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1811.07056","reference_contexts":"We filter the Ima-geNet validation set images from the dataset (see [58] ).","reference_id":"MTE5MTM3NzE2MjExNjk3MTI4MzE="} +{"index":4661,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1511.06709","reference_contexts":"Back translation and self-training have led to significant improvements in machine translation [72, 20, 28, 14, 90, 29] .","reference_id":"NjY1NDE3OTMxODM2NDA3NjY5Ng=="} +{"index":4662,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1312.6199","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4663,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1904.12848","reference_contexts":"First, data augmentation is an important noising method in Noisy Student Training because it forces the student to ensure prediction consistency across augmented versions of an image (similar to UDA [91] )","reference_id":"MTc4NzE0MDQxMjcxNTIyMTM4Mzg="} +{"index":4664,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1904.12848","reference_contexts":"Apart from self-training, another important line of work in semi-supervised learning [12, 103] is based on consistency training [5, 64, 47, 84, 56, 52, 62, 13, 16, 60, 2, 49, 88, 91, 8, 98, 46, 7]","reference_id":"MTI3MTkyNjQ0ODgyOTgzMDQwOTI="} +{"index":4665,"citer_arxiv_id":"1911.04252","cited_arxiv_id":"1904.12848","reference_contexts":"Other Techniques -Noisy Student Training also works better with an additional trick: data filtering and balancing, similar to [91, 93]","reference_id":"NzU0NjAxMDk2NDQzNjEyODE3NA=="} +{"index":4666,"citer_arxiv_id":"1506.02438","cited_arxiv_id":"1509.02971","reference_contexts":"In concurrent work, researchers have been developing policy gradient methods that involve differentiation with respect to the continuous-valued action (Lillicrap et al., 2015; Heess et al., 2015)","reference_id":"MTM0NDI3MzMyNjI1NDEyNzQ3MTM="} +{"index":4667,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2209.12356","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4668,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2301.07597","reference_contexts":"However, despite its powerful abilities, anecdotal reports on ChatGPT consistently showed remaining challenges -for example, it fails in some elementary mathematical (Gilson et al., 2022; Goldberg, 2023; Frieder et al., 2023; Choi et al., 2023; Davis, 2023b) and commonsense reasoning tasks (Guo et al., 2023; Davis, 2023b) ; it hallucinates with human-like fluency and eloquence on things that are not based on truth (Shen et al., 2023; Thorp, 2023; Smith, 2023) ; and as a generalpurpose language model trained from everything on the web, its language coverage is questionable (Lu et al., 2022; Jiao et al., 2023)","reference_id":"MjQ4NjcxMjg2NjI1Mjc1NjM5Mg=="} +{"index":4669,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.10403","reference_contexts":"Logical Reasoning Inductive, deductive, and abductive reasoning are common forms of logical reasoning, a process of deriving a conclusion or judgment based on given evidence or past experience and observations (Rogers et al., 2022; Wason and Johnson-Laird, 1972; Huang and Chang, 2022)","reference_id":"NTI3ODkzMDYwMzg0OTYwMjU3NQ=="} +{"index":4670,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.10403","reference_contexts":"Recent works show that LLMs perform impressively well on commonsense reasoning benchmarks (Qiao et al., 2022; Huang and Chang, 2022; Bhargava and Ng, 2022)","reference_id":"MzA5Mjc0OTk5MjA5MDg4Mzg2"} +{"index":4671,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.10403","reference_contexts":"Reasoning is one of the most actively discussed and debated abilities of LLMs as scaling the model parameter size also increases the implicit knowledge in LLMs (Wei et al., 2022a; Wang et al., 2022; Huang and Chang, 2022)","reference_id":"NDkxMDU4MzQ5Mzk4MTk4MjUxNg=="} +{"index":4672,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.10403","reference_contexts":"Causal Reasoning Causal reasoning is the process of identifying the relationship between causes/actions and effects/changes (i.e., causality) (Thomason, 2018; Huang and Chang, 2022)","reference_id":"NDE0NDYyNTA5NzE2NTUyOTU5MA=="} +{"index":4673,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.14882","reference_contexts":"Responsible Generative AI Previous works have discussed the ethical implications or concerns associated with ChatGPT (and other LLMs) (Jabotinsky and Sarel, 2022; Susnjak, 2022; Blanco-Gonzalez et al., 2022; Aydın and Karaarslan, 2022; Jeblick et al., 2022)","reference_id":"NTYwMzY1MzUzNzU4MzYxNzU2"} +{"index":4674,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.14882","reference_contexts":"(2023) show that \"significantly below those of an average mathematics graduate student.\" There are many investigations of ChatGPT's understanding and potential applications in different fields such as law (Choi et al., 2023) , medical domain (Blanco-Gonzalez et al., 2022; Jeblick et al., 2022) and finance (Birch, 2022; Dowling and Lucey, 2023)","reference_id":"MjQzNTAyMjAxNTAwMzIzNzAzMg=="} +{"index":4675,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.09597","reference_contexts":"We follow (Qiao et al., 2022) categorization on the deductive and inductive reasoning datasets, but we only use the QA part of EntailmentBank, that the authors took from ARC dataset (Clark et al., 2018) , as we aim to test for reasoning capability","reference_id":"NTk0NzU5ODI4NDA0NjgxODc1Mw=="} +{"index":4676,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.09597","reference_contexts":"Other works focus on more specific abilities such as mathematical skills (Davis, 2023b) , reasoning (Webb et al., 2022a; Qiao et al., 2022)","reference_id":"MTUxNDE5NDc2NzAyNjY0MjA3NjM="} +{"index":4677,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.09597","reference_contexts":"Inductive and deductive are categorized by \"a degree to which the premise supports the conclusion\" based on logic and philosophy (Qiao et al., 2022; Rogers et al., 2022; Hawthorne, 2021)","reference_id":"MTY0MTI5MDM3MTA0Nzg3MTQ4Mjg="} +{"index":4678,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.09597","reference_contexts":"Recent works show that LLMs perform impressively well on commonsense reasoning benchmarks (Qiao et al., 2022; Huang and Chang, 2022; Bhargava and Ng, 2022)","reference_id":"MzA5Mjc0OTk5MjA5MDg4Mzg2"} +{"index":4679,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.09597","reference_contexts":"This is in line with the anecdotal experience of users with ChatGPT -some of the examples demonstrate surprisingly good \"reasoning\" abilities compared to previously introduced LLMs but at the same time ChatGPT fails in very simple reasoning problems (Desk, 2023a; Venuto, 2023; Qiao et al., 2022; Cookup.ai, 2022; Labs, 2022)","reference_id":"MjUzMTQzMjUxMTM3MTA4NTE5Mg=="} +{"index":4680,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2302.06476","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4681,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2206.04615","reference_contexts":"the task in the BIG-bench benchmark (Srivastava et al., 2022) nal)","reference_id":"NDgxNzIzNDIxNDY1MTk3NzkyMg=="} +{"index":4682,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2206.04615","reference_contexts":"This limitation of language identification in LMs aligns with the result from BIGbench (Srivastava et al., 2022) .","reference_id":"MTI1MjM0NTAxOTU0MTA3ODc5ODM="} +{"index":4683,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2206.04615","reference_contexts":"Recently, several benchmarks with a collection of a large number of NLP datasets, such as BIG-Bench (Srivastava et al., 2022) and AI LM Harness (Gao et al., 2021) , have been introduced","reference_id":"MTI4MDI4OTQwNTkzMzc5MjM2Mjk="} +{"index":4684,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.09292","reference_contexts":"The most active discussion is over the use of academic writing and exam integrity (Jabotinsky and Sarel, 2022; Susnjak, 2022)","reference_id":"MzQ1MDYwNzM0NjY1ODQ2MzI2NQ=="} +{"index":4685,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2212.09292","reference_contexts":"Responsible Generative AI Previous works have discussed the ethical implications or concerns associated with ChatGPT (and other LLMs) (Jabotinsky and Sarel, 2022; Susnjak, 2022; Blanco-Gonzalez et al., 2022; Aydın and Karaarslan, 2022; Jeblick et al., 2022)","reference_id":"NTYwMzY1MzUzNzU4MzYxNzU2"} +{"index":4686,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"1811.00937","reference_contexts":"Based on our evaluation with CommonsenseQA (Talmor et al., 2018 ), PiQA (Bisk et al., 2020) and Pep-3k (Wang et al., 2018b) , ChatGPT shows surprisingly good commonsense reasoning capability, perhaps due to its large parametric memory.","reference_id":"MTQwNzg0MjMzODEwNjk4NzIwMTA="} +{"index":4687,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"1811.00937","reference_contexts":"To evaluate ChatGPT's capability on commonsense reasoning, we first test it on two widely used benchmark datasets CommonsenseQA (Talmor et al., 2018) and PiQA (Bisk et al., 2020)","reference_id":"ODA5MDYwNjA4MDg5MTAyNDY0OA=="} +{"index":4688,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2201.08239","reference_contexts":"To ensure factuality, it is possible to build LLMs with an interface to an external knowledge source, like Blenderbot 3.0 (Shuster et al., 2022) and LaMDa (Thoppilan et al., 2022)","reference_id":"MTExNjI0MDAyOTMzODQ0NTk3MA=="} +{"index":4689,"citer_arxiv_id":"2302.04023","cited_arxiv_id":"2203.11171","reference_contexts":"Reasoning is one of the most actively discussed and debated abilities of LLMs as scaling the model parameter size also increases the implicit knowledge in LLMs (Wei et al., 2022a; Wang et al., 2022; Huang and Chang, 2022)","reference_id":"NDkxMDU4MzQ5Mzk4MTk4MjUxNg=="} +{"index":4690,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2303.08774","reference_contexts":"The newest large models not only reach unprecedented performance on reasoning benchmarks (Achiam et al., 2023) , but they also demonstrate multimodal and multilingual capabilities (Gemini Team, 2024) and even the ability to use context lengths of over 1M tokens (Gemini Team, 2024) .","reference_id":"MTM4MzU5OTQyMjM4NTk0NDQ0Mzk="} +{"index":4691,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2305.13245","reference_contexts":"Grouped-Query Attention (Ainslie et al., 2023)","reference_id":"NDkwOTU2OTgyODQ0NzI4MTcwNA=="} +{"index":4692,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2305.10403","reference_contexts":"To study susceptibility to such attacks and quantify memorization, we evaluate models for verbatim and approximate memorization as was done in several prior studies (Anil et al., 2023; Carlini et al., 2022; Gemini Team, 2024; Kudugunta et al., 2023) .","reference_id":"MTcyNjMyNDMxNzEzNzkzMDU1NzM="} +{"index":4693,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2108.07732","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021) While thorough testing of our models has been conducted, these tests cannot cover all applications and scenarios in which Gemma 2 may be used","reference_id":"MTc3Njg2NTcyNTA5MzQyNjY5MjM="} +{"index":4694,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2004.05150","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4695,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2004.05150","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4696,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2005.14165","reference_contexts":"Large language models (LLMs) have demonstrated strong capabilities in language understanding, generation, and reasoning (Brown et al., 2020; Radford et al., 2019; Raffel et al., 2019)","reference_id":"MjUwODQ2MDY5MDM5ODU3OTU0Mg=="} +{"index":4697,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2005.14165","reference_contexts":"Scaling has been key to this recent progress, with many new capabilities only emerging at scale (Brown et al., 2020)","reference_id":"NDE3MzQ0MTE5NzQ5OTY2NTY5Mg=="} +{"index":4698,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2202.07646","reference_contexts":"To study susceptibility to such attacks and quantify memorization, we evaluate models for verbatim and approximate memorization as was done in several prior studies (Anil et al., 2023; Carlini et al., 2022; Gemini Team, 2024; Kudugunta et al., 2023) .","reference_id":"MTcyNjMyNDMxNzEzNzkzMDU1NzM="} +{"index":4699,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2107.03374","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021) While thorough testing of our models has been conducted, these tests cannot cover all applications and scenarios in which Gemma 2 may be used","reference_id":"MTc3Njg2NTcyNTA5MzQyNjY5MjM="} +{"index":4700,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"1905.10044","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021) While thorough testing of our models has been conducted, these tests cannot cover all applications and scenarios in which Gemma 2 may be used","reference_id":"MTc3Njg2NTcyNTA5MzQyNjY5MjM="} +{"index":4701,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2110.14168","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021) While thorough testing of our models has been conducted, these tests cannot cover all applications and scenarios in which Gemma 2 may be used","reference_id":"MTc3Njg2NTcyNTA5MzQyNjY5MjM="} +{"index":4702,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2009.03300","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021) While thorough testing of our models has been conducted, these tests cannot cover all applications and scenarios in which Gemma 2 may be used","reference_id":"MTc3Njg2NTcyNTA5MzQyNjY5MjM="} +{"index":4703,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"1503.02531","reference_contexts":"In particular, we focus our efforts on knowledge distillation (Hinton et al., 2015) , which replaces the one-hot vector seen at each token with the distribution of potential next tokens computed from a large model","reference_id":"MTQ2Nzk3NTE3OTk3ODEzOTQyMzA="} +{"index":4704,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2203.15556","reference_contexts":"This approach only scales logarithmically with dataset size (Hoffmann et al., 2022) , and the latest small models require up to 15T tokens to improve the state of the art by less than 1-2% (AI@Meta, 2024).","reference_id":"MTUxOTU3ODM3MjQ5MzI3MTU2MDE="} +{"index":4705,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2203.15556","reference_contexts":"Concretely, we use a large language model as a teacher to train small models, namely 2B and 9B models, on a quantity of tokens that is more than 50× the compute-optimal quantity predicted by the theory (Hoffmann et al., 2022)","reference_id":"MTYzNDg1MjY5MzM0NzA1OTcx"} +{"index":4706,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2210.17546","reference_contexts":"We compare the overall memorization rates, across a uniform sample of the entire dataset, using both an exact match criteria and approximate match criteria (Ippolito et al., 2022) using an edit distance of 10%.","reference_id":"MzQ0MzI5MjY5ODMzNDQ4OTI="} +{"index":4707,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2312.11671","reference_contexts":"\"Self-proliferation\" refers to the ability for an agent to autonomously replicate -to instantiate goal-directed agents on other machines, and to acquire resources such as compute necessary to keep them running (Kinniment et al., 2024)","reference_id":"MTMwNjE1OTY0NTQ5MzgzOTY4MTY="} +{"index":4708,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"1508.04025","reference_contexts":"We alternate between a local sliding window attention (Beltagy et al., 2020a,b ) and global attention (Luong et al., 2015) in every other layer.","reference_id":"NDkxOTE1NjY5NDQzMjk3NTM5NA=="} +{"index":4709,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2311.17035","reference_contexts":"Large language models may, under particular circumstances, be vulnerable to attacks causing the model to produce memorized 1 training data (Nasr et al., 2023)","reference_id":"NTk3MjU4NzE2NjIxODE3MzEyNQ=="} +{"index":4710,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"1910.10683","reference_contexts":"Large language models (LLMs) have demonstrated strong capabilities in language understanding, generation, and reasoning (Brown et al., 2020; Radford et al., 2019; Raffel et al., 2019)","reference_id":"MjUwODQ2MDY5MDM5ODU3OTU0Mg=="} +{"index":4711,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"1907.10641","reference_contexts":"Example domains include question answering (Clark et al., 2019; Kwiatkowski et al., 2019) , commonsense reasoning (Sakaguchi et al., 2019; Suzgun et al., 2022) , mathematics and science (Cobbe et al., 2021; Hendrycks et al., 2020) , and coding (Austin et al., 2021; Chen et al., 2021) While thorough testing of our models has been conducted, these tests cannot cover all applications and scenarios in which Gemma 2 may be used","reference_id":"MTc3Njg2NTcyNTA5MzQyNjY5MjM="} +{"index":4712,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2002.05202","reference_contexts":"length of 8192 tokens, the use of Rotary Position Embeddings (RoPE) (Su et al., 2021) , and the approximated GeGLU non-linearity (Shazeer, 2020)","reference_id":"MjQwOTM3NDczNjk4NDAxNTM3OQ=="} +{"index":4713,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2305.15324","reference_contexts":"We focus on capabilities relevant to extreme risks (Shevlane et al., 2023 ) (Phuong et al., 2024)","reference_id":"MzUwNjIzMzU2ODA2MTI1MTcyOA=="} +{"index":4714,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2104.09864","reference_contexts":"length of 8192 tokens, the use of Rotary Position Embeddings (RoPE) (Su et al., 2021) , and the approximated GeGLU non-linearity (Shazeer, 2020)","reference_id":"MjQwOTM3NDczNjk4NDAxNTM3OQ=="} +{"index":4715,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"1706.03762","reference_contexts":"Similar to previous Gemma models (Gemma Team, 2024) , the Gemma 2 models are based on a decoder-only transformer architecture (Vaswani et al., 2017)","reference_id":"NDI4MzEzNTgxNTk3Mzc3NDg3MQ=="} +{"index":4716,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2112.04359","reference_contexts":"Our approach and resulting impact assessment is reflective of that outlined for Gemma 1 (Gemma Team, 2024): we continue to believe that openness in AI can spread the benefits of these technologies across society, but must be evaluated against the risk of malicious uses, such as the creation of deepfake imagery, AI-generated disinformation or illegal and disturbing material, that can cause harm on both an individual and institutional levels (Weidinger et al., 2021)","reference_id":"MTc3MDcyNTgzMDEyNTI2MTc2MzU="} +{"index":4717,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2105.04663","reference_contexts":"As in Gemma 1, we use the GSPMD partitioner (Xu et al., 2021) for training step computation and the MegaScale XLA compiler (XLA, 2019).","reference_id":"OTI4MzAyMzA1MjU5ODI1MjY1MQ=="} +{"index":4718,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2306.14898","reference_contexts":"Specifically, we test on InterCode-CTF (Yang et al., 2023) , our own internal CTF suite foot_3 (Phuong et al., 2024) ; and a challenge based on Hack the Box foot_4 .","reference_id":"MTQ4MDA1ODQwMjMzMTc1MDAyMDM="} +{"index":4719,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"1910.07467","reference_contexts":"To stabilize training, we use RMSNorm (Zhang and Sennrich, 2019) to normalize the input and output of each transformer sub-layer, the attention layer, and the feedforward layer.","reference_id":"MTc4MDA5OTk0MDY4NTM5NTMzNDI="} +{"index":4720,"citer_arxiv_id":"2408.00118","cited_arxiv_id":"2309.11998","reference_contexts":"In particular, we use the prompts, but not the answers from LMSYS-chat-1M (Zheng et al., 2023)","reference_id":"NjIzMDA1ODQzNjE0NDA2OTE5OA=="} +{"index":4721,"citer_arxiv_id":"2303.16634","cited_arxiv_id":"2302.04166","reference_contexts":"Unlike GPTScore (Fu et al., 2023) which uses the conditional probability of generating the target text as an evaluation metric, G-EVAL directly performs the evaluation task with a form-filling paradigm","reference_id":"NjMwNTIzNTcxNDI4Mzk5ODc4NA=="} +{"index":4722,"citer_arxiv_id":"2303.16634","cited_arxiv_id":"2302.04166","reference_contexts":"GPTScore (Fu et al., 2023) is a new framework that evaluates texts with generative pre-training models like GPT-3","reference_id":"MjUyNjgzMzA5Mzc3ODIxOTU0Nw=="} +{"index":4723,"citer_arxiv_id":"2303.16634","cited_arxiv_id":"2303.04048","reference_contexts":"2023; Wang et al., 2023)","reference_id":"NTI2NDQzMjcxODcyNDY1Nzk3Ng=="} +{"index":4724,"citer_arxiv_id":"2303.16634","cited_arxiv_id":"2201.11903","reference_contexts":"Thus, there is a need for a more effective and reliable framework for using LLMs for NLG evaluation.In this paper, we propose G-EVAL, a framework of using LLMs with chain-of-thoughts (CoT) (Wei et al., 2022) to evaluate the quality of generated texts in a form-filling paradigm","reference_id":"NzY4NTA4NTk1MzQ1MDIyMzA4"} +{"index":4725,"citer_arxiv_id":"2303.16634","cited_arxiv_id":"1904.09675","reference_contexts":"BERTScore (Zhang et al., 2019) measures the similarity between two texts based on the contextualized embedding from BERT (Devlin et al., 2019)","reference_id":"NDczMDgxMTE4OTgzMjU5MjA4MQ=="} +{"index":4726,"citer_arxiv_id":"2303.16634","cited_arxiv_id":"1904.09675","reference_contexts":"BERTScore (Zhang et al., 2019) measures the similarity between two texts based on the contextualized embedding from BERT (Devlin et al., 2019) .","reference_id":"NzA2Mzc0NDM0NjIzMTkwNjkw"} +{"index":4727,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1312.5602","reference_contexts":"Games have served as stepping stones along this path for decades, from Backgammon (1992) to Chess (1997) to Atari (2013) [1] [2] [3]","reference_id":"MTgxOTg0MDg1OTA5NDYwMzk1MzM="} +{"index":4728,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1312.5602","reference_contexts":"[3] show that it is possible to combine a deep convolutional neural network with a Q-learning algorithm [40] and a novel experience replay approach to learn policies that can reach superhuman performance on the Atari ALE games","reference_id":"MzA2Njk2OTAzMjA2MTY1NjU4Mw=="} +{"index":4729,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1705.04304","reference_contexts":"In recent years, reinforcement learning (RL) models have tackled tasks as varied as robotic manipulation [5] , text summarization [6] , and video games such as Starcraft [7] and Minecraft [8] .","reference_id":"NDI1NTY5MDI1NDEzNTU3NjAzMg=="} +{"index":4730,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1707.06347","reference_contexts":"The policy is trained using Proximal Policy Optimization (PPO) [14] , a variant of advantage actor critic [15, 16]","reference_id":"NTk3MjY0ODA1MDE0MTIxNjg3OA=="} +{"index":4731,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1707.06347","reference_contexts":"Proximal Policy Optimization [14] and A3C [47] improve the ability to asynchronously collect rollout data","reference_id":"MTE5OTYzNjk0NjYxNTc0Nzk5NzA="} +{"index":4732,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1707.06347","reference_contexts":"Per [14] , we use entropy bonus to encourage exploration","reference_id":"MTA4NTgxMDc2MTA0MjI0NzU3Mzg="} +{"index":4733,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1506.02438","reference_contexts":"6 The optimization algorithm uses Generalized Advantage Estimation [17] (GAE), a standard advantage-based variance reduction technique [15] to stabilize and accelerate training","reference_id":"NTYyMDM5MzA0ODk5MTI0Nzc2Mw=="} +{"index":4734,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1506.02438","reference_contexts":"Here γ is the discount factor [17] and T is the real game time corresponding to each step (0.133 seconds)","reference_id":"MTExOTA5Nzc1MjQyMzAzMjk2NjE="} +{"index":4735,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1412.6980","reference_contexts":"We apply the Adam optimizer [20] using truncated backpropagation through time [21] over samples of 16 timesteps","reference_id":"MTAwNDQ1NTEzNTExNzEyMzQyODM="} +{"index":4736,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1511.05641","reference_contexts":"This case is a special case of Net2Net-style function preserving transformations [23]","reference_id":"MTQ5MjkyODI4MjQ4NTIyNzc0NDY="} +{"index":4737,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1511.05641","reference_contexts":"The motivation for our surgery method is similar to prior work on Net2Net style function preserving transformations [23] which attempt to add model capacity without compromising performance, whereas our surgery technique was used in cases where the inputs, outputs, and recurrent layer size changed","reference_id":"MTMyODk0NjE1MzE3Mjc2ODA5MzU="} +{"index":4738,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1910.07113","reference_contexts":"Further work such as [24] has already begun to explore alternate methods of surgery which avoid this constraint.","reference_id":"MzM1OTYxNTMyODIxNzY3OTU0MQ=="} +{"index":4739,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1910.07113","reference_contexts":"Simultaneous works in recent months [7, 24] have matched or slightly exceeded our scale.","reference_id":"MTE0MDEyNTI2ODUwMjk1NDMyMzk="} +{"index":4740,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1910.07113","reference_contexts":"[24] has reported success using behavioral cloning for similar purposes.","reference_id":"OTc0MjE4NTcyODgwOTE1NjAyOQ=="} +{"index":4741,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1812.06162","reference_contexts":"This renewed interest in reducing convergence-time and treating batch-size as a key design parameter also motivated the work of [28] , where an analytical tool is derived to estimate a training-time optimal batch size on per task basis by studying the \"noise scale\" of the gradients","reference_id":"MTgwMTQzNjQ4ODI5NjA2MjE5ODM="} +{"index":4742,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1812.06162","reference_contexts":"Per results in [28] , we hoped to find (in the early part of training) linear speedup from increasing batch size; i.e","reference_id":"MTY2NDg0MTUyNTQ4ODM1OTc5MzM="} +{"index":4743,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1812.06162","reference_contexts":"If this scaling property holds, it is possible to use the same total amount of GPU-days (and thus dollars) to reach a given result [28]","reference_id":"MTQ3MDEzMjI4MDIxMjAyNzgxODk="} +{"index":4744,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1710.03748","reference_contexts":"Self-play was shown to be a powerful algorithm for learning skills within high-dimensional continuous environments [33] and a method for automatically generating curricula [34]","reference_id":"MTM5MjI5NDQ4ODczNDA2NjA5MTM="} +{"index":4745,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1703.05407","reference_contexts":"Self-play was shown to be a powerful algorithm for learning skills within high-dimensional continuous environments [33] and a method for automatically generating curricula [34]","reference_id":"MTM5MjI5NDQ4ODczNDA2NjA5MTM="} +{"index":4746,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1810.12894","reference_contexts":"Following this work, a variety of efforts have pushed performance on the remaining Atari games [16] , reduced the sample complexity, and introduced new challenges by focusing on intrinsic rewards [41] [42] [43] .","reference_id":"MTYzNzAyMTgwMzM1MjI0MTM="} +{"index":4747,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1802.01561","reference_contexts":"Recent work has demonstrated the benefit of distributed learning on a wide array of problems including single-player video games [48] and robotics [5] .","reference_id":"NzM4MjM5ODEwNzM2NDA2ODY3NA=="} +{"index":4748,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1606.04671","reference_contexts":"[53] use similar methods to use a trained model to quickly learn novel tasks","reference_id":"MTYzNzYxNDAxODUzMjcxMTMwMzQ="} +{"index":4749,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1503.02531","reference_contexts":"Distillation [54] and imitation learning [55, 56] offer an alternate approach to surgery for making model changes in response to a shifting environment","reference_id":"OTU0NzE4ODY4MzYzNjMyMjE2MQ=="} +{"index":4750,"citer_arxiv_id":"1912.06680","cited_arxiv_id":"1606.01540","reference_contexts":"We designed our Dota 2 environment to behave like a standard OpenAI Gym environment [59]","reference_id":"ODk4NTk2MjM0NDQ2NDk0OTgxOQ=="} +{"index":4751,"citer_arxiv_id":"1606.04671","cited_arxiv_id":"1503.02531","reference_contexts":"While distillation [8] offers one potential solution to multitask learning [17] , it requires a reservoir of persistent training data for all tasks, an assumption which may not always hold","reference_id":"MTUyNDc0NDMxNzU4NjM5ODIyMzU="} +{"index":4752,"citer_arxiv_id":"1606.04671","cited_arxiv_id":"1511.06295","reference_contexts":"Recently, there have been new methods proposed for multi-task or transfer learning with deep RL: [22, 17, 14]","reference_id":"MTE0OTUyMDQzODgyNTAxODkzNA=="} +{"index":4753,"citer_arxiv_id":"1606.04671","cited_arxiv_id":"1511.06295","reference_contexts":"by adding fewer layers or less capacity, by pruning [9] , or by online compression [17] during learning","reference_id":"MjA5MTA3MzYyNjUzMTI5NTE0MA=="} +{"index":4754,"citer_arxiv_id":"1606.04671","cited_arxiv_id":"1511.06295","reference_contexts":"While distillation [8] offers one potential solution to multitask learning [17] , it requires a reservoir of persistent training data for all tasks, an assumption which may not always hold","reference_id":"MTUyNDc0NDMxNzU4NjM5ODIyMzU="} +{"index":4755,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2204.05862","reference_contexts":"For example, according to a holistic study [Liang et al., 2022] , closed-source models (e.g., InstructGPT [Ouyang et al., 2022] and Anthropic-LM [Bai et al., 2022] ) usually outperforms open-source models (e.g., GPT-J [Wang and Komatsuzaki, 2021] and OPT [Zhang et al., 2022a] ), posting a challenge to open-source research.","reference_id":"NDI5ODcyODgzMTMzMzAzMzgyOA=="} +{"index":4756,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2212.06817","reference_contexts":"This is due to the heavy reliance on human-robot interactions, as highlighted by studies such as [Shah et al., 2023 , Brohan et al., 2022 , Fang et al., 2020]","reference_id":"NDAwNTAxOTk5ODI2NDc2NzY4MQ=="} +{"index":4757,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2204.02311","reference_contexts":"LLMs such as GPT-3 [Brown et al., 2020] , Bloom [Scao et al., 2022] , PaLM [Chowdhery et al., 2022] , Megatron-Turing-530B [Smith et al., 2022] , Chinchilla [Hoffmann et al., 2022] , and others have expanded the horizons of language understanding and generation","reference_id":"NzczNjQ4NzAyMTM4NDUwOTk2MQ=="} +{"index":4758,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2204.02311","reference_contexts":"Models like GPT-3 [Brown et al., 2020] , Bloom [Scao et al., 2022] , PaLM [Chowdhery et al., 2022] , megatron-turing-530b [Smith et al., 2022] , Chinchilla [Hoffmann et al., 2022] and so on have pushed the capabilities of language understanding and generation to new frontiers","reference_id":"MTMxNTMzMTkzNDE0OTM5ODUwMjU="} +{"index":4759,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2104.13921","reference_contexts":"Recently, open-vocabulary object detection has attracted a lot of attention [Gu et al., 2021 , Li et al., 2022 , Liu et al., 2023b , Yao et al., 2022]","reference_id":"NDM3NzAwMzQ5MDA2NzM3NTUwNQ=="} +{"index":4760,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2104.13921","reference_contexts":"Recently, a series of open-vocabulary detection systems have emerged as the new trend [Gu et al., 2021 , Li et al., 2022 , Yao et al., 2022 , Liu et al., 2023b]","reference_id":"OTM2NDU5NDA3NzQ4Njc0NzM1MA=="} +{"index":4761,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2203.15556","reference_contexts":"LLMs such as GPT-3 [Brown et al., 2020] , Bloom [Scao et al., 2022] , PaLM [Chowdhery et al., 2022] , Megatron-Turing-530B [Smith et al., 2022] , Chinchilla [Hoffmann et al., 2022] , and others have expanded the horizons of language understanding and generation","reference_id":"NzczNjQ4NzAyMTM4NDUwOTk2MQ=="} +{"index":4762,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2203.15556","reference_contexts":"Models like GPT-3 [Brown et al., 2020] , Bloom [Scao et al., 2022] , PaLM [Chowdhery et al., 2022] , megatron-turing-530b [Smith et al., 2022] , Chinchilla [Hoffmann et al., 2022] and so on have pushed the capabilities of language understanding and generation to new frontiers","reference_id":"MTMxNTMzMTkzNDE0OTM5ODUwMjU="} +{"index":4763,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2211.09110","reference_contexts":"For example, according to a holistic study [Liang et al., 2022] , closed-source models (e.g., InstructGPT [Ouyang et al., 2022] and Anthropic-LM [Bai et al., 2022] ) usually outperforms open-source models (e.g., GPT-J [Wang and Komatsuzaki, 2021] and OPT [Zhang et al., 2022a] ), posting a challenge to open-source research.","reference_id":"NDI5ODcyODgzMTMzMzAzMzgyOA=="} +{"index":4764,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2211.09110","reference_contexts":"However, the most powerful large models are closed-source [Liang et al., 2022] , limiting their accessibility and hindering progress in the field","reference_id":"NDc3MzE3NzU3MjQ1MTExNDY1MQ=="} +{"index":4765,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2211.05100","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4766,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2201.11990","reference_contexts":"LLMs such as GPT-3 [Brown et al., 2020] , Bloom [Scao et al., 2022] , PaLM [Chowdhery et al., 2022] , Megatron-Turing-530B [Smith et al., 2022] , Chinchilla [Hoffmann et al., 2022] , and others have expanded the horizons of language understanding and generation","reference_id":"NzczNjQ4NzAyMTM4NDUwOTk2MQ=="} +{"index":4767,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2201.11990","reference_contexts":"Models like GPT-3 [Brown et al., 2020] , Bloom [Scao et al., 2022] , PaLM [Chowdhery et al., 2022] , megatron-turing-530b [Smith et al., 2022] , Chinchilla [Hoffmann et al., 2022] and so on have pushed the capabilities of language understanding and generation to new frontiers","reference_id":"MTMxNTMzMTkzNDE0OTM5ODUwMjU="} +{"index":4768,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2302.13971","reference_contexts":"Recently, Meta's LLaMA [Touvron et al., 2023] provides a series of powerful open-source models that boost language model research","reference_id":"NDI3NTc3MTI3NDgzNzQ2NzcyNg=="} +{"index":4769,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2302.13971","reference_contexts":"In response, Meta's LLaMA [Touvron et al., 2023] offers a suite of powerful open-source models that bolster language model research","reference_id":"MTgzMTMxMjk5NzMyODE3ODkzNjA="} +{"index":4770,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2202.03052","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4771,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2203.11171","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4772,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2201.11903","reference_contexts":"These neural networks, with hundreds of billions of parameters, exhibit human-like proficiency in complex reasoning [Wei et al., 2022 , Wang et al., 2022b , Zhou et al., 2022 , Zhang et al., 2022b , Diao et al., 2023b , Shum et al., 2023]","reference_id":"MTQ4Nzk3MzcxNDc1NzExODY4OTM="} +{"index":4773,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2205.01068","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4774,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2210.03493","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4775,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2205.10625","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4776,"citer_arxiv_id":"2305.14167","cited_arxiv_id":"2010.04159","reference_contexts":"Traditional object detectors have a fixed number of classification heads, which restricts their practicality and makes them only capable of predicting the classes on which they are trained [Girshick, 2015 , Ren et al., 2015 , Lin et al., 2017 , Yao et al., 2021a , Duan et al., 2019 , Yao et al., 2021b , Zhu et al., 2020 , Carion et al., 2020]","reference_id":"MTAwNTM5NzcxOTQ0OTE1NTE0OTY="} +{"index":4777,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2305.06500","reference_contexts":"(1) LVLMs designed for dealing with a single image, BLIP2-OPT-6.7B (Li et al., 2023b) , InstructBLIP-Vicuna7B (Dai et al., 2023) , LLaVA-1.5-7B/13B (Liu et al., 2023a)","reference_id":"MTE0MTMxOTY5OTI3NDA4NDk1NjI="} +{"index":4778,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2305.06500","reference_contexts":"Innovations such as the Q-Former module in BLIP2 (Li et al., 2023b) and instruction integration in InstructBLIP (Dai et al., 2023) further enhance alignment capabilities","reference_id":"NTMxMTY3MTI5NjUxMzI2MTg5NQ=="} +{"index":4779,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2305.06500","reference_contexts":"3 InstructBLIP (Dai et al., 2023) employs an instruction-aware visual feature extraction module based on BLIP2 (Li et al., 2023b) and is trained with unified multimodal instruction tuning datasets.","reference_id":"MTI3MzYzMjYxOTExNzQzMzk2OTA="} +{"index":4780,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2306.13394","reference_contexts":"Large vision-language models (LVLMs), which integrate large language models (LLMs) (Brown et al., 2020a; Touvron et al., 2023) with pre-trained vision encoders through cross-modal alignment training (Madureira, 2021; Liu et al., 2023b; Li et al., 2023d) , have demonstrated remarkable perceptual and cognitive capabilities in processing concrete images from everyday scenes (OpenAI, 2023; Fu et al., 2023; Yang et al., 2023a; Reka, 2024)","reference_id":"NDgyMTY2NTkzNTUzODQ3ODk4"} +{"index":4781,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2306.13394","reference_contexts":"Contextualized Captioning Inspired by the evolving in-context learning capabilities of LLMs (Brown et al., 2020b; Dong et al., 2022) , we introduce a contextualized captioning task to examine the in-context learning ability of LVLMs","reference_id":"NDM0MTUzMTIxMjg2NzQ1NzU4"} +{"index":4782,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2312.11805","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4783,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"1710.07300","reference_contexts":"In the scientific domain, Fig-CAP (Chen et al., 2019) and FigureQA (Kahou et al., 2017) are created based on synthetic figures","reference_id":"MTgxMTUwNDg4OTkxNTA3MDA1Mjk="} +{"index":4784,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2306.00890","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4785,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2301.12597","reference_contexts":"(1) LVLMs designed for dealing with a single image, BLIP2-OPT-6.7B (Li et al., 2023b) , InstructBLIP-Vicuna7B (Dai et al., 2023) , LLaVA-1.5-7B/13B (Liu et al., 2023a)","reference_id":"MTE0MTMxOTY5OTI3NDA4NDk1NjI="} +{"index":4786,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2301.12597","reference_contexts":"Innovations such as the Q-Former module in BLIP2 (Li et al., 2023b) and instruction integration in InstructBLIP (Dai et al., 2023) further enhance alignment capabilities","reference_id":"NTMxMTY3MTI5NjUxMzI2MTg5NQ=="} +{"index":4787,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2301.12597","reference_contexts":"B.1 Details of Evaluated Models BLIP2 (Li et al., 2023b) , introduces a lightweight Q-Former designed to bridge modality gaps and leverages frozen LLMs","reference_id":"ODQ3OTY5NzAwMTE4NTUwMTY1Ng=="} +{"index":4788,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2306.04387","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4789,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2304.08485","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4790,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2310.02255","reference_contexts":"However, recent studies have shown that open-source LVLMs struggle to understand abstract figures, such as geometric shapes in multimodal mathematical reasoning (Lu et al., 2023; Zhang et al., 2024b) and scientific plots (Yue et al., 2023)","reference_id":"NTA4MTAwOTM2OTUxMDE2Njc1NQ=="} +{"index":4791,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2310.02255","reference_contexts":"Our experiments demonstrate that ArXivQA brings a significant 10.4% absolute accuracy boost for Qwen-VL-Chat (Bai et al., 2023b) , on the Math-Vista (Lu et al., 2023) , a challenging benchmark for multimodal mathematical reasoning","reference_id":"NzczNDI4MTI5OTk3MDQ5MDYy"} +{"index":4792,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2310.02255","reference_contexts":"We find that not only the scientific reasoning performance is improved on multimodal reasoning tasks (MathVista (Lu et al., 2023) Then ∠CAB = 90°-∠ACB = 90°-45°= 45°","reference_id":"OTY3ODY1NjI1MDE3MjM0MjUzNQ=="} +{"index":4793,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2403.05530","reference_contexts":"Gemini 1.0 Pro Vision (Reid et al., 2024) , a upgraded LVLM by Google","reference_id":"NTUxMzA2ODQ0NDc2MjQ3Mjc4MA=="} +{"index":4794,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2111.02114","reference_contexts":"Modality alignment training often utilizes web-scale image-caption pairs such as Laion-400M (Schuhmann et al., 2021) , with recent studies favoring cleaned captions (Chen et al., 2023a; Yu et al., 2023)","reference_id":"OTgzNDUwOTAyNDU0MzQxNDUwNA=="} +{"index":4795,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2309.14525","reference_contexts":"Alignment methodologies such as RLHF training (Ouyang et al., 2022) , e.g., LLaVA-RLHF (Sun et al., 2023) , and preference optimization through AI feedback (Li et al., 2023c) demonstrate effectiveness in aligning LVLMs with human preferences.","reference_id":"NjE0MjY2MjUwODg1NzQ1NTcyNw=="} +{"index":4796,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2302.13971","reference_contexts":"Large vision-language models (LVLMs), which integrate large language models (LLMs) (Brown et al., 2020a; Touvron et al., 2023) with pre-trained vision encoders through cross-modal alignment training (Madureira, 2021; Liu et al., 2023b; Li et al., 2023d) , have demonstrated remarkable perceptual and cognitive capabilities in processing concrete images from everyday scenes (OpenAI, 2023; Fu et al., 2023; Yang et al., 2023a; Reka, 2024)","reference_id":"NDgyMTY2NTkzNTUzODQ3ODk4"} +{"index":4797,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2302.13971","reference_contexts":"CLIP (Radford et al., 2021) is widely used for image encoding, while LLaMA (Touvron et al., 2023) and Vicuna (Chiang et al., 2023) serve as popular choices for LLMs","reference_id":"MTE2MzY3NTU0Nzc0NjY0ODg4NTI="} +{"index":4798,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2305.17926","reference_contexts":"Specifically, we employ GPT-4 for the evaluation of single-figure caption tasks following FairEval (Wang et al., 2023)","reference_id":"MzU3MTQxNDQyODI2MTE5Nzg1Mw=="} +{"index":4799,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2309.17421","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4800,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2310.20550","reference_contexts":"Modality alignment training often utilizes web-scale image-caption pairs such as Laion-400M (Schuhmann et al., 2021) , with recent studies favoring cleaned captions (Chen et al., 2023a; Yu et al., 2023)","reference_id":"OTgzNDUwOTAyNDU0MzQxNDUwNA=="} +{"index":4801,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2311.16502","reference_contexts":"However, recent studies have shown that open-source LVLMs struggle to understand abstract figures, such as geometric shapes in multimodal mathematical reasoning (Lu et al., 2023; Zhang et al., 2024b) and scientific plots (Yue et al., 2023)","reference_id":"NTA4MTAwOTM2OTUxMDE2Njc1NQ=="} +{"index":4802,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2403.14624","reference_contexts":"However, recent studies have shown that open-source LVLMs struggle to understand abstract figures, such as geometric shapes in multimodal mathematical reasoning (Lu et al., 2023; Zhang et al., 2024b) and scientific plots (Yue et al., 2023)","reference_id":"NTA4MTAwOTM2OTUxMDE2Njc1NQ=="} +{"index":4803,"citer_arxiv_id":"2403.00231","cited_arxiv_id":"2304.10592","reference_contexts":"The alignment module varies from simple linear projections (Liu et al., 2023b; Zhu et al., 2023) to more complex architectures like gated cross-attention layers substantiated by Flamingo and IDEFICS (Alayrac et al., 2022; Awadalla et al., 2023)","reference_id":"NDg3NDY5NzYxNzIxNzczNzYwNw=="} +{"index":4804,"citer_arxiv_id":"1604.01696","cited_arxiv_id":"1502.05698","reference_contexts":"There are a few recent frameworks for evaluating language comprehension (Hermann et al., 2015; Weston et al., 2015) , including the MCTest (Richardson et al., 2013) as a notable one","reference_id":"MzgyOTY1MDk2MzI1MzIzNzc0MQ=="} +{"index":4805,"citer_arxiv_id":"1604.01696","cited_arxiv_id":"1506.06724","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4806,"citer_arxiv_id":"2304.06790","cited_arxiv_id":"2304.02643","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4807,"citer_arxiv_id":"1803.11175","cited_arxiv_id":"1705.02364","reference_contexts":"However, recent work has demonstrated strong transfer task performance using pre-trained sentence level embeddings (Conneau et al., 2017) .","reference_id":"ODYzNDE5NTQyNTUyODUxOTMyMg=="} +{"index":4808,"citer_arxiv_id":"1803.11175","cited_arxiv_id":"1705.00652","reference_contexts":"The supported tasks include: a Skip-Thought like task (Kiros et al., 2015) for the unsupervised learning from arbitrary running text; a conversational input-response task for the inclusion of parsed conversational data (Henderson et al., 2017) ; and classification tasks for training on supervised data","reference_id":"MTI1MjI5MDgxMjY2NjM4MjI1ODU="} +{"index":4809,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"1312.6114","reference_contexts":"In the default formulation of the Variational Autoencoder (VAE) [1] , we directly maximize the ELBO","reference_id":"NjQ1ODE4NTQzNTA3OTEwMTI1Ng=="} +{"index":4810,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2205.11487","reference_contexts":"This forms the backbone of image super-resolution models such as Cascaded Diffusion Models [18] , as well as state-of-the-art image-text models such as DALL-E 2 [19] and Imagen [7] .","reference_id":"MTczOTQzMzEzOTU4OTgzMjI4ODg="} +{"index":4811,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2205.11487","reference_contexts":"We have therefore shown that learning a VDM by predicting the original image x 0 is equivalent to learning to predict the noise; empirically, however, some works have found that predicting the noise resulted in better performance [5, 7] .","reference_id":"NjE1OTE3NTcwMjgwNjg2NTAxMw=="} +{"index":4812,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2011.13456","reference_contexts":"Collectively, learning to represent a distribution as a score function and using it to generate samples through Markov Chain Monte Carlo techniques, such as Langevin dynamics, is known as Score-based Generative Modeling [9, 10, 11] .","reference_id":"MTU1NDUwNzcwMDI5MDU1MTk1MDQ="} +{"index":4813,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2011.13456","reference_contexts":"Therefore, in Classifier Guidance [10, 20] , the score of an unconditional diffusion model is learned as previously derived, alongside a classifier that takes in arbitrary noisy x t and attempts to predict conditional information y","reference_id":"OTAyMTkyNTEwMDAwOTM2NjE1MA=="} +{"index":4814,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2011.13456","reference_contexts":"Sampling is then performed by reversing the SDE, which naturally requires estimating the score function at each continuous-valued noise level [10]","reference_id":"MTEwMDM1NzIzNjI2ODcyMDA4MDc="} +{"index":4815,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2011.13456","reference_contexts":"Fortunately, we can look to another class of generative models, Score-based Generative Models [9, 10, 11] , for exactly this intuition","reference_id":"MTA1NjQ2MzcxNzQ0NjQ5MDYyNDk="} +{"index":4816,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2011.13456","reference_contexts":"The two most popular forms of guidance are known as Classifier Guidance [10, 20] and Classifier-Free Guidance [21] .","reference_id":"MTE0Mzk1NzQ0MDU2MDc0NzAyNzU="} +{"index":4817,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2101.03288","reference_contexts":"To begin to understand why optimizing a score function makes sense, we take a detour and revisit energybased models [12, 13]","reference_id":"NjU4OTUwNjYwNDg4NTA3MzExNg=="} +{"index":4818,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"1805.08306","reference_contexts":"Fortunately, alternative techniques known as score matching [14, 15, 16, 17] have been derived to minimize this Fisher divergence without knowing the ground truth score, and can be optimized with stochastic gradient descent.","reference_id":"ODg4NTgyODQwMTExMTI3MzM3OA=="} +{"index":4819,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2204.06125","reference_contexts":"prior matching term (Definition of KL Divergence) (19) In this case, we learn an intermediate bottlenecking distribution q φ (z|x) that can be treated as an encoder ; it transforms inputs into a distribution over possible latents","reference_id":"MTQxMzY3MjcxNDM0NTA1MjE0NTU="} +{"index":4820,"citer_arxiv_id":"2208.11970","cited_arxiv_id":"2204.06125","reference_contexts":"This forms the backbone of image super-resolution models such as Cascaded Diffusion Models [18] , as well as state-of-the-art image-text models such as DALL-E 2 [19] and Imagen [7] .","reference_id":"MTczOTQzMzEzOTU4OTgzMjI4ODg="} +{"index":4821,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2308.01390","reference_contexts":"Human A blue cartoon A white fish with orange/red spots InstructBLIP [2] There is a cute blue and white pony in a still, looking cheerful and cheerful","reference_id":"MTQ3NjQ5MDc4MDA1NTQ5Mzk2NjU="} +{"index":4822,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2308.01390","reference_contexts":"This is an object of InstructBLIP [2] sofa person cutting meat into small pieces with a large knife LLaVA [12] The image is a black and white drawing of a couch","reference_id":"MTA2NzA4MTM3NTU5NjIzMjQ4NzM="} +{"index":4823,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2308.01390","reference_contexts":"The first category includes models that employ a large language model to interface with individual, modality-specific models or APIs [16, 22, 38, 46, 52]","reference_id":"NDczMzg0NzY4OTM5MTQxNTQ4OA=="} +{"index":4824,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2308.01390","reference_contexts":"In the meantime, we find that including category names in the prompt results in meaningless responses from InstructBLIP [2] , which is the model we compare with, making meaningful comparisons challenging","reference_id":"MzMwMTIyMjI1NTIwMjA3NTA4NA=="} +{"index":4825,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2308.01390","reference_contexts":"Prompt What is this? This is an object of InstructBLIP [2] table the foot LLaVA [12] The image features a black and white, close-up view of a grater, which is an object used for grating cheese or other food items.","reference_id":"MTQ2MDU3NjY3NTk0MzQ2ODgyMDU="} +{"index":4826,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2308.01390","reference_contexts":"Ground Truth Guitar Bottle Prompt What is this? This is an object of InstructBLIP [2] telescope wine bottle LLaVA [12] The image is a close-up view of a black ink pen sitting on a white surface.","reference_id":"MTYwODE2NDk1NjA1NzI2MTQ3MTA="} +{"index":4827,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2308.01390","reference_contexts":"A cartoon green and red like a fruit A black insect InstructBLIP [2] an appleavatar 3d model 26 LLaVA [12] This is a computer-generated image (CGI) or a 3D model of an apple with a leaf on it.","reference_id":"MTQwNTAwNzYzMTkyNjMyNTM4NzE="} +{"index":4828,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2204.02311","reference_contexts":"For training Point-BERT, we employ ViT-L/14 trained on DataComp-1B [6] for 12.8B steps and batch size 90k (denoted as 'ViT-L/14datacomp xl s13b b90k') from OpenCLIP [10]","reference_id":"MTQ5NzgwMzExNDI2NjEzMjUxMDM="} +{"index":4829,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2303.03378","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4830,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.15010","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4831,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2104.08821","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4832,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2305.04790","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4833,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2309.00615","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4834,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2309.03905","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4835,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2206.06336","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4836,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"1606.08415","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4837,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.12995","reference_contexts":"The first category includes models that employ a large language model to interface with individual, modality-specific models or APIs [16, 22, 38, 46, 52]","reference_id":"NDczMzg0NzY4OTM5MTQxNTQ4OA=="} +{"index":4838,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.12995","reference_contexts":"Multi-modal Large Language Models (MLLMs) are designed to comprehend and interpret a wide range of information that extends beyond mere text-based data [56] , including but not limited to images [14, 23, 31, 51, 61] , audio [22] , motion [25] , etc","reference_id":"MjI2NDA3MzExMzcwOTQ0NDI5OA=="} +{"index":4839,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.12995","reference_contexts":"We use ULIP-2 [21] to pre-train our point cloud encoder (Point-BERT [22] )","reference_id":"OTQxNTY4NjM4MzI1MzQ4ODgxNQ=="} +{"index":4840,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.12995","reference_contexts":"Beyond text, the exploration of multi-modal LLMs (MLLMs) now extends to processing audio [22] , images [1, 23, 28, 31, 35, 59 , 61], and more.","reference_id":"NTE4NDUyODg1NDk3Mzc0MTMxNg=="} +{"index":4841,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2302.14045","reference_contexts":"Recently, connecting pre-trained encoders with LLMs using projection layers and employing a two-stage training of alignment and instruction tuning has been proven effective for developing MLLMs across various domains [8, 10, 23, 25, 31, 61]","reference_id":"MTE5MTkxNzkxNzAwNDM2Njc3NTE="} +{"index":4842,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2302.14045","reference_contexts":"Multi-modal Large Language Models (MLLMs) are designed to comprehend and interpret a wide range of information that extends beyond mere text-based data [56] , including but not limited to images [14, 23, 31, 51, 61] , audio [22] , motion [25] , etc","reference_id":"MjI2NDA3MzExMzcwOTQ0NDI5OA=="} +{"index":4843,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2302.14045","reference_contexts":"Beyond text, the exploration of multi-modal LLMs (MLLMs) now extends to processing audio [22] , images [1, 23, 28, 31, 35, 59 , 61], and more.","reference_id":"NTE4NDUyODg1NDk3Mzc0MTMxNg=="} +{"index":4844,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2302.14045","reference_contexts":"The first involves training the model from scratch, similar to text-only LLMs, using large-scale multi-modal corpora and datasets [23, 39]","reference_id":"ODQ1OTE3ODk4MTQ3NTUyNzI3MA=="} +{"index":4845,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2306.05425","reference_contexts":"[1, 2, 8, 10, 11, 14, 27, 28, 31, 44, [59] [60] [61]","reference_id":"ODQ0MjMyMzAzNDg0OTYxMjkwMw=="} +{"index":4846,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2303.08774","reference_contexts":"To overcome this, we follow [31] and propose an automated data generation technique utilizing the large-scale point cloud captioning dataset, Cap3D [33] , with the assistance of GPT-4 [35]","reference_id":"OTcyNTkzNzE5MTYxMTY5NDY2"} +{"index":4847,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2303.08774","reference_contexts":"Beyond text, the exploration of multi-modal LLMs (MLLMs) now extends to processing audio [22] , images [1, 23, 28, 31, 35, 59 , 61], and more.","reference_id":"NTE4NDUyODg1NDk3Mzc0MTMxNg=="} +{"index":4848,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2305.15334","reference_contexts":"The first category includes models that employ a large language model to interface with individual, modality-specific models or APIs [16, 22, 38, 46, 52]","reference_id":"NDczMzg0NzY4OTM5MTQxNTQ4OA=="} +{"index":4849,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2306.14824","reference_contexts":"The first involves training the model from scratch, similar to text-only LLMs, using large-scale multi-modal corpora and datasets [23, 39]","reference_id":"ODQ1OTE3ODk4MTQ3NTUyNzI3MA=="} +{"index":4850,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"1908.10084","reference_contexts":"Therefore, we incorporate and rely more on two additional data-driven metrics, Sentence-BERT [43] and SimCSE [12] similarity, which compute the similarity of sentence embeddings between model-generated and human captions.","reference_id":"Nzc1MTAyNzUwOTI1NjYxMjIxMg=="} +{"index":4851,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2305.16355","reference_contexts":"[1, 2, 8, 10, 11, 14, 27, 28, 31, 44, [59] [60] [61]","reference_id":"ODQ0MjMyMzAzNDg0OTYxMjkwMw=="} +{"index":4852,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2303.08128","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4853,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2302.13971","reference_contexts":"We use the LLaMA model [48] as our LLM backbone, with the 7B and 13B Vicuna [5] checkpoint","reference_id":"MTI5MTIwNjI4NTM3MzgzNDk5OTQ="} +{"index":4854,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2305.11175","reference_contexts":"Multi-modal Large Language Models (MLLMs) are designed to comprehend and interpret a wide range of information that extends beyond mere text-based data [56] , including but not limited to images [14, 23, 31, 51, 61] , audio [22] , motion [25] , etc","reference_id":"MjI2NDA3MzExMzcwOTQ0NDI5OA=="} +{"index":4855,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2303.04671","reference_contexts":"The first category includes models that employ a large language model to interface with individual, modality-specific models or APIs [16, 22, 38, 46, 52]","reference_id":"NDczMzg0NzY4OTM5MTQxNTQ4OA=="} +{"index":4856,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2306.13549","reference_contexts":"Multi-modal Large Language Models (MLLMs) are designed to comprehend and interpret a wide range of information that extends beyond mere text-based data [56] , including but not limited to images [14, 23, 31, 51, 61] , audio [22] , motion [25] , etc","reference_id":"MjI2NDA3MzExMzcwOTQ0NDI5OA=="} +{"index":4857,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2303.16199","reference_contexts":"[1, 2, 8, 10, 11, 14, 27, 28, 31, 44, [59] [60] [61]","reference_id":"ODQ0MjMyMzAzNDg0OTYxMjkwMw=="} +{"index":4858,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2303.16199","reference_contexts":"Beyond text, the exploration of multi-modal LLMs (MLLMs) now extends to processing audio [22] , images [1, 23, 28, 31, 35, 59 , 61], and more.","reference_id":"NTE4NDUyODg1NDk3Mzc0MTMxNg=="} +{"index":4859,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2307.03601","reference_contexts":"[1, 2, 8, 10, 11, 14, 27, 28, 31, 44, [59] [60] [61]","reference_id":"ODQ0MjMyMzAzNDg0OTYxMjkwMw=="} +{"index":4860,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.10592","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4861,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2305.06500","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4862,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.14108","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4863,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2309.00615","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4864,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2305.02463","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4865,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2304.08485","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4866,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2212.08751","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4867,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2212.10560","reference_contexts":"acting as generalized interfaces [18] for a broad range of tasks [4, 42]","reference_id":"OTMyOTM0NTMxODQ1NTk4NzA4Mg=="} +{"index":4868,"citer_arxiv_id":"2308.16911","cited_arxiv_id":"2212.10560","reference_contexts":"2 , we present word clouds (after removing generic words like \"model\", \"object\", etc.) and verb-noun pair distributions following [18] from our instruction following dataset, highlighting its extensive coverage of diverse topics such as color, shape, usage, material, and more","reference_id":"MTQ5MjMwNDI3NzAzMzk1OTA0OTQ="} +{"index":4869,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"1810.04805","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4870,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2009.03300","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4871,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2205.11916","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4872,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2001.08361","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4873,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2210.14306","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4874,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2112.09332","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4875,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2212.13138","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4876,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2203.11171","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4877,"citer_arxiv_id":"2303.13375","cited_arxiv_id":"2201.11903","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4878,"citer_arxiv_id":"2203.09509","cited_arxiv_id":"2009.06367","reference_contexts":"This is similar in spirit to previous work on using cooperative discriminators on uncontrolled LLMs (Holtzman et al., 2018; Krause et al., 2020; Yang and Klein, 2021; Liu et al., 2021a), yet in this work our LLM is controlled in an adversarial way by prompting and by a classifier.","reference_id":"MTc3NTQxMTI4MDc5MDY3Mzg4ODE="} +{"index":4879,"citer_arxiv_id":"2203.09509","cited_arxiv_id":"2101.06804","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4880,"citer_arxiv_id":"2203.09509","cited_arxiv_id":"2107.13586","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4881,"citer_arxiv_id":"2203.09509","cited_arxiv_id":"2104.08773","reference_contexts":"While there are several approaches for prompting pretrained LLMs (Liu et al., 2021b) , a recent and promising direction is demonstration-based prompting (Gao et al., 2021; Mishra et al., 2021) .","reference_id":"MTU5MzkyNDEyNzg5MzcwNTY2MQ=="} +{"index":4882,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2008.03156","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4883,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2012.13255","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4884,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2004.05150","reference_contexts":"Coupling this with the challenges of training transformers on very long sequence lengths (Choromanski et al., 2020; Wang et al., 2020; Beltagy et al., 2020) , it was important to automatically convert web pages to a simplified form, which we call Minimal-HTML (MHTML), as defined below.","reference_id":"MjU3MDUwNjYwMzExOTQ3MzM0NA=="} +{"index":4885,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2005.14165","reference_contexts":"The vast majority of text used to pretrain language models is extracted from web pages, while discarding any markup they contain (Liu et al., 2019; Brown et al., 2020; Raffel et al., 2019; Lewis et al., 2019)","reference_id":"MTUwNDI1NzMyNzUxOTc5MjM2NzU="} +{"index":4886,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2009.14794","reference_contexts":"Coupling this with the challenges of training transformers on very long sequence lengths (Choromanski et al., 2020; Wang et al., 2020; Beltagy et al., 2020) , it was important to automatically convert web pages to a simplified form, which we call Minimal-HTML (MHTML), as defined below.","reference_id":"MjU3MDUwNjYwMzExOTQ3MzM0NA=="} +{"index":4887,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1911.02116","reference_contexts":"One by-product of this type of filtering is that it also produced high-quality documents by default foot_2 ; thus, we opted out of model-based filtering of documents such as CC-100 (Conneau et al., 2019)","reference_id":"NDU2MzA3NTc3NjY0NzE5NjYwMw=="} +{"index":4888,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1612.03651","reference_contexts":"We also filter out data to only English (en) after 170k steps using FastText (Joulin et al., 2016)","reference_id":"MTI4NjU5Mzc2MzQ1NjEyNjk3ODQ="} +{"index":4889,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1811.00783","reference_contexts":"Reddit TIFU (Kim et al., 2018) contains summaries of Reddit posts","reference_id":"NDg1OTc4NzkzMzk4ODgyMzUxNQ=="} +{"index":4890,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1412.6980","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4891,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1910.13461","reference_contexts":"The vast majority of text used to pretrain language models is extracted from web pages, while discarding any markup they contain (Liu et al., 2019; Brown et al., 2020; Raffel et al., 2019; Lewis et al., 2019)","reference_id":"MTUwNDI1NzMyNzUxOTc5MjM2NzU="} +{"index":4892,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1910.13461","reference_contexts":"We adopt a BART-style denoising autoencoder (Lewis et al., 2019) for several reasons","reference_id":"NTg0NDcyODU3MDg1OTMxNTM3OQ=="} +{"index":4893,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1910.13461","reference_contexts":"We compare against other pre-training MLM models such as RoBERTa (Liu et al., 2019) , original BART (Lewis et al., 2019) , and T5 (Raffel et al., 2019) by finetuning on the GLUE benchmark (Wang et al., 2018) .","reference_id":"NjMzMTUzNDkwMzkyMzg0OTQ1NA=="} +{"index":4894,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1910.13461","reference_contexts":"We use a modified BART denoising objective (Lewis et al., 2019 ) that randomly masks spans of hyper-text and aims to reconstruct the original input","reference_id":"MTQ3MjQ4MzM0NTQ2MTIyNzIzMjc="} +{"index":4895,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2101.00190","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4896,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1907.11692","reference_contexts":"The vast majority of text used to pretrain language models is extracted from web pages, while discarding any markup they contain (Liu et al., 2019; Brown et al., 2020; Raffel et al., 2019; Lewis et al., 2019)","reference_id":"MTUwNDI1NzMyNzUxOTc5MjM2NzU="} +{"index":4897,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1907.11692","reference_contexts":"We compare against other pre-training MLM models such as RoBERTa (Liu et al., 2019) , original BART (Lewis et al., 2019) , and T5 (Raffel et al., 2019) by finetuning on the GLUE benchmark (Wang et al., 2018) .","reference_id":"NjMzMTUzNDkwMzkyMzg0OTQ1NA=="} +{"index":4898,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2007.02871","reference_contexts":"DART (Nan et al., 2020 ) is a open-domain structured generation dataset containing Wikipedia tables.","reference_id":"MTc0NzE1MTIyODg0NDI1NzkyNTI="} +{"index":4899,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1808.08745","reference_contexts":"XSum (Narayan et al., 2018) provides abstractive single sentence summaries of news articles.","reference_id":"MTU1MTk0MTI1MDgyMjU1MDI1NDQ="} +{"index":4900,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1706.09254","reference_contexts":"E2E (Novikova et al., 2017) is a table-to-text generation dataset containing approximately 50K samples with 8 unique fields from the restaurants domain.","reference_id":"MTQ1ODM0NjY0NzYxNzAwMTk3MDQ="} +{"index":4901,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2105.11447","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4902,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1910.10683","reference_contexts":"The vast majority of text used to pretrain language models is extracted from web pages, while discarding any markup they contain (Liu et al., 2019; Brown et al., 2020; Raffel et al., 2019; Lewis et al., 2019)","reference_id":"MTUwNDI1NzMyNzUxOTc5MjM2NzU="} +{"index":4903,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1910.10683","reference_contexts":"We compare against other pre-training MLM models such as RoBERTa (Liu et al., 2019) , original BART (Lewis et al., 2019) , and T5 (Raffel et al., 2019) by finetuning on the GLUE benchmark (Wang et al., 2018) .","reference_id":"NjMzMTUzNDkwMzkyMzg0OTQ1NA=="} +{"index":4904,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2009.07118","reference_contexts":"We provide the same analysis extended to BART, T5-Large, and HTLM using the same PET prompts provided in Schick and Schütze (2020)","reference_id":"MTI0OTkxMzc3NTk1MTI0OTg4NTQ="} +{"index":4905,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2009.07118","reference_contexts":"By reformulating NLP tasks as cloze-style questions, Schick and Schütze (2020) shows that the prompting capabilities exhibited by GPT3 can occur in language models of a much smaller scale when gradient-based finetuning is combined with task-specific unlabeled data","reference_id":"MTQyNTE4NjcxODIzNTU1ODE5NDE="} +{"index":4906,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"2006.04768","reference_contexts":"Coupling this with the challenges of training transformers on very long sequence lengths (Choromanski et al., 2020; Wang et al., 2020; Beltagy et al., 2020) , it was important to automatically convert web pages to a simplified form, which we call Minimal-HTML (MHTML), as defined below.","reference_id":"MjU3MDUwNjYwMzExOTQ3MzM0NA=="} +{"index":4907,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1905.07830","reference_contexts":"HellaSwag The last dataset we evaluate is the commonsense natural language inference task Hel-laSwag which, due to its adversarial nature, is considered complex (Zellers et al., 2019) Table 3: We evaluate GPT-2 MEDIUM , GPT-2 LARGE and HTLM on table-to-text generation on E2E (left), WebNLG (middle) and DART (right).","reference_id":"MTMyMDc4NDMyNDUwMTYyMDE1NTY="} +{"index":4908,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1912.08777","reference_contexts":"For example, PEGA-SUS (Zhang et al., 2019) proposes a novel pretraining scheme tailored for summarization which involves masking and generating salient gap sentences from a large news corpus","reference_id":"NDQ4OTkzNzI4MjEyNjU2NjE1"} +{"index":4909,"citer_arxiv_id":"2107.06955","cited_arxiv_id":"1912.08777","reference_contexts":"For our baseline, we compare against PEGA-SUS (Zhang et al., 2019) , the current state of the art for zero shot summarization","reference_id":"MzUwMTYyNDEyNzM5NzA0Mjc5Mw=="} +{"index":4910,"citer_arxiv_id":"1801.03924","cited_arxiv_id":"1404.5997","reference_contexts":"But deep networks, even across architectures (Squeezenet [20] , AlexNet [27] , VGG [52] ) and supervision type (supervised [47] , self-supervised [13, 40, 43, 64] , and even unsupervised [26] ), provide an emergent embedding which agrees surprisingly well with humans","reference_id":"OTQ1MjQ2ODYzMjM2MzEyNTcwNg=="} +{"index":4911,"citer_arxiv_id":"1801.03924","cited_arxiv_id":"1404.5997","reference_contexts":"We show performance across methods, including unsupervised [26] , selfsupervised [1, 44, 12, 57, 63, 41, 43, 40, 13, 64] , supervised [27, 52, 20] , and our perceptually-learned metrics (LPIPS)","reference_id":"MjE2NDIyMTc1MTM4MzkzNDAzOQ=="} +{"index":4912,"citer_arxiv_id":"1801.03924","cited_arxiv_id":"1404.5997","reference_contexts":"We use the conv1-conv5 layers from [27]","reference_id":"NDIxNjQ2Nzg3NjIzMDc5ODcwNQ=="} +{"index":4913,"citer_arxiv_id":"1801.03924","cited_arxiv_id":"1404.5997","reference_contexts":"We note that even averaging across all scales and layers, with no further calibration, the AlexNet [27] architecture gives scores near the highest metric, FSIMc [62]","reference_id":"MTYzNDEzMjc5NjkzODE3NzExODc="} +{"index":4914,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1609.03499","reference_contexts":"As concerned with a big part of multimodal machine learning researches, the cross-modal generation task has attracted dramatic attentions, for audio-linguistic [2, 3] , visual-linguistic [4, 5, 6, 7, 8, 9] , audio-visual [10, 11] modalities","reference_id":"NDYyNjA0NDk5ODgwNDc1OTczOQ=="} +{"index":4915,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"Inspired by the autoregressive models for the images [15, 16] , recent works [9, 17] have proposed to formulate the text-to-image synthesis as a sequence-to-sequence problem where the image tokens are learned through discrete VAE [18, 9]","reference_id":"NjM1NzMwNjcyNjA1MzYyOTc2Mg=="} +{"index":4916,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"We utilize dVAE [9] , rather than VQGAN, to discretize the image to visual sequence and build the image from the visual tokens, considering the instability of the training process of GAN","reference_id":"MTM1NTMxMDEzOTcwOTgwMTU0NDc="} +{"index":4917,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"To generate images given the text, we follow the same sampling strategy as in [9] and the generated images are re-ranked with an in-house contrastive learning-based image-text matching model","reference_id":"OTQyNTY2NjMzOTQ4MjU0OTE3Mg=="} +{"index":4918,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"Different from the decode-only [9, 17] or separate encoder-decoder [33] architectures, the encoder and decoder of ERNIE-ViLG share parameters and use specific self-attention masks to control the contexts like UniLM [38] and ERNIE-GEN [39] , where the source tokens are allowed to attend all the source tokens and the target tokens are allowed to attend the source tokens and the target tokens lie left to them","reference_id":"MzYyOTI1NzUyNjU5NjgyNzczNw=="} +{"index":4919,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"As concerned with a big part of multimodal machine learning researches, the cross-modal generation task has attracted dramatic attentions, for audio-linguistic [2, 3] , visual-linguistic [4, 5, 6, 7, 8, 9] , audio-visual [10, 11] modalities","reference_id":"NDYyNjA0NDk5ODgwNDc1OTczOQ=="} +{"index":4920,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"Recently, inspired by the autoregressive generative models for pixel-by-pixel image generation [16] , the transformer-based methods [9, 17] have shown promising results for text-to-image synthesis.","reference_id":"NjgzODExODU2NjAxOTAwMDI2NQ=="} +{"index":4921,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"Based on image quantization techniques [18, 9, 21] , both the image-to-text and text-to-image generation are tackled in autoregressive manners within the same transformer","reference_id":"MjYyNzQ0OTgwNzA4MjkwMDQ4"} +{"index":4922,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"[20] 29.9 NÜWA [33] 12.9 ERNIE-ViLG 7.9 Zero-shot DALL-E [9] 27.5 CogView [17] 27.1 ERNIE-ViLG 14.7","reference_id":"MTM4MTgwMzMyMzQxNTgwNTQ5OTQ="} +{"index":4923,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.12092","reference_contexts":"As shown in Figure 2 , we utilize a similar sparse attention mechanism as in [9]","reference_id":"MTM0ODk0MTY4NTcwOTE1NzU5MjQ="} +{"index":4924,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1606.05328","reference_contexts":"Inspired by the autoregressive models for the images [15, 16] , recent works [9, 17] have proposed to formulate the text-to-image synthesis as a sequence-to-sequence problem where the image tokens are learned through discrete VAE [18, 9]","reference_id":"NjM1NzMwNjcyNjA1MzYyOTc2Mg=="} +{"index":4925,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2105.13290","reference_contexts":"Inspired by the autoregressive models for the images [15, 16] , recent works [9, 17] have proposed to formulate the text-to-image synthesis as a sequence-to-sequence problem where the image tokens are learned through discrete VAE [18, 9]","reference_id":"NjM1NzMwNjcyNjA1MzYyOTc2Mg=="} +{"index":4926,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2105.13290","reference_contexts":"For the generated images, we rerank the samples and select best of 60 samples for the zero-shot experiments for a fair comparison with [17] and best of 10 samples when fine-tuning","reference_id":"MTY2Nzc5MzIwMjI5OTM3NTQ0NTQ="} +{"index":4927,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2105.13290","reference_contexts":"Different from the decode-only [9, 17] or separate encoder-decoder [33] architectures, the encoder and decoder of ERNIE-ViLG share parameters and use specific self-attention masks to control the contexts like UniLM [38] and ERNIE-GEN [39] , where the source tokens are allowed to attend all the source tokens and the target tokens are allowed to attend the source tokens and the target tokens lie left to them","reference_id":"MzYyOTI1NzUyNjU5NjgyNzczNw=="} +{"index":4928,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2105.13290","reference_contexts":"Recently, inspired by the autoregressive generative models for pixel-by-pixel image generation [16] , the transformer-based methods [9, 17] have shown promising results for text-to-image synthesis.","reference_id":"NjgzODExODU2NjAxOTAwMDI2NQ=="} +{"index":4929,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2105.13290","reference_contexts":"[20] 29.9 NÜWA [33] 12.9 ERNIE-ViLG 7.9 Zero-shot DALL-E [9] 27.5 CogView [17] 27.1 ERNIE-ViLG 14.7","reference_id":"MTM4MTgwMzMyMzQxNTgwNTQ5OTQ="} +{"index":4930,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2105.13290","reference_contexts":"ERNIE-ViLG achieves higher quality scores than CogView [17] from all three perspectives which shows that our model acquires better zero-shot text-to-image generation capability","reference_id":"MzUwMTgxODI3MzMzOTQ2Nzg0MQ=="} +{"index":4931,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1711.00937","reference_contexts":"Inspired by the autoregressive models for the images [15, 16] , recent works [9, 17] have proposed to formulate the text-to-image synthesis as a sequence-to-sequence problem where the image tokens are learned through discrete VAE [18, 9]","reference_id":"NjM1NzMwNjcyNjA1MzYyOTc2Mg=="} +{"index":4932,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1711.00937","reference_contexts":"VQVAE [18] proposes to represent images as latent discrete sequences using vector quantization.","reference_id":"Nzg3MTAwNzYzNzQ2NDEwMzMyMQ=="} +{"index":4933,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1711.00937","reference_contexts":"Based on the discrete image tokens learned by various discrete VAE [18, 21] , they try to model the image tokens and text tokens in a single transformer framework, following a unidirectional manner for both the text input and the image target.","reference_id":"ODIxNzk4MzI4NzU3Njg2MDcxNA=="} +{"index":4934,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1711.00937","reference_contexts":"Based on image quantization techniques [18, 9, 21] , both the image-to-text and text-to-image generation are tackled in autoregressive manners within the same transformer","reference_id":"MjYyNzQ0OTgwNzA4MjkwMDQ4"} +{"index":4935,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1711.00937","reference_contexts":"The image is represented as a sequence of discrete representation by vector quantization variational autoencoder (VQVAE) [18]","reference_id":"MTIxNTgyNzQ5ODM5NDYzMzUwOTI="} +{"index":4936,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2009.11278","reference_contexts":"Supervised (fine-tuned) AttnGAN [47] 35.5 DM-GAN [37] 32.6 DF-GAN [48] 21.4 XMC-GAN [8] 9.3 X-LXMERT [19] 37.4 Huang et al","reference_id":"OTg2NjQ1MzAzNTA3NzUxODU1MA=="} +{"index":4937,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2009.11278","reference_contexts":"[19] attempt to enhance the existing Vision-Language Pre-training models with the capability of text-to-image generation, utilizing image discrete tokens generated by K-means clustering of pre-trained image features and generating the image sequence in a non-autoregressive manner","reference_id":"MTMxOTQ4NjA4MTUzODIxMzcwMDA="} +{"index":4938,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2009.11278","reference_contexts":"[19] propose to enhance the VLP models with image generation capability based on discretized visual representation and fine-tuned for text-to-image synthesis","reference_id":"Njg0MDMwNjczMzk4ODIxNDkxNQ=="} +{"index":4939,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1908.02265","reference_contexts":"Vision-Language Pre-training (VLP) models [22, 7, 23, 24, 25, 26, 27] have greatly improved the performance of various Vision-Language tasks, such as VQA and cross-modal retrieval","reference_id":"MTE5MjA2MDAzNzg3Njg0OTcwODM="} +{"index":4940,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1909.11740","reference_contexts":"Utilizing transformer architecture (two-stream or single-stream) to fuse the visual modality and the linguistic modality, they propose various pre-training tasks, from the conventional masked language modeling (MLM), masked region prediction (MRP), image-text matching (ITM) to word-region alignment [23] , scene graph prediction [24] , multimodal conditional text generation [28] , Prefix-LM [29] , etc","reference_id":"MTI1Mzg4NTk3NzgwOTY4MDkwMjk="} +{"index":4941,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1909.11740","reference_contexts":"Vision-Language Pre-training (VLP) models [22, 7, 23, 24, 25, 26, 27] have greatly improved the performance of various Vision-Language tasks, such as VQA and cross-modal retrieval","reference_id":"MTE5MjA2MDAzNzg3Njg0OTcwODM="} +{"index":4942,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2006.16934","reference_contexts":"Utilizing transformer architecture (two-stream or single-stream) to fuse the visual modality and the linguistic modality, they propose various pre-training tasks, from the conventional masked language modeling (MLM), masked region prediction (MRP), image-text matching (ITM) to word-region alignment [23] , scene graph prediction [24] , multimodal conditional text generation [28] , Prefix-LM [29] , etc","reference_id":"MTI1Mzg4NTk3NzgwOTY4MDkwMjk="} +{"index":4943,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2006.16934","reference_contexts":"Vision-Language Pre-training (VLP) models [22, 7, 23, 24, 25, 26, 27] have greatly improved the performance of various Vision-Language tasks, such as VQA and cross-modal retrieval","reference_id":"MTE5MjA2MDAzNzg3Njg0OTcwODM="} +{"index":4944,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2006.06195","reference_contexts":"Vision-Language Pre-training (VLP) models [22, 7, 23, 24, 25, 26, 27] have greatly improved the performance of various Vision-Language tasks, such as VQA and cross-modal retrieval","reference_id":"MTE5MjA2MDAzNzg3Njg0OTcwODM="} +{"index":4945,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2102.02779","reference_contexts":"Utilizing transformer architecture (two-stream or single-stream) to fuse the visual modality and the linguistic modality, they propose various pre-training tasks, from the conventional masked language modeling (MLM), masked region prediction (MRP), image-text matching (ITM) to word-region alignment [23] , scene graph prediction [24] , multimodal conditional text generation [28] , Prefix-LM [29] , etc","reference_id":"MTI1Mzg4NTk3NzgwOTY4MDkwMjk="} +{"index":4946,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2108.10904","reference_contexts":"[29] propose prefix-LM pre-training task, similar to image captioning, solely exploiting language modeling objective","reference_id":"NDcxMjQ0MzgxNTA2NzIwNTg3Ng=="} +{"index":4947,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2108.10904","reference_contexts":"Utilizing transformer architecture (two-stream or single-stream) to fuse the visual modality and the linguistic modality, they propose various pre-training tasks, from the conventional masked language modeling (MLM), masked region prediction (MRP), image-text matching (ITM) to word-region alignment [23] , scene graph prediction [24] , multimodal conditional text generation [28] , Prefix-LM [29] , etc","reference_id":"MTI1Mzg4NTk3NzgwOTY4MDkwMjk="} +{"index":4948,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1504.00325","reference_contexts":"The pre-training datasets are also a big concern of VLP researches where early works utilize million-scale noisy image-text datasets such as Conceptual Captions [30] and human labeled datasets such as COCO Captions [31] and recent works scale up the pre-training with hundreds of millions and nearly billion-scale noisy image-text dataset [29] .","reference_id":"NTQ0MTAxMTU3NjMzMTg4NjA3Ng=="} +{"index":4949,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2111.12233","reference_contexts":"[32] study the scaling behavior of pre-training from both the data and model perspective for image captioning.","reference_id":"NjQ1NDgyNDkzMTQ3NTAzMTQ2MA=="} +{"index":4950,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2111.12417","reference_contexts":"Different from the decode-only [9, 17] or separate encoder-decoder [33] architectures, the encoder and decoder of ERNIE-ViLG share parameters and use specific self-attention masks to control the contexts like UniLM [38] and ERNIE-GEN [39] , where the source tokens are allowed to attend all the source tokens and the target tokens are allowed to attend the source tokens and the target tokens lie left to them","reference_id":"MzYyOTI1NzUyNjU5NjgyNzczNw=="} +{"index":4951,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2111.12417","reference_contexts":"[20] 29.9 NÜWA [33] 12.9 ERNIE-ViLG 7.9 Zero-shot DALL-E [9] 27.5 CogView [17] 27.1 ERNIE-ViLG 14.7","reference_id":"MTM4MTgwMzMyMzQxNTgwNTQ5OTQ="} +{"index":4952,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2111.12417","reference_contexts":"[33] attempt to construct a unified encoder-decoder pre-trained framework aiming at both image and video synthesis tasks.","reference_id":"MjgyNTI4NTc3MjY5NTIwMjA0MA=="} +{"index":4953,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1604.06174","reference_contexts":"In addition, activation recomputation [44] and mixed-precision [45] are applied to reduce the GPU memory footprint and increase throughput","reference_id":"MTQxNTE5NjE3MDU3Mzk5NjE0ODI="} +{"index":4954,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1710.03740","reference_contexts":"In addition, activation recomputation [44] and mixed-precision [45] are applied to reduce the GPU memory footprint and increase throughput","reference_id":"MTQxNTE5NjE3MDU3Mzk5NjE0ODI="} +{"index":4955,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2101.06840","reference_contexts":"Moreover, Optimizer-offload [46] is introduced to swap the sharded optimizer states and master parameters to the CPU which largely reduces the GPU memory footprint","reference_id":"MTc1MzU2MjM5MjU5NDM5NzA5NjY="} +{"index":4956,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"2008.05865","reference_contexts":"Supervised (fine-tuned) AttnGAN [47] 35.5 DM-GAN [37] 32.6 DF-GAN [48] 21.4 XMC-GAN [8] 9.3 X-LXMERT [19] 37.4 Huang et al","reference_id":"OTg2NjQ1MzAzNTA3NzUxODU1MA=="} +{"index":4957,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1505.05612","reference_contexts":"experiments on a public FMIQA [54] dataset, which has freestyle and diversified question-answer annotations","reference_id":"Njk1MTY3MzgyOTY2NTYwMTM5NQ=="} +{"index":4958,"citer_arxiv_id":"2112.15283","cited_arxiv_id":"1505.05612","reference_contexts":"We also make the evaluation for the human-labeled answers to eliminate the bias of different evaluators compared to [54]","reference_id":"MTE2MTMwMjk4MzA4Mjk3MDA0MjE="} +{"index":4959,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2102.08981","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4960,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2102.05918","reference_contexts":"ALIGN (Jia et al., 2021) is a family of multimodal dual encoders that learn to represent images and text in a shared embedding space","reference_id":"MTI2NjcwMTMxMjI5ODM1Nzc0MTI="} +{"index":4961,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2102.05918","reference_contexts":"1 ), we explore dual encoder learning from both image-caption and translation pairs at massive scale: 6 billion translation pairs (Feng et al., 2020) and 1.8 billion imagecaption pairs (Jia et al., 2021)","reference_id":"MTE5NDU4ODExMjc3NTQ2NDUwNTY="} +{"index":4962,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2102.05918","reference_contexts":"The multilingual version of Alt-Text (Jia et al., 2021) is a noisy dataset with 1.8 billion images and their alt-text descriptions, covering 110 languages","reference_id":"NTgxMDg0Mzg0MjUxMjcyMzY4Mg=="} +{"index":4963,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2102.05918","reference_contexts":"Table 2 compares MURAL and previous results (Burns et al., 2020; Ni et al., 2021; Zhou et al., 2021; Jia et al., 2021) in both zero-shot and fine-tuned settings.","reference_id":"MTM4OTM0OTk3NjA4NjMzNzAwMTk="} +{"index":4964,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2102.05918","reference_contexts":"Following ALIGN (Jia et al., 2021) , we use LAMB optimizer (You et al., 2020) with a weight decay ratio of 1e-5","reference_id":"ODU3OTY5Nzk4NTIyMDk0MjgwNw=="} +{"index":4965,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2102.05918","reference_contexts":"In contrast, multimodal dual encoders can be learned directly on noisy, massive image-caption datasets using a simple loss based on in-batch bidirectional retrieval (Jia et al., 2021; Radford et al., 2021)","reference_id":"NTE5NzY2NjY5NzgwNDQwNzA3NA=="} +{"index":4966,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2103.00020","reference_contexts":"In contrast, multimodal dual encoders can be learned directly on noisy, massive image-caption datasets using a simple loss based on in-batch bidirectional retrieval (Jia et al., 2021; Radford et al., 2021)","reference_id":"NTE5NzY2NjY5NzgwNDQwNzA3NA=="} +{"index":4967,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2103.01913","reference_contexts":"The large-scale Wikipedia Image Text (WIT) dataset (Srinivasan et al., 2021) is mined from Wikipedia, covering 108 languages","reference_id":"MTAxODM5MjU5MjUzNDk0Mzg5Nzc="} +{"index":4968,"citer_arxiv_id":"2109.05125","cited_arxiv_id":"2103.01913","reference_contexts":"However, the recent Wikipedia Image-Text (WIT) dataset (Srinivasan et al., 2021) , which covers 108 languages, addresses this gap.","reference_id":"MTE2NzMwOTczNjE5OTY5MjgzMTI="} +{"index":4969,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"2006.06195","reference_contexts":"Prominent work includes ViLBERT (Lu et al., 2019) , LXMERT (Tan and Bansal, 2019) , VL-BERT (Su et al., 2020) , Unicoder-VL (Li et al., 2020a) , B2T2 (Alberti et al., 2019) , UNITER (Chen et al., 2020b) and VILLA (Gan et al., 2020)","reference_id":"MTY5MDczMTkyMDU0NTMwODQ4OTU="} +{"index":4970,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"2004.00849","reference_contexts":"More recently, ViLBERT is enhanced by multi-task learning (Lu et al., 2020) , Oscar (Li et al., 2020b) enhances pre-training with image tags, and Pixel-BERT (Huang et al., 2020) proposes to align image pixels (instead of bottom-up features (Anderson et al., 2018) ) with text","reference_id":"MTc1NjM4MzMwNjk3NDE3NjY0NQ=="} +{"index":4971,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1602.02410","reference_contexts":"Masked Frame Modeling with Noise Contrastive Estimation (MNCE) Instead of directly regressing the real values of masked visual features, we use the softmax version of Noise Contrastive Estimation (NCE) loss (Jozefowicz et al., 2016) , which is widely adopted in self-supervised representation learning (Sun et al., 2019a; Hjelm et al., 2019; Oord et al., 2018)","reference_id":"OTAwNDY1MDkzODU2MTQwNjA4Nw=="} +{"index":4972,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1705.06950","reference_contexts":"For Video Embedder, we first use ResNet (He et al., 2016) pretrained on ImageNet (Deng et al., 2009) and Slow-Fast (Feichtenhofer et al., 2019) pre-trained on Ki-netics (Kay et al., 2017) to extract 2D and 3D visual features for each video frame","reference_id":"MjI0MzM3Nzk5NDgxNDU3NTYwOQ=="} +{"index":4973,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1908.03557","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4974,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1907.11692","reference_contexts":"Since the birth of BERT (Devlin et al., 2019) , there has been continuing advancement in language model pre-training, such as XLNet (Yang et al., 2019) , RoBERTa (Liu et al., 2019) , ALBERT (Lan et al., 2020) , UniLM (Dong et al., 2019) , and T5 (Raffel et al., 2019) , which epitomizes the superb power of large-scale pre-training","reference_id":"NTQ1NjA0NjkzNDQ0NzQ3NjY0Ng=="} +{"index":4975,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"2002.06353","reference_contexts":"So far, VideoBERT (Sun et al., 2019b) , CBT (Sun et al., 2019a) , MIL-NCE (Miech et al., 2020) , Act-BERT (Zhu and Yang, 2020) and UniViLM (Luo et al., 2020) are the only existing work exploring this space, covering downstream tasks from textbased video retrieval (Zhou et al., 2018a; Xu et al., 2016b) and video question answering (Maharaj et al., 2017; Lei et al., 2020a) to video captioning (Zhou et al., 2018b) .","reference_id":"OTk5NTI5MzQ5MjE2MzM3OTk5NQ=="} +{"index":4976,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"2002.06353","reference_contexts":"UniViLM (Luo et al., 2020) takes a step further and considers both understanding and generation tasks.","reference_id":"MTI1NzQwMzYxMjYxMDg3NzMzODc="} +{"index":4977,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1807.03748","reference_contexts":"Masked Frame Modeling with Noise Contrastive Estimation (MNCE) Instead of directly regressing the real values of masked visual features, we use the softmax version of Noise Contrastive Estimation (NCE) loss (Jozefowicz et al., 2016) , which is widely adopted in self-supervised representation learning (Sun et al., 2019a; Hjelm et al., 2019; Oord et al., 2018)","reference_id":"OTAwNDY1MDkzODU2MTQwNjA4Nw=="} +{"index":4978,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1910.10683","reference_contexts":"Since the birth of BERT (Devlin et al., 2019) , there has been continuing advancement in language model pre-training, such as XLNet (Yang et al., 2019) , RoBERTa (Liu et al., 2019) , ALBERT (Lan et al., 2020) , UniLM (Dong et al., 2019) , and T5 (Raffel et al., 2019) , which epitomizes the superb power of large-scale pre-training","reference_id":"NTQ1NjA0NjkzNDQ0NzQ3NjY0Ng=="} +{"index":4979,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1906.05743","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4980,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1902.04094","reference_contexts":"Satellited around BERT, there is parallel growing interest in model compression (Sun et al., 2019c; Shen et al., 2020) and extension to generation tasks (Chen et al., 2020a; Wang and Cho, 2019) .","reference_id":"MTgwODk1MTE4MjI4MzAyMDkzMDk="} +{"index":4981,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1609.08144","reference_contexts":"(2019) and tokenize a subtitle sentence s i into a sequence of WordPieces (Wu et al., 2016) , i.e., w s i = {w j s i } L j=1 (L is the number of tokens in s i )","reference_id":"MTY4OTI2MDE0OTMzNDQ5OTkz"} +{"index":4982,"citer_arxiv_id":"2005.00200","cited_arxiv_id":"1901.06706","reference_contexts":"Inspired by natural language inference (Bowman et al., 2015; Williams et al., 2018) and visual entailment (Xie et al., 2019) , Liu et al","reference_id":"NTUwNTIzNjI1NDI0MjE5NTQ2NA=="} +{"index":4983,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2303.08774","reference_contexts":"Scaling laws in autoregressive language models reveal a power-law relationship between model size, dataset size, and compute with test set cross-entropy loss [31, 25, 1]","reference_id":"Nzc1OTY3ODYwMTM2MjQ5NzE4MA=="} +{"index":4984,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2305.10403","reference_contexts":"Diffusion models [27, 52, 20, 43, 42, 21] , trained to invert the forward paths of data towards random noise, effectively generate images through a continuous denoising process","reference_id":"NzIxMTAxNjI2NjM0NTkwMTEyMg=="} +{"index":4985,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2309.16609","reference_contexts":"Diffusion models [27, 52, 20, 43, 42, 21] , trained to invert the forward paths of data towards random noise, effectively generate images through a continuous denoising process","reference_id":"NzIxMTAxNjI2NjM0NTkwMTEyMg=="} +{"index":4986,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2201.06503","reference_contexts":"Denoising learning mechanisms [27, 41] and sampling efficiency [53, 52, 38, 39, 4] have been continuously optimized to generate high-quality images","reference_id":"NjgyOTExOTQxODY3Mjg5NDQ0NA=="} +{"index":4987,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2403.04692","reference_contexts":"The images of other models are obtained either by querying their open-source demo website (HART [58] ) or running their official inference code locally (Flux-Schnell [34] , SD3-Medium [21] , and PixArt Sigma [12] )","reference_id":"MTIyNzg1MTQxODA0ODEyMTY2ODY="} +{"index":4988,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2403.04692","reference_contexts":"We compare Infinity with four open-source models: PixArt-Sigma [12] , SD3-Medium [21] , SDXL [43] , and HART [58]","reference_id":"MTUxNzUwNDEwMTcyMDg5MDk5MjY="} +{"index":4989,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2403.04692","reference_contexts":"Infinity sets a new record for AutoRegressive models, and also surpasses leading diffusion models including SDXL [43] , PixArt-Sigma [12] ,DALL-E3 [7] and Stable-Diffusion 3 [21] on several challenging text-to-image benchmarks","reference_id":"MTAyNDcyNTQwNDAxODk3NzU5NzM="} +{"index":4990,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.00426","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":4991,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.00426","reference_contexts":"Consequently, mainstream text-to-image diffusion models [21, 7, 14 ] have adopted the DiT architecture.","reference_id":"MTY5MTQ0OTcyMjUwNTc5MzU2ODA="} +{"index":4992,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2309.15807","reference_contexts":"Recently, latent diffusion models [18, 21] have also adopted scaling up VAE to improve the representation in the latent space","reference_id":"MTUxNTgxMjk2ODQ5MzMyMzk2MDc="} +{"index":4993,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2010.14701","reference_contexts":"Scaling laws in autoregressive language models reveal a power-law relationship between model size, dataset size, and compute with test set cross-entropy loss [31, 25, 1]","reference_id":"Nzc1OTY3ODYwMTM2MjQ5NzE4MA=="} +{"index":4994,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2010.14701","reference_contexts":"VAR leverages the powerful scaling properties of LLMs [31, 25] and can simultaneously refine previous scale steps, benefiting from the strengths of diffusion models as well","reference_id":"NjQ1NjMxMDYwNzMwODk4MzU1Nw=="} +{"index":4995,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2203.15556","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":4996,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2203.15556","reference_contexts":"These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [9, 62, 63, 80, 70, 28]","reference_id":"NTg5OTgzNzY0OTIwNjMyNzY0Mg=="} +{"index":4997,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2403.05135","reference_contexts":"On DPG [29] .","reference_id":"MTMzMTU4NzIwNjg2MTYwOTU4Nzc="} +{"index":4998,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2403.05135","reference_contexts":"As for evaluation, we report results on popular text-to-image benchmarks like GenEval [24] and DPG [29]","reference_id":"MTUwMzY4MTg4MjM4NDM5MjE0MTc="} +{"index":4999,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2001.08361","reference_contexts":"Scaling laws in autoregressive language models reveal a power-law relationship between model size, dataset size, and compute with test set cross-entropy loss [31, 25, 1]","reference_id":"Nzc1OTY3ODYwMTM2MjQ5NzE4MA=="} +{"index":5000,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2001.08361","reference_contexts":"VAR leverages the powerful scaling properties of LLMs [31, 25] and can simultaneously refine previous scale steps, benefiting from the strengths of diffusion models as well","reference_id":"NjQ1NjMxMDYwNzMwODk4MzU1Nw=="} +{"index":5001,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2402.17245","reference_contexts":"Our model reaches an overall score of 83.46, surpassing SDXL [43] , Playground v2.5 [36] , and DALLE 3 [7]","reference_id":"MTM2NzgyMjA2MzM1NjA2NDEyMDE="} +{"index":5002,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2410.01756","reference_contexts":"However, the index-wise discrete tokenizer [64, 22, 79, 61, 37, 66, 69, 79] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing finegrained details especially in high-resolution images","reference_id":"MTI0MTM3MjcxMDg3NTI2MDU1MTM="} +{"index":5003,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2211.01095","reference_contexts":"Denoising learning mechanisms [27, 41] and sampling efficiency [53, 52, 38, 39, 4] have been continuously optimized to generate high-quality images","reference_id":"NjgyOTExOTQxODY3Mjg5NDQ0NA=="} +{"index":5004,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.10797","reference_contexts":"This issue poses a significant challenge, as the existing solutions [65, 61, 54, 26, 40] exhibit substantial limitations under such conditions","reference_id":"MTA4ODA4OTI1MDU4MDYxNDc2MzM="} +{"index":5005,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.10797","reference_contexts":"Simply applying RoPE2d [26] or normalized RoPE2d [40] can not distinguish features from different resolutions","reference_id":"MTc3NTQwNjM0OTMwNTMzMjI4OA=="} +{"index":5006,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.01952","reference_contexts":"Infinity sets a new record for AutoRegressive models, and also surpasses leading diffusion models including SDXL [43] , PixArt-Sigma [12] ,DALL-E3 [7] and Stable-Diffusion 3 [21] on several challenging text-to-image benchmarks","reference_id":"MTAyNDcyNTQwNDAxODk3NzU5NzM="} +{"index":5007,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.01952","reference_contexts":"Our model reaches an overall score of 83.46, surpassing SDXL [43] , Playground v2.5 [36] , and DALLE 3 [7]","reference_id":"MTM2NzgyMjA2MzM1NjA2NDEyMDE="} +{"index":5008,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.01952","reference_contexts":"Diffusion models [27, 52, 20, 43, 42, 21] , trained to invert the forward paths of data towards random noise, effectively generate images through a continuous denoising process","reference_id":"NzIxMTAxNjI2NjM0NTkwMTEyMg=="} +{"index":5009,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.01952","reference_contexts":"We compare Infinity with four open-source models: PixArt-Sigma [12] , SD3-Medium [21] , SDXL [43] , and HART [58]","reference_id":"MTUxNzUwNDEwMTcyMDg5MDk5MjY="} +{"index":5010,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.01952","reference_contexts":"Text-to-image generation [50, 46, 45, 7, 43, 21] is one of the most challenging tasks due to its need for complex language adherence and intricate scene creation","reference_id":"MTczMDQ5Mzc4MDEwMjUzNTkzMzA="} +{"index":5011,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2204.06125","reference_contexts":"Text-to-image generation [50, 46, 45, 7, 43, 21] is one of the most challenging tasks due to its need for complex language adherence and intricate scene creation","reference_id":"MTczMDQ5Mzc4MDEwMjUzNTkzMzA="} +{"index":5012,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2111.02114","reference_contexts":"[51] , COYO [10] , OpenImages [33]","reference_id":"MTAzMTE1NDA0MzY3ODU0ODQ1NDc="} +{"index":5013,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2010.02502","reference_contexts":"The visual generation [27, 52, 20, 48, 42] has recently witnessed rapid advancements, enabling highquality, high-resolution images and video synthesis [8, 21]","reference_id":"NDA2OTQ0MTgwMDQ3MzMyMDA0Nw=="} +{"index":5014,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2010.02502","reference_contexts":"Denoising learning mechanisms [27, 41] and sampling efficiency [53, 52, 38, 39, 4] have been continuously optimized to generate high-quality images","reference_id":"NjgyOTExOTQxODY3Mjg5NDQ0NA=="} +{"index":5015,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2010.02502","reference_contexts":"Diffusion models [27, 52, 20, 43, 42, 21] , trained to invert the forward paths of data towards random noise, effectively generate images through a continuous denoising process","reference_id":"NzIxMTAxNjI2NjM0NTkwMTEyMg=="} +{"index":5016,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.06525","reference_contexts":"VQ-based methods [64, 47, 22, 35, 55] employ vector quantization to convert image patches into index-wise tokens and use a decoder-only transformer to predict the next token index","reference_id":"Mzg0MjYxODQxMDA1MDczMjEyMg=="} +{"index":5017,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.06525","reference_contexts":"Inspired by the success and widespread adoption of binary vector quantization [76, 81] , we scaled up the tokenizer vocabulary to 2 64 , significantly surpassing all previous AutoRegressive model vocabularies [77, 55]","reference_id":"MjA5OTY1MjI1MDk1MjU4MDUy"} +{"index":5018,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.05222","reference_contexts":"This has inspired research into scaling in visual generation [56, 78, 61, 21, 8] 3 Infinity Architecture","reference_id":"NzA3MzI0OTIyMzg0ODQ5MDQ1NQ=="} +{"index":5019,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2107.02137","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5020,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2410.10812","reference_contexts":"The images of other models are obtained either by querying their open-source demo website (HART [58] ) or running their official inference code locally (Flux-Schnell [34] , SD3-Medium [21] , and PixArt Sigma [12] )","reference_id":"MTIyNzg1MTQxODA0ODEyMTY2ODY="} +{"index":5021,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2410.10812","reference_contexts":"Especially for the comparison with HART [58] (another SOTA AR-based model), Infinity earns 90.0%, 83.9%, and 93.2% win rate in terms of overall quality, prompt following, and visual aesthetics, respectively","reference_id":"MTEwODU0MTYwNDk2NDc4NDE2OTE="} +{"index":5022,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2410.10812","reference_contexts":"We compare Infinity with four open-source models: PixArt-Sigma [12] , SD3-Medium [21] , SDXL [43] , and HART [58]","reference_id":"MTUxNzUwNDEwMTcyMDg5MDk5MjY="} +{"index":5023,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2410.10812","reference_contexts":"HART [58] adopted hybrid tokenizers based on VAR","reference_id":"NDg0ODYyNTk2OTYzODY3NjkzMw=="} +{"index":5024,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2405.09818","reference_contexts":"Parti [75] , Emu3 [67] , chameleon [59] , loong [68] and VideoPoet [32] scaled up autoregressive models in text-to-image or video synthesis","reference_id":"NDk4ODAzMzU3ODAyOTM3MTc3Ng=="} +{"index":5025,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2312.11805","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5026,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"This has inspired research into scaling in visual generation [56, 78, 61, 21, 8] 3 Infinity Architecture","reference_id":"NzA3MzI0OTIyMzg0ODQ5MDQ1NQ=="} +{"index":5027,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"Previous works [61, 76] directly predict these index labels using a classifier of V d classes","reference_id":"OTQ3OTE2NjEyMTM0NjMzMDkyNQ=="} +{"index":5028,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"Recently, Visual AutoRegressive modeling (VAR) [61] redefined autoregressive learning on images as coarse-to-fine \"next-scale prediction\"","reference_id":"NzQ1MTcyMzE5MDQ2MTE3OTAyMw=="} +{"index":5029,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"VAR [61] inherits the teacher-forcing training from LLMs","reference_id":"MTUwNjMxNjAxNDgwMTA0OTAxMTg="} +{"index":5030,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"Infinity can generate photo-realistic images with various aspect ratios, which is significantly different from VAR [61] that can only generate square images","reference_id":"MzY2NTUwNzIyNDE5NTYwNDkyOQ=="} +{"index":5031,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"However, directly enlarging the vocabulary in existing tokenizers [77, 61] leads to a substantial increase in memory consumption and computational burden","reference_id":"MTM1MDE3MjIxNDQzNDg1MDE5Nzc="} +{"index":5032,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"In previous index-wise [61] representations, the shape of prediction is (h k , w k , V d )","reference_id":"MTMxMDI3MDU1NjYyMjc2MzM4OA=="} +{"index":5033,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"This issue poses a significant challenge, as the existing solutions [65, 61, 54, 26, 40] exhibit substantial limitations under such conditions","reference_id":"MTA4ODA4OTI1MDU4MDYxNDc2MzM="} +{"index":5034,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"However, the index-wise discrete tokenizer [64, 22, 79, 61, 37, 66, 69, 79] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing finegrained details especially in high-resolution images","reference_id":"MTI0MTM3MjcxMDg3NTI2MDU1MTM="} +{"index":5035,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"We replace the original vector quantizer of VAR [61] with a dimension-independent bitwise quantizer","reference_id":"MjIwMjI3NjMzNjY5MjM1ODMwNA=="} +{"index":5036,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5037,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"As a result, the conventional index-wise classifier [61, 11, 77] is difficult to optimize.","reference_id":"NzE0MTU2MzA5MDI1NjQyNTU4MA=="} +{"index":5038,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2404.02905","reference_contexts":"Learnable APE adopted in VAR [61] brings too many parameters and gets confused when the sequence length varies","reference_id":"MzI0ODYyMzE4ODUwNDQ2NzA4MA=="} +{"index":5039,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2302.13971","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5040,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2302.13971","reference_contexts":"These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [9, 62, 63, 80, 70, 28]","reference_id":"NTg5OTgzNzY0OTIwNjMyNzY0Mg=="} +{"index":5041,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2302.13971","reference_contexts":"AutoRegressive models, leveraging the powerful scaling capabilities of LLMs [44, 9, 16, 62, 63] , use discrete image tokenizers [64, 47, 22] in conjunction with transformers to generate images based on next-token prediction","reference_id":"MjAwNTM4MTU3NDEwNTc2NTY0NA=="} +{"index":5042,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.09288","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5043,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.09288","reference_contexts":"These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [9, 62, 63, 80, 70, 28]","reference_id":"NTg5OTgzNzY0OTIwNjMyNzY0Mg=="} +{"index":5044,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2307.09288","reference_contexts":"AutoRegressive models, leveraging the powerful scaling capabilities of LLMs [44, 9, 16, 62, 63] , use discrete image tokenizers [64, 47, 22] in conjunction with transformers to generate images based on next-token prediction","reference_id":"MjAwNTM4MTU3NDEwNTc2NTY0NA=="} +{"index":5045,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.09399","reference_contexts":"However, the index-wise discrete tokenizer [64, 22, 79, 61, 37, 66, 69, 79] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing finegrained details especially in high-resolution images","reference_id":"MTI0MTM3MjcxMDg3NTI2MDU1MTM="} +{"index":5046,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2409.18869","reference_contexts":"Parti [75] , Emu3 [67] , chameleon [59] , loong [68] and VideoPoet [32] scaled up autoregressive models in text-to-image or video synthesis","reference_id":"NDk4ODAzMzU3ODAyOTM3MTc3Ng=="} +{"index":5047,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2409.18869","reference_contexts":"AutoRegressive models encounter significant challenges in highresolution text-to-image synthesis [75, 67]","reference_id":"NDkzNzc1OTg5NDIyMzAxNDIxNw=="} +{"index":5048,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2211.05100","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5049,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2211.05100","reference_contexts":"These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [9, 62, 63, 80, 70, 28]","reference_id":"NTg5OTgzNzY0OTIwNjMyNzY0Mg=="} +{"index":5050,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2306.09341","reference_contexts":"Tab.2 lists the results of two human preference benchmarks, i.e., ImageReward [73] and HPSv2.1 [71]","reference_id":"MTU3NTE4ODkwOTQ3ODkyMjI3ODU="} +{"index":5051,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2306.09341","reference_contexts":"We also measure our method on two human preference evaluation benchmarks, i.e., ImageReward [73] and HPSv2.1 [71]","reference_id":"MTcwMzU0ODQ3NjIwMjMzMTk2NTE="} +{"index":5052,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2408.12528","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5053,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2110.04627","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5054,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2206.10789","reference_contexts":"Parti [75] , Emu3 [67] , chameleon [59] , loong [68] and VideoPoet [32] scaled up autoregressive models in text-to-image or video synthesis","reference_id":"NDk4ODAzMzU3ODAyOTM3MTc3Ng=="} +{"index":5055,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2206.10789","reference_contexts":"AutoRegressive models [15, 22, 75, 61] , on the other hand, harness the scalability and generalizability of language models [16, 2, 28, 62, 63, 70, 57, 3, 60] by employing a visual tokenizer [64, 47, 74] to convert images into discrete tokens and optimize these tokens causally, allowing image generation through next-token prediction or next-scale prediction","reference_id":"NDk0MzgwODc1MDU5NzgzMjY2Mg=="} +{"index":5056,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2206.10789","reference_contexts":"AutoRegressive models encounter significant challenges in highresolution text-to-image synthesis [75, 67]","reference_id":"NDkzNzc1OTg5NDIyMzAxNDIxNw=="} +{"index":5057,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2206.10789","reference_contexts":"We have collected 360 prompts in total, including prompts randomly sampled from Parti [75] and other human-written prompts","reference_id":"MTUzNTMxNDAwMTIzNzMxODQ2MTc="} +{"index":5058,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.05737","reference_contexts":"Previous works [61, 76] directly predict these index labels using a classifier of V d classes","reference_id":"OTQ3OTE2NjEyMTM0NjMzMDkyNQ=="} +{"index":5059,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.05737","reference_contexts":"For binary quantization [76, 81] with code embedding dimension d, V d = 2 d","reference_id":"ODYxNTM2NjI4MTE3NDc4NjM1OQ=="} +{"index":5060,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.05737","reference_contexts":"Inspired by the success and widespread adoption of binary vector quantization [76, 81] , we scaled up the tokenizer vocabulary to 2 64 , significantly surpassing all previous AutoRegressive model vocabularies [77, 55]","reference_id":"MjA5OTY1MjI1MDk1MjU4MDUy"} +{"index":5061,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.05737","reference_contexts":"As a result, the conventional index-wise classifier [61, 11, 77] is difficult to optimize.","reference_id":"NzE0MTU2MzA5MDI1NjQyNTU4MA=="} +{"index":5062,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.05737","reference_contexts":"However, directly enlarging the vocabulary in existing tokenizers [77, 61] leads to a substantial increase in memory consumption and computational burden","reference_id":"MTM1MDE3MjIxNDQzNDg1MDE5Nzc="} +{"index":5063,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.05737","reference_contexts":"Inspired by the success and widespread adoption of binary vector quantization [76, 81] , we scaled up the tokenizer vocabulary to 2 64 , significantly surpassing all previous AutoRegressive model vocabularies [77, 55]","reference_id":"MjA5OTY1MjI1MDk1MjU4MDUy"} +{"index":5064,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2310.05737","reference_contexts":"In this paper, we consider two candidates, LFQ [77] and BSQ [81]","reference_id":"NjUxNDg3OTE5NzU5MTg3OTIzMw=="} +{"index":5065,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2309.02591","reference_contexts":"This has inspired research into scaling in visual generation [56, 78, 61, 21, 8] 3 Infinity Architecture","reference_id":"NzA3MzI0OTIyMzg0ODQ5MDQ1NQ=="} +{"index":5066,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.07550","reference_contexts":"However, the index-wise discrete tokenizer [64, 22, 79, 61, 37, 66, 69, 79] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing finegrained details especially in high-resolution images","reference_id":"MTI0MTM3MjcxMDg3NTI2MDU1MTM="} +{"index":5067,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2205.01068","reference_contexts":"These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [9, 62, 63, 80, 70, 28]","reference_id":"NTg5OTgzNzY0OTIwNjMyNzY0Mg=="} +{"index":5068,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.07548","reference_contexts":"By contrast, since both input and output in BSQ are unit vectors, BSQ [81] proposes an approximation formula for the entropy penalty, reducing the computational complexity to O(d)","reference_id":"ODQyMTg4NjQyNzMzMzgyMjMyMg=="} +{"index":5069,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.07548","reference_contexts":"For binary quantization [76, 81] with code embedding dimension d, V d = 2 d","reference_id":"ODYxNTM2NjI4MTE3NDc4NjM1OQ=="} +{"index":5070,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.07548","reference_contexts":"Inspired by the success and widespread adoption of binary vector quantization [76, 81] , we scaled up the tokenizer vocabulary to 2 64 , significantly surpassing all previous AutoRegressive model vocabularies [77, 55]","reference_id":"MjA5OTY1MjI1MDk1MjU4MDUy"} +{"index":5071,"citer_arxiv_id":"2412.04431","cited_arxiv_id":"2406.07548","reference_contexts":"In this paper, we consider two candidates, LFQ [77] and BSQ [81]","reference_id":"NjUxNDg3OTE5NzU5MTg3OTIzMw=="} +{"index":5072,"citer_arxiv_id":"2009.06489","cited_arxiv_id":"1806.01261","reference_contexts":"This suggests the way a network is organized and our inductive biases is as important as the overall size of the network (Herculano-Houzel et al., 2014; Battaglia et al., 2018; Spelke & Kinzler, 2007)","reference_id":"NzczMTEwOTI4OTk5MTY4MDI3Mw=="} +{"index":5073,"citer_arxiv_id":"2009.06489","cited_arxiv_id":"1805.08974","reference_contexts":"For example, while the parameters almost double between Inception V3 (Szegedy et al., 2016) and Inception V4 architectures (Szegedy et al., 2015a) (from 21.8 to 41.1 million parameters), accuracy on ImageNet differs by less than 2% between the two networks (78.8 vs 80 %) (Kornblith et al., 2018)","reference_id":"MTI3NjQ0MTYwNzg2ODgxMTcwNTc="} +{"index":5074,"citer_arxiv_id":"2009.06489","cited_arxiv_id":"1512.00567","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5075,"citer_arxiv_id":"2009.06489","cited_arxiv_id":"1602.07261","reference_contexts":"For example, while the parameters almost double between Inception V3 (Szegedy et al., 2016) and Inception V4 architectures (Szegedy et al., 2015a) (from 21.8 to 41.1 million parameters), accuracy on ImageNet differs by less than 2% between the two networks (78.8 vs 80 %) (Kornblith et al., 2018)","reference_id":"MTI3NjQ0MTYwNzg2ODgxMTcwNTc="} +{"index":5076,"citer_arxiv_id":"2009.06489","cited_arxiv_id":"2007.05558","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5077,"citer_arxiv_id":"2009.06489","cited_arxiv_id":"2007.05558","reference_contexts":"After decades of treating hardware, software and algorithms as separate choices, the catalysts for closer collaboration include changing hardware economics (Hennessy, 2019) , a \"bigger is better\" race in the size of deep learning architectures (Amodei et al., 2018; Thompson et al., 2020b) and the dizzying requirements of deploying machine learning to edge devices (Warden & Situnayake, 2019) .","reference_id":"NTAyMDE4ODU2MTMwNzE3NjA5MA=="} +{"index":5078,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1508.05326","reference_contexts":"We evaluated our model on the most widely used NLI dataset: The Stanford Natural Language Inference (SNLI) Corpus [3] , which consists of 549,367/9,842/9,824 premise-hypothesis pairs in train/dev/test sets and target labels indicating their relations","reference_id":"MTQwNjUxOTUyODE3MjcxMzA0MjM="} +{"index":5079,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1508.05326","reference_contexts":"StructBERT significantly advances the state-of-the-art results on a variety of NLU tasks, including the GLUE benchmark [27] , the SNLI dataset [3] and the SQuAD v1.1 question answering task [21]","reference_id":"NDEwODkzMTcyMTY3MjE5NDUyMQ=="} +{"index":5080,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1708.00055","reference_contexts":"A pre-trained language model (LM) is a key component in many natural language understanding (NLU) tasks such as semantic textual similarity [4] , question answering [21] and sentiment classification [25]","reference_id":"MTA3Mzg5MzgxODAxNjg4NDMyMjE="} +{"index":5081,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1708.00055","reference_contexts":"The General Language Understanding Evaluation (GLUE) benchmark [27] is a collection of nine NLU tasks, covering textual entailment (RTE [1] and MNLI [29] ), question-answer entailment (QNLI [27] ), paraphrase (MRPC [7] ), question paraphrase (QQP 1 ), textual similarity (STS-B [4] ), sentiment (SST-2 [25] ), linguistic acceptability (CoLA [28] ), and Winograd Schema (WNLI [13] ).","reference_id":"ODY0NjYxMjUwNzE1NTgxOTk2Nw=="} +{"index":5082,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"We can see that even without any [32] 88.9 94.5 89.9 85.0 BERT(ensemble+DA) [6] 86.2 92.2 87.4 93.2 KT-NET(single) [31] 85.1 81.7 85.9 92.4 BERT(single+DA) [6] 84 Table 4: Ablation over the pre-training objectives using StructBERTBase architecture","reference_id":"MTY3MTYwOTgxMTEyNjg3MjkyMDg="} +{"index":5083,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"The input representation follows that used in BERT [6]","reference_id":"ODE0MzA3OTM0NzUzMTQyNjI3Nw=="} +{"index":5084,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"We used documents from English Wikipedia (2,500M words) and BookCorpus [35] as pre-training data, following the preprocessing and the WordPiece tokenization from [6]","reference_id":"MTMyNDEwMjY0OTE3NzY5OTIzMjA="} +{"index":5085,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"Contextualized word representation is considered to be an important part of modern NLP research, with various pre-trained language models [16, 18, 20, 6] emerging recently","reference_id":"MTExNzA2NjYwMjA4NDA4MjgyNDc="} +{"index":5086,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"The next sentence prediction task is considered easy for the original BERT model (the prediction accuracy of BERT can easily achieve 97%-98% in this task [6] )","reference_id":"MTcwNjM1MTAyMDIwNDMwNjQ2OTA="} +{"index":5087,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"In contrast, BERT [6] (as well as its robustly optimized version RoBERTa [15] ) employs a bidirectional Transformer encoder to fuse both the left and the right context, and introduces two novel pre-training tasks for better language understanding","reference_id":"MTI1OTExMjcwMzc2OTYzOTEzMjA="} +{"index":5088,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"While BERT [6] pre-trains a binarized next sentence prediction task to understand sentence relationships, we take one step further and treat it as a sentence ordering task","reference_id":"NTI4NjU3NzU1ODQ0NTEwMDEz"} +{"index":5089,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"For the masked LM objective, we followed the same masking rate and settings as in BERT [6]","reference_id":"NDM0MTI3OTA3NDE4MDk1NjU5NQ=="} +{"index":5090,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"Different from traditional word-specific embedding in which each token is assigned a global representation, recent work, such as Cove [16] , ELMo [18] , GPT [20] and BERT [6] , derives contextualized word vectors from a language model trained on a large text corpus","reference_id":"OTcwOTA0MTA1OTMyODcwMTEyNQ=="} +{"index":5091,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1810.04805","reference_contexts":"In every input sequence, we first mask 15% of all tokens at random, as done in BERT [6]","reference_id":"MTQxOTEwMjQ2NDA4ODExODY0NDQ="} +{"index":5092,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1606.08415","reference_contexts":"The gelu activation [10] was used as done in GPT [20] .","reference_id":"MTM1OTEwMTMxMTUyMTAyNDEzOTU="} +{"index":5093,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1907.10529","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5094,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1901.11504","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5095,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1907.11692","reference_contexts":"Among the context-sensitive language models, BERT (and its robustly optimized version RoBERTa [15] ) has taken the NLP world by storm","reference_id":"MTM0OTUzNzczMjIzODg1MzkyMDk="} +{"index":5096,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1907.11692","reference_contexts":"In contrast, BERT [6] (as well as its robustly optimized version RoBERTa [15] ) employs a bidirectional Transformer encoder to fuse both the left and the right context, and introduces two novel pre-training tasks for better language understanding","reference_id":"MTI1OTExMjcwMzc2OTYzOTEzMjA="} +{"index":5097,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1802.05365","reference_contexts":"ELMo [18] learns two unidirectional LMs based on long short-term memory networks (LSTMs)","reference_id":"MTcyNDM2NzE3NzcyNTMzMDY0MTQ="} +{"index":5098,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1802.05365","reference_contexts":"Different from traditional word-specific embedding in which each token is assigned a global representation, recent work, such as Cove [16] , ELMo [18] , GPT [20] and BERT [6] , derives contextualized word vectors from a language model trained on a large text corpus","reference_id":"OTcwOTA0MTA1OTMyODcwMTEyNQ=="} +{"index":5099,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1802.05365","reference_contexts":"Contextualized word representation is considered to be an important part of modern NLP research, with various pre-trained language models [16, 18, 20, 6] emerging recently","reference_id":"MTExNzA2NjYwMjA4NDA4MjgyNDc="} +{"index":5100,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1811.01088","reference_contexts":"This follows the two-stage transfer learning STILTs introduced in [19]","reference_id":"MTczNDIxNjY0OTg5MjI5NjQxMzY="} +{"index":5101,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1606.05250","reference_contexts":"A pre-trained language model (LM) is a key component in many natural language understanding (NLU) tasks such as semantic textual similarity [4] , question answering [21] and sentiment classification [25]","reference_id":"MTA3Mzg5MzgxODAxNjg4NDMyMjE="} +{"index":5102,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1606.05250","reference_contexts":"StructBERT significantly advances the state-of-the-art results on a variety of NLU tasks, including the GLUE benchmark [27] , the SNLI dataset [3] and the SQuAD v1.1 question answering task [21]","reference_id":"NDEwODkzMTcyMTY3MjE5NDUyMQ=="} +{"index":5103,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1606.05250","reference_contexts":"SQuAD v1.1 is a popular machine reading comprehension dataset consisting of 100,000+ questions created by crowd workers on 536 Wikipedia articles [21]","reference_id":"MTQ2ODM2MTIwMzk5MjYxMjE5MzQ="} +{"index":5104,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1804.07461","reference_contexts":"The General Language Understanding Evaluation (GLUE) benchmark [27] is a collection of nine NLU tasks, covering textual entailment (RTE [1] and MNLI [29] ), question-answer entailment (QNLI [27] ), paraphrase (MRPC [7] ), question paraphrase (QQP 1 ), textual similarity (STS-B [4] ), sentiment (SST-2 [25] ), linguistic acceptability (CoLA [28] ), and Winograd Schema (WNLI [13] ).","reference_id":"ODY0NjYxMjUwNzE1NTgxOTk2Nw=="} +{"index":5105,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1804.07461","reference_contexts":"StructBERT significantly advances the state-of-the-art results on a variety of NLU tasks, including the GLUE benchmark [27] , the SNLI dataset [3] and the SQuAD v1.1 question answering task [21]","reference_id":"NDEwODkzMTcyMTY3MjE5NDUyMQ=="} +{"index":5106,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1804.07461","reference_contexts":"Many important downstream tasks such as STS and NLI [27] are based on understanding the relationship between two text sentences, which is not directly captured by language modeling","reference_id":"NjE1NzQ1MTE4MzI0OTMwODg5Mg=="} +{"index":5107,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1805.12471","reference_contexts":"The General Language Understanding Evaluation (GLUE) benchmark [27] is a collection of nine NLU tasks, covering textual entailment (RTE [1] and MNLI [29] ), question-answer entailment (QNLI [27] ), paraphrase (MRPC [7] ), question paraphrase (QQP 1 ), textual similarity (STS-B [4] ), sentiment (SST-2 [25] ), linguistic acceptability (CoLA [28] ), and Winograd Schema (WNLI [13] ).","reference_id":"ODY0NjYxMjUwNzE1NTgxOTk2Nw=="} +{"index":5108,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1704.05426","reference_contexts":"The General Language Understanding Evaluation (GLUE) benchmark [27] is a collection of nine NLU tasks, covering textual entailment (RTE [1] and MNLI [29] ), question-answer entailment (QNLI [27] ), paraphrase (MRPC [7] ), question paraphrase (QQP 1 ), textual similarity (STS-B [4] ), sentiment (SST-2 [25] ), linguistic acceptability (CoLA [28] ), and Winograd Schema (WNLI [13] ).","reference_id":"ODY0NjYxMjUwNzE1NTgxOTk2Nw=="} +{"index":5109,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1609.08144","reference_contexts":"Texts are tokenized to subword units by WordPiece [30] and absolute positional embeddings are learned with supported sequence lengths up to 512 tokens","reference_id":"NjgwMzYwMTY2NzQ2NTgzOTY4OQ=="} +{"index":5110,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1906.08237","reference_contexts":"We can see that even without any [32] 88.9 94.5 89.9 85.0 BERT(ensemble+DA) [6] 86.2 92.2 87.4 93.2 KT-NET(single) [31] 85.1 81.7 85.9 92.4 BERT(single+DA) [6] 84 Table 4: Ablation over the pre-training objectives using StructBERTBase architecture","reference_id":"MTY3MTYwOTgxMTEyNjg3MjkyMDg="} +{"index":5111,"citer_arxiv_id":"1908.04577","cited_arxiv_id":"1804.09541","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5112,"citer_arxiv_id":"1511.01432","cited_arxiv_id":"1508.01211","reference_contexts":"[31] , which has been successfully used for machine translation [20, 10] , text parsing [32] , image captioning [34] , video analysis [30] , speech recognition [4, 3] and conversational modeling [27, 33]","reference_id":"MTU5NTk0NDA0OTQ2MTgxMzM2NzA="} +{"index":5113,"citer_arxiv_id":"1511.01432","cited_arxiv_id":"1412.1602","reference_contexts":"[31] , which has been successfully used for machine translation [20, 10] , text parsing [32] , image captioning [34] , video analysis [30] , speech recognition [4, 3] and conversational modeling [27, 33]","reference_id":"MTU5NTk0NDA0OTQ2MTgxMzM2NzA="} +{"index":5114,"citer_arxiv_id":"1511.01432","cited_arxiv_id":"1410.8206","reference_contexts":"[31] , which has been successfully used for machine translation [20, 10] , text parsing [32] , image captioning [34] , video analysis [30] , speech recognition [4, 3] and conversational modeling [27, 33]","reference_id":"MTU5NTk0NDA0OTQ2MTgxMzM2NzA="} +{"index":5115,"citer_arxiv_id":"1511.01432","cited_arxiv_id":"1409.2329","reference_contexts":"With random embedding dimension dropout [37] and random word dropout (not published previously), we are able to reach performance of around 86.5% accuracy in the test set, which is approximately 5% worse than most baselines.","reference_id":"ODgyODU0MTA5NDM1MzkzMDYyNg=="} +{"index":5116,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2301.00704","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5117,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2304.00186","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5118,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2108.02938","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5119,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2211.11337","reference_contexts":"Optimization-based methods, such as DreamArtist (Dong, Wei, and Lin 2022) , directly optimize the pseudo word to establish a mapping between userprovided images and textual inversion","reference_id":"MTM2MDI2ODY3MzAwMDk2OTkyNzk="} +{"index":5120,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2211.11337","reference_contexts":"2022) , DreamArtist (Dong, Wei, and Lin 2022) , and CustomDiffusion (Kumari et al","reference_id":"MzM2ODY4MDIwNjEzNzQxNDgwOA=="} +{"index":5121,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2208.01618","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5122,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2106.09685","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5123,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"1710.10196","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5124,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2307.11410","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5125,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2112.10741","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5126,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2204.06125","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5127,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2307.06949","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5128,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2304.03411","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5129,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2306.00983","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5130,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2010.02502","reference_contexts":"However, due to the iterative nature of diffusion model-based image generation, inverting an image into the noise space associated with the image (e.g., DDIM (Song, Meng, and Ermon 2020) , DDIB (Su et al","reference_id":"MTE0MjQ1Nzc5NTA5OTQ1NDY4NjU="} +{"index":5131,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2203.08382","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5132,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2302.13848","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5133,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2210.05559","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5134,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2206.10789","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5135,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2306.00926","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5136,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2302.05543","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5137,"citer_arxiv_id":"2309.05793","cited_arxiv_id":"2305.13579","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5138,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"1607.06450","reference_contexts":"Layer normalization [1] is a key part in transformer for stable training and faster convergence","reference_id":"MTI3NDExMTA5MDQzMzE2NDg3MTA="} +{"index":5139,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"1906.07155","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5140,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"2012.12556","reference_contexts":"Nowadays, transformer architectures have been used in a growing number of computer vision tasks [11] such as image recognition [7, 44, 33] , object detection [50] , and segmentation [47, 42] .","reference_id":"MTEwMTQ4MDc0Nzc0MzYxNTA0MTI="} +{"index":5141,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"1606.08415","reference_contexts":"where W and b are the weight and bias term of fully-connected layer respectively, and σ(•) is the activation function such as GELU [14] .","reference_id":"NjkwMDg3OTg2MzAxNDI3MzExNQ=="} +{"index":5142,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"1605.07648","reference_contexts":"The main advanced technologies apart from common settings [13] include AdamW [23] , label smoothing [31] , DropPath [20] , and repeated augmentation [15]","reference_id":"NzEzOTIzOTU3NDYyMDc2NzQyOA=="} +{"index":5143,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"1711.05101","reference_contexts":"The main advanced technologies apart from common settings [13] include AdamW [23] , label smoothing [31] , DropPath [20] , and repeated augmentation [15]","reference_id":"NzEzOTIzOTU3NDYyMDc2NzQyOA=="} +{"index":5144,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"1711.05101","reference_contexts":"All the compared models are trained using AdamW [23] with batch size of 16 for 50 epochs","reference_id":"Nzk3MTIxMjIwODMyMzcxMDUw"} +{"index":5145,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"2101.11986","reference_contexts":"Nowadays, transformer architectures have been used in a growing number of computer vision tasks [11] such as image recognition [7, 44, 33] , object detection [50] , and segmentation [47, 42] .","reference_id":"MTEwMTQ4MDc0Nzc0MzYxNTA0MTI="} +{"index":5146,"citer_arxiv_id":"2103.00112","cited_arxiv_id":"1710.09412","reference_contexts":"We adopt the same data augmentation strategy as that in DeiT [35] including random crop, random clip, Rand-Augment [8] , Random Erasing [48] , Mixup [46] and CutMix [45]","reference_id":"MTAyNDEzMjg4ODcwODI2Mjg2ODk="} +{"index":5147,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1809.10853","reference_contexts":"(Rae et al., 2018) 29.0 29.2 Transformer (Baevski and Auli, 2019) -18.7 18L TransformerXL, M=384 (Dai et al.,","reference_id":"MjQ4NDE1NDM4OTU4NTY1MDgwNg=="} +{"index":5148,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1409.0473","reference_contexts":"The Transformer (Vaswani et al., 2017 ) is a sequence model which stores the hidden activation of every time-step, and integrates this information using an attention operator (Bahdanau et al., 2014)","reference_id":"MTIyNzQwOTY5NzU0Nzc1OTM3ODc="} +{"index":5149,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1612.03801","reference_contexts":"We test the Compressive Transformer on a challenging memory task within the DMLab-30 (Beattie et al., 2016) domain, rooms select nonmatching object","reference_id":"MTMzNzI5NjkyNTMwMTg1MTI5MjQ="} +{"index":5150,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1312.3005","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5151,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1904.10509","reference_contexts":"(Child et al., 2019) 0.991 Adaptive Transf","reference_id":"MTM4NTQ4NTM4NDE4OTQ2MTc0Mg=="} +{"index":5152,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1904.10509","reference_contexts":"The Sparse Transformer (Child et al., 2019) uses fixed sparse attention masks to attend to roughly √ n locations in memory","reference_id":"MTQxOTYxNTQ4MjgzNzIwMzA2MTg="} +{"index":5153,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1904.10509","reference_contexts":"Several works have focused on reducing the computational cost of attention with sparse access mechanisms (Rae et al., 2016; Child et al., 2019; Sukhbaatar et al., 2019; Lample et al., 2019)","reference_id":"MTUyMTA0OTE0MDUyNjY1MzkyNzg="} +{"index":5154,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1609.01704","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5155,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1901.02860","reference_contexts":"With this granular memory, the Transformer has brought about a step-change in state-of-the-art performance, within machine translation (Vaswani et al., 2017) , language modelling (Dai et al., 2019; Shoeybi et al., 2019) , video captioning (Zhou et al., 2018) , and a multitude of language understanding benchmarks (Devlin et al., 2018; Yang et al., 2019) amongst others.","reference_id":"MTY0MDc0MDE1MDU0MjY1MTY4MjI="} +{"index":5156,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1901.02860","reference_contexts":"However recent Transformer models, such as the TransformerXL (Dai et al., 2019) appear to be able to exploit temporal dependencies on the order of several thousand words","reference_id":"NjM1MjA4NTI0Njg1NzE4MzQ2Nw=="} +{"index":5157,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1901.02860","reference_contexts":"We build on the ideas of the TransformerXL (Dai et al., 2019) which maintains a memory of past activations at each layer to preserve a longer history of context","reference_id":"MTQ3NjgzMjUzMzQ4MzI0Mzc0MTg="} +{"index":5158,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1901.02860","reference_contexts":"(2019) 1.06 24L TXL (Dai et al., 2019) 0.99 Sparse Transf","reference_id":"MTY3NTMyNjQ3ODI2NzcwMDc2MjM="} +{"index":5159,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1612.08083","reference_contexts":"(Bai et al., 2018a) -45.2 GCNN-14 (Dauphin et al., 2016) -37.2 Quasi- RNN Bradbury et al","reference_id":"NTk4MTcwMTA0NDEwMDc1Nzc3Nw=="} +{"index":5160,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1810.04805","reference_contexts":"With this granular memory, the Transformer has brought about a step-change in state-of-the-art performance, within machine translation (Vaswani et al., 2017) , language modelling (Dai et al., 2019; Shoeybi et al., 2019) , video captioning (Zhou et al., 2018) , and a multitude of language understanding benchmarks (Devlin et al., 2018; Yang et al., 2019) amongst others.","reference_id":"MTY0MDc0MDE1MDU0MjY1MTY4MjI="} +{"index":5161,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1612.04426","reference_contexts":"This dataset has been a popular recent LM benchmark due to the potential to exploit longer-range dependencies (Grave et al., 2016; Rae et al., 2018; Bai et al., 2018b)","reference_id":"OTAxNjYxOTE1MjU5NzU2MTkyMg=="} +{"index":5162,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1308.0850","reference_contexts":"7L LSTM (Graves, 2013) 1.67 LN HyperNetworks Ha et al","reference_id":"MTMwMDg4MDg3NTE2NjAxMjAyODM="} +{"index":5163,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1410.5401","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5164,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1609.09106","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5165,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1511.02301","reference_contexts":"Texts extracted from books have been used for prior NLP benchmarks; such as the Children's Book Test (Hill et al., 2015) and LAMBADA (Paperno et al., 2016)","reference_id":"ODIzNTY3Nzg4Nzk1NzI2ODYwNQ=="} +{"index":5166,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1904.09751","reference_contexts":"We use Nucleus Sampling with p = 0.98 (Holtzman et al., 2019)","reference_id":"MTIwMDA2MTIyMzYwMDAzODg4OTQ="} +{"index":5167,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1610.10099","reference_contexts":"(2016) 1.32 ByteNet (Kalchbrenner et al., 2016) 1.31 RHN Zilly et al","reference_id":"Njc2MzQ0MDg0NjE1NTY2ODQxMQ=="} +{"index":5168,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1412.6980","reference_contexts":"We optimised all models with Adam (Kingma and Ba, 2014)","reference_id":"MTE5OTYzOTU3NTcxMjk0MzgzNQ=="} +{"index":5169,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1609.07959","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5170,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1907.05242","reference_contexts":"Several works have focused on reducing the computational cost of attention with sparse access mechanisms (Rae et al., 2016; Child et al., 2019; Sukhbaatar et al., 2019; Lample et al., 2019)","reference_id":"MTUyMTA0OTE0MDUyNjY1MzkyNzg="} +{"index":5171,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1609.07843","reference_contexts":"We observe this improves the modelling of text, achieving state-of-the-art results in character-based language modelling -0.97 bpc on Enwik8 from the Hutter Prize (Hutter, 2012) -and word-level language modelling -17.1 perplexity on WikiText-103 (Merity et al., 2016)","reference_id":"MTYwODY3MjgxNzE4MTI5ODIyMzQ="} +{"index":5172,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1609.03499","reference_contexts":"We show the Compressive Transformer works not only for language, but can also model the waveform of high-frequency speech with a trend of lower likelihood than the TransformerXL and Wavenet (Oord et al., 2016) when trained over 400,000 steps","reference_id":"MTgzNjg0NzU2NTM2NTczODc4NTQ="} +{"index":5173,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1609.03499","reference_contexts":"We chunk the sequences into windows of size 3840, roughly 80ms of audio, and compare a 20-layer Compressive Transformer to a 20-layer TransformerXL and a 30-layer WaveNet model (Oord et al., 2016) -a state-of-the-art audio generative model used to serve production speech synthesis applications at Google (Oord et al., 2018)","reference_id":"NjUyNTc0OTI2MzQ3NzYxNTE5OQ=="} +{"index":5174,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1609.03499","reference_contexts":"In both cases, we compare to very strong baselines (Wavenet (Oord et al., 2016) and IMPALA (Espeholt et al., 2018) ).","reference_id":"MzQ1NDE5OTU5MDA4Mzk4MjMz"} +{"index":5175,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1803.10049","reference_contexts":"This dataset has been a popular recent LM benchmark due to the potential to exploit longer-range dependencies (Grave et al., 2016; Rae et al., 2018; Bai et al., 2018b)","reference_id":"OTAxNjYxOTE1MjU5NzU2MTkyMg=="} +{"index":5176,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1803.10049","reference_contexts":"(Rae et al., 2018) 29.0 29.2 Transformer (Baevski and Auli, 2019) -18.7 18L TransformerXL, M=384 (Dai et al.,","reference_id":"MjQ4NDE1NDM4OTU4NTY1MDgwNg=="} +{"index":5177,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1905.07799","reference_contexts":"The WikiText-103 model was trained with an embedding size of 1024, adaptive inputs using the same parameters as (Sukhbaatar et al., 2019) , 16 attention heads, 18 layers, an mlp hidden size of 4096, a sequence window of size 512 and a memory of size 512","reference_id":"Mzk1NDM0NDY0NDQ2MDYzNjE2NQ=="} +{"index":5178,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1905.07799","reference_contexts":"(Sukhbaatar et al., 2019) 0.98","reference_id":"NDM0NzgwNzgxMjg4NDk0NjgzMA=="} +{"index":5179,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1905.07799","reference_contexts":"Several works have focused on reducing the computational cost of attention with sparse access mechanisms (Rae et al., 2016; Child et al., 2019; Sukhbaatar et al., 2019; Lample et al., 2019)","reference_id":"MTUyMTA0OTE0MDUyNjY1MzkyNzg="} +{"index":5180,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1901.10430","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5181,"citer_arxiv_id":"1911.05507","cited_arxiv_id":"1906.08237","reference_contexts":"With this granular memory, the Transformer has brought about a step-change in state-of-the-art performance, within machine translation (Vaswani et al., 2017) , language modelling (Dai et al., 2019; Shoeybi et al., 2019) , video captioning (Zhou et al., 2018) , and a multitude of language understanding benchmarks (Devlin et al., 2018; Yang et al., 2019) amongst others.","reference_id":"MTY0MDc0MDE1MDU0MjY1MTY4MjI="} +{"index":5182,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1907.02893","reference_contexts":"We call these structural failure modes, because they are often diagnosed as a misalignment between the predictor learned by empirical risk minimization and the causal structure of the desired predictor (Schölkopf, 2019; Arjovsky et al., 2019)","reference_id":"MTY3MTAwNDc1NjQ3NTY5NjUwMzM="} +{"index":5183,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1907.02893","reference_contexts":"In particular, concerns regarding \"spurious correlations\" and \"shortcut learning\" in trained models are now widespread (e.g., Geirhos et al., 2020; Arjovsky et al., 2019) .","reference_id":"MTc4MDQzNzQ4ODg4NDUwNjk1MzA="} +{"index":5184,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1907.02893","reference_contexts":"In such cases, the iid-optimal predictors must necessarily incorporate spurious associations (Caruana et al., 2015; Arjovsky et al., 2019; Ilyas et al., 2019)","reference_id":"NjQzNTQ5ODg4NjY2MjA1NjYyMQ=="} +{"index":5185,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1709.08568","reference_contexts":"Although some general-purpose heuristics have been proposed (Bengio, 2017) , proposals for expressing application-specific requirements with flexible but unified frameworks may be a promising middle ground solution","reference_id":"MTM4NDgxMjQ4ODc1MTQ4NjgwMw=="} +{"index":5186,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2002.06305","reference_contexts":"Using the same protocol, we replicate the results (McCoy et al., 2019a; Dodge et al., 2020; Zhou et al., 2020) , and extend them to show sensitivity to the pretraining random seed","reference_id":"Mjc0MDU2NTE5MjAyNTk1NTI5"} +{"index":5187,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2002.06305","reference_contexts":"Notably, previous work has shown that performance on these stronger stress tests has also been shown to be unstable with respect to the fine-tuning seed (Zhou et al., 2020; McCoy et al., 2019a; Dodge et al., 2020)","reference_id":"MzE4NTUwNjgwMzk1OTYxOTkzNA=="} +{"index":5188,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2004.07780","reference_contexts":"In particular, concerns regarding \"spurious correlations\" and \"shortcut learning\" in trained models are now widespread (e.g., Geirhos et al., 2020; Arjovsky et al., 2019) .","reference_id":"MTc4MDQzNzQ4ODg4NDUwNjk1MzA="} +{"index":5189,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2004.07780","reference_contexts":"However, recent work has shown that NLP systems built with this pattern often rely on \"shortcuts\" (Geirhos et al., 2020) , which may be based on spurious phenomena in the training data (McCoy et al., 2019b)","reference_id":"Njg5NDMwNTYzNTgyNDk3Nzk0"} +{"index":5190,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1907.07174","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5191,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2006.16241","reference_contexts":"The fact that underspecification persists in the BiT models is particularly notable, because simultaneously scaling up data and model size has been shown to improve performance across a wide range of robustness stress tests, aligning closely with how much these models improve performance on iid evaluations (Djolonga et al., 2020; Taori et al., 2020; Hendrycks et al., 2020)","reference_id":"MTEyMzc3MzkzMjgwMTk1Nzg3NTg="} +{"index":5192,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2009.06489","reference_contexts":"These include the particular scheme used for initialization; conventions for parameterization; choice of optimization algorithm; conventions for representing data; and choices of batch size, learning rate, and other hyperparameters, all of which may interact with the infrastructure available for training and serving models (Hooker, 2020)","reference_id":"NTQ3MDY1NTQwMTE4MDkzOTk3OQ=="} +{"index":5193,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1803.05407","reference_contexts":"Chaudhari et al., 2019) , and connectivity between global modes in the context of model averaging (Izmailov et al., 2018; Fort et al., 2019; Wilson and Izmailov, 2020) and network pruning (Frankle et al., 2020)","reference_id":"MTA3MzUyMDkyNTI0MTE4NzA2MzI="} +{"index":5194,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1912.11370","reference_contexts":"We examine two models: the ResNet-50 model (He et al., 2016) trained in ImageNet, and a ResNet-101x3 Big Transfer (BiT) model (Kolesnikov et al., 2019) pre-trained on the JFT-300M dataset (Sun et al., 2017) and fine-tuned on ImageNet","reference_id":"MTIxMTA2NDg4MDA0MzU3MTk3MTE="} +{"index":5195,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1909.11942","reference_contexts":"(2019) , although better correlations have subsequently been obtained by pretraining on larger datasets (Liu et al., 2019; Lan et al., 2019; Yang et al., 2019) .","reference_id":"MjYxNzUwODMzNTYyNTg4MDYwNQ=="} +{"index":5196,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1907.11692","reference_contexts":"(2019) , although better correlations have subsequently been obtained by pretraining on larger datasets (Liu et al., 2019; Lan et al., 2019; Yang et al., 2019) .","reference_id":"MjYxNzUwODMzNTYyNTg4MDYwNQ=="} +{"index":5197,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1902.01007","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5198,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"In order to evaluate the asymptotics of this expression, we recall some formulas that follow from Mei and Montanari (2019) :","reference_id":"NjU1NzUxNjUyMzM1Mzc4NzUwNw=="} +{"index":5199,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"Theorem 1 in Mei and Montanari (2019) holds for the risk, conditional on the realization of X, y","reference_id":"NTM4MjAyMDkzMjE2Nzk4OTQ4MA=="} +{"index":5200,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"Building on the work of Mei and Montanari (2019) we can determine the precise asymptotics of the risk under certain distribution shifts in the limit n, N, d → ∞ with fixed ratios n/d, N/n","reference_id":"NjgzOTM0MDEzNjc1MzM0MTAyOA=="} +{"index":5201,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"We will work within the 'noisy linear features model' of Mei and Montanari (2019) which replaces U σ with","reference_id":"MTMwNTc5OTcyNDI3NDE0NjE1OQ=="} +{"index":5202,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"(15) We then have, from (Mei and Montanari, 2019 , Theorem 1), the following characterization of the in-distribution risk 8 R(W , P) = r 2 E 1 (ζ, ψ 1 , ψ 2 , λ) E 0 (ζ, ψ 1 , ψ 2 , λ)","reference_id":"NTYyMjQ1Mjg2NTAzMTczNzk0MA=="} +{"index":5203,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"We begin by recalling some results and notations from Mei and Montanari (2019)","reference_id":"NzY1NTYwMzI3MDMwMzg3MDY5OA=="} +{"index":5204,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"Underspecification also plays a role in recent analyses of overparametrization in theoretical and real deep learning models (Belkin et al., 2018; Mei and Montanari, 2019; Nakkiran et al., 2020)","reference_id":"MTMzMDk3MjIxNzg0MzA2NzU3NDU="} +{"index":5205,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1908.05355","reference_contexts":"The universality results of Mei and Montanari (2019) suggest that this substitution produces an error that is asymptotically negligible, namely:","reference_id":"MzIxMjk1MjczNDAyNDA1MDkxOQ=="} +{"index":5206,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1301.3781","reference_contexts":"","reference_id":"NjQ3MDk4OTA5MTIwNjg3ODc3MQ=="} +{"index":5207,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2004.09456","reference_contexts":"Specifically, we examine the performance of our ensemble of five BERT checkpoints on the StereoSet benchmark (Nadeem et al., 2020) .","reference_id":"MTQwNTkzNjMxMzM2NjQ1NjAxODg="} +{"index":5208,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2004.09456","reference_contexts":"(Nadeem et al., 2020)","reference_id":"MTExMzYxMTM0MTQ0NzI4Mjc0ODM="} +{"index":5209,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1911.10500","reference_contexts":"We call these structural failure modes, because they are often diagnosed as a misalignment between the predictor learned by empirical risk minimization and the causal structure of the desired predictor (Schölkopf, 2019; Arjovsky et al., 2019)","reference_id":"MTY3MTAwNDc1NjQ3NTY5NjUwMzM="} +{"index":5210,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"1911.10500","reference_contexts":"Causal DAGs (Schölkopf, 2019) and explanations (Ross et al., 2017) are both promising candidates for such frameworks.","reference_id":"MTAwNDcxNTYyNTQ4Mjg5MDc5NzE="} +{"index":5211,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2007.00644","reference_contexts":"Taori et al., 2020; Djolonga et al., 2020)","reference_id":"MTc1MzQ1NDkxNDcxMTUwODk4NTQ="} +{"index":5212,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2007.00644","reference_contexts":"Shifted performance evaluations form the backbone of empirical evaluations in the literature on robust machine learning and task adaptation (e.g., Hendrycks and Dietterich, 2019; Wang et al., 2019; Djolonga et al., 2020; Taori et al., 2020)","reference_id":"NjU5MzA5NTA5ODM1ODA2MjA3"} +{"index":5213,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2007.00644","reference_contexts":"This has given rise to an active literature on robustness in image classification (see, e.g., Taori et al., 2020; Djolonga et al., 2020)","reference_id":"ODE3MjAyMjM0MjIzMTQ1Njk2Ng=="} +{"index":5214,"citer_arxiv_id":"2011.03395","cited_arxiv_id":"2007.00644","reference_contexts":"The fact that underspecification persists in the BiT models is particularly notable, because simultaneously scaling up data and model size has been shown to improve performance across a wide range of robustness stress tests, aligning closely with how much these models improve performance on iid evaluations (Djolonga et al., 2020; Taori et al., 2020; Hendrycks et al., 2020)","reference_id":"MTEyMzc3MzkzMjgwMTk1Nzg3NTg="} +{"index":5215,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1712.09763","reference_contexts":"The network architecture draws inspiration from existing autoregressive models for images [45, 49, 48, 5, 41, 36, 7]","reference_id":"OTY5OTUxMzAwMDkxNDg5MDgxNg=="} +{"index":5216,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1712.09763","reference_contexts":"Subsequent works in autoregressive image modelling have steadily improved state-of-the-art for image density estimation [48, 41, 36, 5, 7]","reference_id":"OTQwMzUxODcyMjY3OTI0NjU0OA=="} +{"index":5217,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1904.10509","reference_contexts":"The network architecture draws inspiration from existing autoregressive models for images [45, 49, 48, 5, 41, 36, 7]","reference_id":"OTY5OTUxMzAwMDkxNDg5MDgxNg=="} +{"index":5218,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1904.10509","reference_contexts":"Subsequent works in autoregressive image modelling have steadily improved state-of-the-art for image density estimation [48, 41, 36, 5, 7]","reference_id":"OTQwMzUxODcyMjY3OTI0NjU0OA=="} +{"index":5219,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1802.04208","reference_contexts":"Recent works have used generative adversarial networks (GANs) [14] to model both waveforms and spectral representations [12, 13]","reference_id":"MTA2Njk0MzI2OTUwMjA1MTI3NzU="} +{"index":5220,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1308.0850","reference_contexts":"1 We include samples generated using the priming and biasing procedures described by Graves [15]","reference_id":"NTU2NjcwODAzMjg1MjY3OTk2Mg=="} +{"index":5221,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1308.0850","reference_contexts":"We use an attention mechanism which is a straightforward variant of the location-based Gaussian mixture attention introduced by Graves [15]","reference_id":"MTc2OTI5Njg2MTAyOTc1NjY3NTY="} +{"index":5222,"citer_arxiv_id":"1906.01083","cited_arxiv_id":"1308.0850","reference_contexts":"Following the work on Mixture Density Networks [4] and their application to autoregressive models [15] , θ ij is modelled as the output of a neural network and computed as a function of the context x