id
stringlengths
66
88
text
stringlengths
46
1.7k
duration
float64
2.32
30
dnsmos
float64
2.8
3.65
sampling_rate
int64
24k
24k
path
stringlengths
79
101
audio_content_token_indices
listlengths
58
749
audio_global_embedding
listlengths
128
128
audio_token_len
int64
58
749
audio_duration
float64
2.32
30
bilibili_data_1622095_BV1dZ421J7JT_p38_BV1dZ421J7JT_p38_m4-dialogue_0622695
[S1] They did. It was a great game. The Jays have a new pitcher and he's fantastic. What about you? Do you like to watch baseball? [S2] Pitcher. [S1] I think baseball games take too long. I don't have the patience for it. I like the fast pace of basketball. I like to watch NBA games.
24.32
2.9664
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p38_BV1dZ421J7JT_p38_m4-dialogue_0622695.mp3
[ 9818, 561, 7936, 6119, 1500, 4548, 5061, 4618, 7872, 3550, 6559, 11163, 11041, 11106, 11622, 2015, 2013, 1005, 3651, 1300, 9443, 12570, 8418, 9825, 11412, 9313, 6757, 12182, 11162, 12499, 1153, 3220, 3299, 2670, 8439, 9015, 9070, 11630, 116...
[ 0.5482810735702515, -5.709695816040039, 0.03976389020681381, 0.3816370666027069, 0.568026602268219, 0.021455101668834686, 0.18662939965724945, -0.04409167170524597, 0.1937228888273239, -3.9149818420410156, 0.29500776529312134, -0.0033186418004333973, -0.053534653037786484, 0.07132484018802...
608
24.32
bilibili_data_1622095_BV1dZ421J7JT_p38_BV1dZ421J7JT_p38_m4-dialogue_0622696
[S1] Pace. [S2] Got it. Next spring, we can watch the NBA finals together on TV. [S1] Sure thing. I look forward to it.
9.96
3.128996
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p38_BV1dZ421J7JT_p38_m4-dialogue_0622696.mp3
[ 6867, 9248, 12563, 9904, 8098, 7602, 10227, 9191, 11156, 8604, 11659, 9621, 9484, 6988, 6483, 3418, 3418, 3418, 2913, 3360, 3946, 7010, 6515, 9953, 12377, 9736, 12499, 12633, 9769, 12498, 11987, 9866, 11924, 10083, 9882, 4635, 12573, 12563, ...
[ 0.669719398021698, -5.814965724945068, -0.015439272858202457, 0.44194069504737854, 0.3749578595161438, 0.011048446409404278, 0.27468159794807434, -0.06476051360368729, 0.19006407260894775, -3.9862165451049805, 0.15848605334758759, 0.0833197608590126, -0.16347967088222504, 0.029708739370107...
249
9.96
bilibili_data_1622095_BV1dZ421J7JT_p39_BV1dZ421J7JT_p39_m4-dialogue_0692567
[S1] Hmm. There's a new action film with The Rock. There's a new animation film from Disney. There are a few horror movies, and there's a romantic comedy with Will Smith. Action. [S2] Animation. [S1] Horror. [S2] Horror. [S1] Romantic comedy.
24.08
2.934948
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p39_BV1dZ421J7JT_p39_m4-dialogue_0692567.mp3
[ 9363, 8342, 8342, 6295, 4247, 3799, 10069, 6878, 4255, 6871, 4309, 8939, 6763, 12626, 12442, 6675, 5729, 10005, 9363, 9781, 9436, 7404, 7265, 3113, 12506, 3609, 1429, 12726, 10582, 2949, 3854, 6080, 358, 1334, 3091, 8716, 9175, 10579, 10388...
[ 0.18859045207500458, -5.766045570373535, 0.12509410083293915, 0.21059550344944, 0.5784578323364258, 0.031011687591671944, 0.5227056741714478, 0.03265034034848213, 0.09263702481985092, -3.915741205215454, 0.17057693004608154, -0.03913243114948273, -0.2747335731983185, 0.15779054164886475, ...
602
24.08
bilibili_data_1622095_BV1dZ421J7JT_p42_BV1dZ421J7JT_p42_m4-dialogue_0046692
[S1] You look excited. [S2] I am. I just bought tickets for Maroon 5's concert in Tokyo in December. [S1] Concert. [S2] Concert. [S1] I heard about their concert tour. Aren't they coming to Taipei, too? Why didn't you buy their Taipei concert tickets? [S2] Concert tour. [S1] Concert tour.
20.16
3.13571
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p42_BV1dZ421J7JT_p42_m4-dialogue_0046692.mp3
[ 7186, 4688, 9987, 10059, 9604, 10515, 10394, 7770, 5330, 10547, 11630, 7329, 12562, 8686, 4055, 9940, 12644, 3562, 5984, 5465, 6096, 467, 829, 4538, 5075, 4494, 3203, 3487, 11679, 11171, 11851, 5378, 9564, 12642, 9450, 11859, 9843, 8986, 72...
[ 0.05708961561322212, -5.867303848266602, 0.03001466765999794, 0.30250483751296997, 0.45875877141952515, -0.04125678166747093, 0.3952864706516266, -0.08710160106420517, 0.12531256675720215, -3.862858772277832, 0.04017629846930504, 0.009598132222890854, -0.07860400527715683, 0.18809832632541...
504
20.16
bilibili_data_1622095_BV1dZ421J7JT_p42_BV1dZ421J7JT_p42_m4-dialogue_0046693
[S1] The Taipei concert tickets were sold out two hours after the tickets were released online. So I decided to try the Tokyo concerts instead. [S2] Sold out. [S1] Sold out. [S2] I see. How much were the tickets? [S1] $200 for floor seat tickets. [S2] Floor seat.
20.56
3.152382
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p42_BV1dZ421J7JT_p42_m4-dialogue_0046693.mp3
[ 9745, 6192, 10892, 12442, 3648, 6519, 11934, 8216, 2944, 7604, 9721, 11708, 11130, 12137, 11667, 6242, 5225, 8360, 9655, 12653, 5015, 4559, 4682, 11371, 9187, 7161, 4538, 11573, 11708, 12077, 11787, 2832, 4066, 10653, 2241, 7250, 8592, 7660, ...
[ 0.41633573174476624, -5.784543514251709, 0.13472647964954376, 0.6330011487007141, 0.4261128008365631, -0.2232176959514618, 0.36406686902046204, -0.01129704900085926, 0.33538272976875305, -3.904693841934204, 0.34042543172836304, -0.19113634526729584, -0.09721614420413971, 0.0946843698620796...
514
20.56
bilibili_data_1622095_BV1dZ421J7JT_p42_BV1dZ421J7JT_p42_m4-dialogue_0046694
[S1] That's a decent price. I paid $300 for Justin Bieber's concert last year. Will you stay a few more days in Tokyo? [S2] I might. Maybe I'll take two days off work and spend more time in Tokyo. [S1] You should. Tokyo is a fun city.
18.48
2.92069
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p42_BV1dZ421J7JT_p42_m4-dialogue_0046694.mp3
[ 10075, 6762, 3802, 11058, 10992, 11883, 5272, 856, 1490, 10750, 4375, 5122, 8745, 7936, 7629, 7365, 10117, 5082, 8160, 8137, 3054, 4015, 6149, 4688, 7259, 10666, 4537, 4338, 2958, 6071, 4981, 12773, 12097, 12053, 6649, 6041, 3026, 5569, 685...
[ 0.23776791989803314, -5.888090133666992, 0.07817849516868591, 0.36539360880851746, 0.42844435572624207, -0.0454690083861351, 0.3508727550506592, -0.04376152902841568, 0.21039530634880066, -3.8664939403533936, 0.1419040560722351, 0.08304032683372498, -0.03757263720035553, 0.1025083586573600...
462
18.48
bilibili_data_1622095_BV1dZ421J7JT_p44_BV1dZ421J7JT_p44_m4-dialogue_0046696
[S1] 1245 PM. However, there are no more standard seats available on that train. The next train after that leaves at 2:00 PM. [S2] How long is the train ride? [S1] Three hours and 45 minutes. [S2] Okay. Two seats for 2:00 PM. [S1] Including the 5% tax, your total comes to €189. [S2] Tax. Here is my credit card.
26.56
3.214309
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p44_BV1dZ421J7JT_p44_m4-dialogue_0046696.mp3
[ 11665, 4465, 243, 11566, 11120, 10992, 8355, 3811, 9081, 8560, 307, 10420, 8587, 8835, 12172, 12036, 3945, 5424, 4023, 3511, 6447, 12783, 4308, 6312, 10930, 5034, 12102, 7493, 3974, 1894, 4973, 4980, 1638, 8347, 12570, 9323, 7266, 12058, 98...
[ 0.02125840075314045, -5.6887617111206055, 0.15180981159210205, 0.34655505418777466, 0.4352559447288513, -0.09629721939563751, 0.4049154222011566, -0.09532783180475235, -0.054389066994190216, -3.862409830093384, 0.15052363276481628, 0.006735582835972309, -0.15448909997940063, 0.180635064840...
664
26.56
bilibili_data_1622095_BV1dZ421J7JT_p45_BV1dZ421J7JT_p45_m4-dialogue_0970269
[S1] Keychain. [S2] Magnet. [S1] I think they are poorly made, and they aren't even made in France. [S2] Good point. What about getting some French desserts? [S1] I would love to bring some macarons home, but they would dry out quickly.
18
2.88665
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p45_BV1dZ421J7JT_p45_m4-dialogue_0970269.mp3
[ 8986, 4392, 10132, 12675, 9538, 11717, 11717, 12055, 11649, 11736, 7634, 11741, 8686, 9190, 7638, 7621, 6606, 4550, 7246, 6213, 6222, 4267, 3763, 12506, 12633, 9825, 11473, 11986, 12417, 11924, 7523, 9883, 4645, 6732, 9908, 6356, 12555, 7261,...
[ 0.35942161083221436, -5.82248592376709, -0.017389735206961632, 0.24313777685165405, 0.4985719621181488, -0.10233939439058304, 0.3301963806152344, -0.011403962969779968, 0.12970878183841705, -3.9058926105499268, 0.21137635409832, 0.04903857782483101, -0.27378496527671814, 0.1249155551195144...
450
18
bilibili_data_1622095_BV1dZ421J7JT_p5_BV1dZ421J7JT_p5_m4-dialogue_0201314
[S1] Wow, you're completely soaked. [S2] Soak. [S1] It's terrible out there. I had an umbrella, but it was useless against the heavy rain.
12.84
2.85592
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p5_BV1dZ421J7JT_p5_m4-dialogue_0201314.mp3
[ 3219, 12571, 11794, 6193, 11979, 9363, 8357, 5732, 3875, 5859, 5355, 8041, 11048, 11625, 11754, 11700, 8114, 10610, 11059, 11571, 11006, 7871, 8319, 5230, 781, 6377, 11987, 11998, 8821, 4765, 12635, 9824, 6192, 12427, 3748, 10051, 10458, 7842...
[ 0.5184734463691711, -5.863399505615234, 0.006453359499573708, 0.5012847185134888, 0.38151106238365173, -0.15754257142543793, 0.22642222046852112, -0.06016193702816963, 0.2120511382818222, -3.9025189876556396, 0.31124943494796753, -0.08405259251594543, -0.01980053447186947, 0.11719952523708...
321
12.84
bilibili_data_1622095_BV1dZ421J7JT_p5_BV1dZ421J7JT_p5_m4-dialogue_0201315
[S1] I can see that. Do you have a change of clothes? [S2] Change of- [S1] ... clothes. [S2] No, I don't. [S1] I have some extra clothes in my office. I started keeping shirts and pants in the office after I got soaked in a rainstorm last summer. [S2] Rainstorm.
22.96
3.24977
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p5_BV1dZ421J7JT_p5_m4-dialogue_0201315.mp3
[ 9635, 3761, 8979, 9258, 8753, 11519, 5502, 9726, 12767, 2315, 11418, 5026, 7599, 4758, 5645, 3334, 2976, 6041, 5528, 6040, 3017, 2944, 3924, 11603, 11531, 11973, 7494, 3203, 3481, 4526, 4972, 4916, 4916, 7540, 2420, 7248, 3120, 7745, 6939, ...
[ 0.324383407831192, -5.698430061340332, 0.12160360813140869, 0.4939461052417755, 0.36038386821746826, -0.2134370505809784, 0.1427895575761795, -0.0013874166179448366, 0.15958808362483978, -3.902622938156128, 0.34063902497291565, -0.1393594592809677, 0.10836968570947647, 0.13442519307136536,...
574
22.96
bilibili_data_1622095_BV1dZ421J7JT_p6_BV1dZ421J7JT_p6_m4-dialogue_0046697
[S1] That's a good idea. [S2] But we'll have to wait about 25 minutes to get our food. [S1] That's okay. I can wait. [S2] What do you feel like eating? [S1] How about Thai food? [S2] Sure. There's a Thai restaurant on my food delivery app. It'll take about 30 minutes to get here.
24.48
3.246412
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p6_BV1dZ421J7JT_p6_m4-dialogue_0046697.mp3
[ 9562, 3345, 4527, 7095, 12218, 12155, 8785, 3041, 2960, 8563, 9508, 3084, 8218, 8737, 11994, 6250, 11603, 3558, 3431, 11110, 12564, 2956, 11574, 8564, 12147, 10052, 1541, 8384, 6611, 10125, 11521, 11465, 11483, 6447, 878, 1908, 1318, 4315, ...
[ 0.4091816842556, -5.730954170227051, -0.05262620002031326, 0.23275138437747955, 0.5333104729652405, 0.03487967699766159, 0.4818568527698517, -0.03687272220849991, 0.15244616568088531, -4.021839618682861, 0.24394115805625916, 0.12969526648521423, -0.2536892294883728, 0.05649888515472412, ...
612
24.48
bilibili_data_1622095_BV1dZ421J7JT_p7_BV1dZ421J7JT_p7_m4-dialogue_0656691
[S1] Hi. How are you? [S2] I'm good, thank you. How's your day going? [S1] It's okay. Two more hours until I get off my shift. How would you like to pay today? Cash or card? Shift. [S2] Credit card. Thanks. [S1] Would you like any bags? [S2] Bag. [S1] No, thank you. I brought my own cloth bag.
28.32
2.891843
24,000
audio/en/bilibili_data_1622095_BV1dZ421J7JT_p7_BV1dZ421J7JT_p7_m4-dialogue_0656691.mp3
[ 7201, 2593, 7412, 1470, 3965, 3965, 3958, 4470, 4534, 4023, 7151, 7654, 9678, 9095, 4950, 7604, 5043, 9570, 9891, 7275, 11987, 12642, 7202, 4145, 9996, 9730, 4617, 9771, 12634, 4186, 2210, 4323, 7671, 4541, 8445, 5222, 5879, 11510, 9525, ...
[ 0.2410508394241333, -5.774068355560303, 0.19051212072372437, 0.03280433639883995, 0.5630646347999573, -0.05821561813354492, 0.2995927929878235, -0.055725548416376114, 0.07685129344463348, -3.8456215858459473, 0.21566306054592133, -0.07999109476804733, -0.1547020971775055, 0.224105626344680...
708
28.32
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512258
[S1] Yeah, uh, first of all, I think that you did a great job in, uh, describing the overall paper and I have almost, uh, no, you know, I have almost nothing to- [S2] No complaints. [S1] Yeah, no complaints regarding that. And, uh, maybe one point regarding the overall, uh, overall point of the paper. And yeah, it, as it's seen from the title, uh, Fourier Convolution
29.52
3.29129
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512258.mp3
[ 6744, 1365, 11588, 11684, 12141, 4023, 4838, 12542, 4030, 6525, 6590, 6006, 5502, 5494, 6372, 5864, 6314, 6764, 5739, 4201, 5561, 485, 8486, 3352, 3137, 2976, 10402, 2730, 2673, 3247, 3247, 5814, 10426, 10264, 5257, 6374, 8958, 6910, 3950, ...
[ 1.0480505228042603, -5.554858684539795, -0.13236895203590393, -0.05074971914291382, 0.5221347808837891, -0.1340174525976181, 0.4218723475933075, 0.016500283032655716, -0.08695684373378754, -3.7462968826293945, -0.09876541793346405, -0.039127059280872345, -0.09129263460636139, 0.15987136960...
738
29.52
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512259
[S1] So these would be the, the Fourier convolutions. Was it already in the form that we see in the paper with the two strands of information- [S2] Yeah. [S1] ... like the global and the local, or did you have to shake things up? [S2] Uh, no, this, uh, this, uh, the right part of this, uh, feature is, uh, reflect the original, uh, form of this fast Fourier convolution as it was proposed by the authors. [S1] Cool. And, um, did it work out of the box?
28.08
3.261655
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512259.mp3
[ 6314, 6713, 1241, 3408, 476, 8286, 2688, 11682, 11585, 6475, 2944, 5348, 10540, 6209, 6184, 9130, 12171, 12173, 4612, 2833, 12198, 11654, 12102, 9094, 3975, 4486, 9094, 1932, 1627, 9745, 8353, 9290, 5754, 2816, 9566, 12037, 10054, 4486, 506...
[ 0.8133490085601807, -5.471554279327393, 0.0044728834182024, -0.12882287800312042, 0.36376431584358215, -0.048430636525154114, 0.3553517162799835, -0.0011993078514933586, -0.17980408668518066, -3.82486629486084, 0.11994793266057968, 0.19646614789962769, -0.005011136643588543, 0.184182405471...
702
28.08
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512262
[S1] Is your, is your algorithm deterministic? [S2] Yeah. [S1] If, if I give it the same input and the same mask, so if I, and this is, is this correct, the, the cleanup.pictures app, that is really your small model that runs here? [S2] No, the, this is the large, large model. [S1] Oh, this is the big model already. [S2] Yeah. [S1] Okay. So here, you know, I've, I've taken this, but what happens, have you ever tried just masking the whole picture? What's kind of like the default output?
29.52
3.140312
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512262.mp3
[ 2601, 11299, 12693, 12107, 6404, 449, 2945, 1346, 12548, 10645, 10340, 7836, 10985, 8017, 8017, 5515, 853, 4708, 11874, 11891, 9370, 8810, 9888, 11493, 12749, 1418, 2888, 9538, 10473, 2899, 4315, 9843, 11314, 9150, 6590, 8511, 7973, 9557, 5...
[ 0.45274603366851807, -5.607014179229736, 0.07395307719707489, -0.1573854237794876, 0.47170281410217285, 0.07110920548439026, 0.2937261760234833, 0.024848591536283493, -0.11464497447013855, -3.8267722129821777, 0.20130087435245514, 0.13974446058273315, 0.10083893686532974, 0.216416701674461...
738
29.52
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512263
[S1] commonly used, uh, this is a commonly used, uh, loss in image-to-image tasks. [S2] Mm-hmm. [S1] It helps to stabilize, stabilize training. [S2] So you, the idea is that the discriminator bases its decisions on features which are perceptually meaningful. So very, very similar to the perceptive loss that you have up here.
23.88
3.12903
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512263.mp3
[ 6184, 3681, 12562, 4588, 10492, 7912, 10778, 10412, 11060, 11573, 8717, 8846, 8845, 3285, 5900, 11669, 11588, 12103, 6916, 7314, 12163, 10577, 10394, 9047, 5074, 1240, 4305, 8224, 5145, 5952, 721, 4693, 7390, 3934, 3926, 3934, 3934, 3934, 3...
[ 0.6646831631660461, -5.6295247077941895, -0.04336542263627052, -0.115854412317276, 0.43483132123947144, -0.08391512185335159, 0.3057698607444763, -0.01525088306516409, -0.10444076359272003, -3.792548418045044, 0.13076439499855042, 0.022470546886324883, 0.10815896093845367, 0.08512721955776...
597
23.88
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512264
[S1] Which seven losses go into my final loss, right? Of the 50 possible losses that I could do, do I try them all or, or they're some, some guidelines? [S2] Actually, I think, uh, all of these losses, except for high receptive field, perceptual loss are pretty common. [S1] Mm-hmm. [S2] And they all,
22.44
2.98572
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512264.mp3
[ 3689, 5688, 2070, 557, 5357, 11731, 10068, 10264, 12162, 12682, 12754, 10193, 4457, 11299, 9779, 5202, 5035, 8169, 8160, 8080, 8677, 11766, 12269, 2779, 2925, 10973, 11283, 11342, 8847, 11407, 7252, 9506, 9825, 6219, 2709, 2707, 5468, 7806, ...
[ 0.6418371200561523, -5.770888328552246, -0.10892504453659058, 0.017586059868335724, 0.5740107297897339, 0.0906253457069397, 0.40141600370407104, 0.024585023522377014, 0.11428800225257874, -3.6953394412994385, 0.07873241603374481, -0.08241575956344604, 0.03422131761908531, 0.176217466592788...
561
22.44
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512266
[S1] ... reconstruct something that we can reconstruct, so we need some loss for reconstruction. And ad-ad-in loss is too restrictive, so we need something that works on features. [S2] Yeah. [S1] Uh, but we worked a lot on... We made a hyper-parameter search, of course, and we changed our
23.56
2.837151
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512266.mp3
[ 5672, 9738, 6758, 4380, 6556, 6246, 11994, 5558, 11509, 8789, 4010, 5592, 6785, 3011, 11629, 11186, 10739, 9006, 11427, 6554, 2984, 2896, 3034, 11198, 8574, 6966, 9309, 6433, 3808, 5862, 8213, 3171, 3171, 802, 6959, 6959, 7242, 8272, 3913, ...
[ 0.8717641830444336, -5.678571701049805, -0.01919393427670002, 0.09066331386566162, 0.4133071005344391, -0.15630970895290375, 0.7064168453216553, 0.017409419640898705, 0.15350086987018585, -3.873483896255493, 0.1774662733078003, -0.36586853861808777, -0.45246386528015137, 0.1857359111309051...
589
23.56
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512267
[S1] Yeah, yeah, because it is natural for the segmentation model to focus more on boundaries of objects instead of their textures. [S2] Yeah. [S1] And, uh, in case of in-painting, good texture can be, uh, learned using only discriminator. [S2] Mm-hmm. [S1] Because there is a lot of freedom in how we can generate fine-grained textures and there is no need to put some, any supervision on that part of, uh, the image. [S2] But the-
28.48
3.305982
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512267.mp3
[ 6817, 1028, 4821, 12181, 11628, 9071, 7022, 6244, 9386, 12338, 8361, 561, 7240, 1612, 12173, 11693, 7095, 6892, 3683, 6754, 6126, 7365, 10858, 7584, 11062, 10483, 10547, 10613, 5964, 2905, 3464, 385, 5918, 11109, 9485, 12611, 12165, 7373, 1...
[ 0.8822304010391235, -5.502527236938477, 0.08922626823186874, -0.1707010567188263, 0.5550698637962341, -0.25809672474861145, 0.3391384780406952, 0.07206784188747406, -0.15710915625095367, -3.795036792755127, 0.19228722155094147, -0.16548795998096466, -0.13972900807857513, -0.030416954308748...
712
28.48
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512269
[S1] ... cool. [S2] Unfortunately, I'm not familiar with, uh, word-shaping from reinforcement learning. [S1] Yeah. [S2] Uh, but our idea here was that basically we have two losses here. The first one is discriminator or the serial and which focused more on fine-grained details and the second is perceptual loss which focused more on, uh, global text, global structures. Yeah. [S1] Mm.
25.04
3.286873
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512269.mp3
[ 6364, 9676, 7333, 11197, 9148, 6443, 5279, 2710, 30, 3350, 6811, 9379, 9811, 8291, 9883, 6170, 8962, 7461, 9907, 6250, 8853, 10077, 9965, 10166, 3965, 6526, 11293, 9360, 5426, 3451, 8037, 9537, 6984, 6361, 5773, 3923, 12215, 8125, 11772, ...
[ 0.9573473334312439, -5.4201555252075195, -0.14909744262695312, -0.02731487713754177, 0.5601029992103577, -0.24507641792297363, 0.36621448397636414, 0.037725165486335754, -0.18749497830867767, -3.7858099937438965, 0.20392285287380219, -0.10044527798891068, -0.02744149975478649, -0.016928760...
626
25.04
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512270
[S1] This up here is really beautiful, right? [S2] Yeah. [S1] What, what picture could I take such that it is absolute garbage? [S2] Yeah, actually, lots of, lots of images, uh, will be, uh, processed bad with our model. Yeah. [S1] I mean, the part of, of course, I can give it a picture that is, you know, very dissimilar to the training data set. [S2] Yeah. [S1] But let's say I actually had a training data set. What, what would be the worst domain or the worst kind of picture? [S2] Yeah, um,
29.92
3.147766
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512270.mp3
[ 9769, 1065, 2963, 11241, 6499, 5528, 786, 3959, 317, 4138, 8497, 7522, 12676, 11163, 2998, 258, 1445, 4374, 448, 5058, 733, 2850, 354, 339, 11590, 11653, 6237, 2771, 4879, 7493, 1057, 6497, 12549, 3863, 2830, 4683, 982, 1310, 3944, 6000, ...
[ 0.5869483351707458, -5.55074405670166, 0.03991246595978737, -0.20395343005657196, 0.5455595254898071, 0.05763059854507446, 0.2600492238998413, 0.041432637721300125, -0.07712685316801071, -3.862276554107666, 0.10211002081632614, 0.11085718125104904, 0.05261065065860748, 0.16811923682689667,...
748
29.92
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512271
[S1] I think- [S2] Yeah. Please, go. [S1] ... it cannot, uh, recreate half of human on something. [S2] Mm-hmm. Yeah, our model focuses mostly on background due to how it was trained and- [S1] Yeah. [S2] ... yeah, it cannot, uh, recover foreground objects really well. [S1] It cannot, uh, do something that requires it to actually know everything about walls and not just take it from picture it sees.
29
2.911455
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512271.mp3
[ 9064, 4383, 3101, 9014, 10681, 12267, 6491, 867, 3310, 6489, 11241, 10648, 12049, 6283, 6802, 6298, 5778, 3931, 6363, 6827, 8803, 9251, 9235, 8819, 6859, 6419, 3365, 6373, 5659, 6689, 7410, 9376, 2601, 4634, 9440, 6187, 4251, 3603, 4130, ...
[ 0.8478728532791138, -5.590081691741943, -0.10900361835956573, -0.004130263347178698, 0.7950568795204163, -0.009513458237051964, 0.6591506600379944, 0.052171241492033005, 0.13233549892902374, -3.8065826892852783, -0.07626237720251083, -0.32012465596199036, -0.2637772560119629, 0.11661227792...
725
29
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512272
[S1] Yeah, I have, I have, um, another question and to the Fourier convolution. So, here we have global information and, and local information, right? As, as sort of two different things. You mention in the paper that other models that have more global information or access to wider information could also work, such as a vision transformer or something like this. [S2] Yeah. [S1] My question is, is there an in-between between
28.92
3.233119
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512272.mp3
[ 8745, 6754, 8725, 9924, 11147, 10739, 6134, 7211, 8702, 11702, 7523, 11707, 8996, 3056, 3747, 8753, 8502, 10021, 9643, 11179, 5979, 1002, 11566, 8175, 4015, 6575, 3503, 4014, 4023, 4462, 5661, 6255, 6751, 9836, 12450, 12395, 562, 8229, 4335...
[ 0.5889399647712708, -5.598641872406006, 0.07529449462890625, -0.2216443121433258, 0.42385974526405334, 0.1623639464378357, 0.2719716727733612, 0.03058655932545662, -0.10139721632003784, -3.7776384353637695, 0.20698139071464539, 0.16661085188388824, 0.14043591916561127, 0.16133876144886017,...
723
28.92
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512273
[S1] Yeah, it doesn't scale up, uh, perfectly, but yeah, it, it, it better than, uh, fully convolution architectures. [S2] Cool. Um, yeah, so where do you think, I mean, m- maybe you don't wanna disclose necessarily, but, but what, what do you, what is the plan for the future? [S1] [LAUGHS]
16.08
3.164055
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512273.mp3
[ 3713, 6990, 9654, 4007, 4998, 9941, 6208, 8157, 10617, 11067, 3857, 401, 8342, 10839, 11787, 8328, 1361, 3809, 3280, 4312, 12443, 5099, 12197, 11675, 11668, 11677, 7983, 10487, 7798, 7774, 2758, 3933, 9900, 11893, 6078, 8062, 5879, 8298, 62...
[ 0.8323982954025269, -5.626744747161865, -0.13656839728355408, -0.025897255167365074, 0.3949830234050751, 0.039904262870550156, 0.30675962567329407, -0.04633191600441933, -0.07699225842952728, -3.8217389583587646, 0.06123282015323639, 0.1653996706008911, 0.05427901819348335, 0.0742170661687...
402
16.08
bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512275
[S1] Yeah. I mean, I, I was almost, I was expecting you to say, we're not happy with our loss. We want more. We want, like, more components to make it be- but it's, it's, I, I think it's pretty cool that, uh, the goal is also to make a system that's kind of as good but simpler. [S2] Yeah. [S1] Um, I think that'll make it also much more accessible. [S2] Cool. [S1] Yeah, um, uh, Roman, Elisa, sorry, Lisa, is that correct? [S2] Yeah.
25.76
3.169234
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p160_BV17h4y1j7aZ_p160_m4-dialogue_0512275.mp3
[ 3229, 8737, 8810, 8345, 9322, 8730, 7382, 12165, 10717, 12207, 1527, 1460, 1846, 4908, 7401, 12386, 12338, 8865, 9825, 6249, 8796, 6889, 5234, 12398, 12774, 3877, 553, 3683, 12196, 8564, 5156, 12693, 12098, 11668, 8206, 8387, 7329, 8753, 11...
[ 0.5494576096534729, -5.7130255699157715, 0.07969602942466736, -0.16266003251075745, 0.3571113049983978, 0.03818568214774132, 0.32909300923347473, 0.04082368314266205, -0.07896967232227325, -3.765490770339966, 0.2673025131225586, 0.25248268246650696, 0.05467677488923073, 0.2800549864768982,...
644
25.76
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758465
[S1] ... through the paper. So, Martin, uh, welcome. Uh, thank you very much for being here. [S2] Hey, I'm happy to be here. [S1] Um, did, was, was it a sort of a good description of what I said so far about Player of Games? [S2] Oh, yes, very, very, very much so.
16.92
3.153081
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758465.mp3
[ 7259, 4325, 7136, 6560, 3360, 885, 3866, 2910, 7758, 5189, 5189, 2694, 1685, 9900, 6705, 2050, 403, 4838, 8296, 7665, 10165, 5085, 10182, 7308, 8296, 3563, 5999, 396, 1411, 4820, 4467, 8873, 5716, 7262, 12404, 7269, 7469, 9326, 5229, 9838...
[ 0.7047213912010193, -5.880446910858154, -0.12534913420677185, 0.06770626455545425, 0.6807475686073303, 0.015060819685459137, 0.27985936403274536, -0.11536175012588501, -0.14039714634418488, -3.7663636207580566, 0.16842107474803925, 0.03230222687125206, -0.0295545756816864, 0.05048104003071...
423
16.92
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758466
[S1] ... that, uh, we will see in this paper, which is it introduced this notion of, uh, of local search in, uh, in poker and these value functions. [S2] Mm-hmm. [S1] And Play of Games is really just putting together AlphaZero in deep step into a single big unified, uh, algorithm. So the, the, the, let's maybe start with the, with the component that, that you, that you just talked about, which is, uh, value function.
25.28
2.834349
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758466.mp3
[ 12683, 9167, 6071, 9254, 8397, 448, 18, 2754, 9591, 12783, 7454, 10841, 6241, 11712, 6502, 8551, 8423, 5790, 3741, 3307, 12614, 7429, 3307, 8511, 8413, 7593, 8080, 3969, 12102, 7429, 9956, 6166, 5514, 10142, 6048, 6208, 8808, 4021, 7653, ...
[ 0.1734119951725006, -5.927214622497559, 0.06734319031238556, -0.43516165018081665, 0.3878454566001892, -0.15108156204223633, 0.5251206159591675, 0.027119005098938942, -0.2998664081096649, -4.0269646644592285, 0.20864038169384003, -0.27901792526245117, -0.21467077732086182, 0.28439769148826...
632
25.28
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758468
[S1] It's, it makes, uh, little to no sense to say what's a value, what's a value of a sub-game or sub-problem in a poker where, where I hold a pair of, of aces. That's pretty much ill-defined, uh, ill-defined sub-game. [S2] Mm-hmm. [S1] But what we, what you need to do is given a, given a public, uh, state, which is, as you say, I come to a table, I see everything that, that I could have observed as a public observer. So that, that's, that, uh, that's basically my state. But given this state, given this observation,
29.92
2.913961
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758468.mp3
[ 12516, 12742, 9737, 8857, 8449, 2904, 3288, 11882, 3218, 6748, 6762, 9812, 8925, 9414, 7688, 8216, 6761, 522, 5214, 8229, 11254, 10219, 12756, 9377, 11875, 4586, 5984, 2881, 2972, 11045, 11046, 10973, 10534, 5926, 5934, 1244, 3741, 6301, 37...
[ 0.1489231437444687, -5.878063201904297, 0.1515377014875412, -0.5165203809738159, 0.5444222688674927, -0.10974038392305374, 0.49046510457992554, -0.018987225368618965, -0.36709660291671753, -4.024369716644287, 0.2543775141239166, -0.30070891976356506, -0.1778961420059204, 0.3222185373306274...
748
29.92
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758470
[S1] Yeah, I see. Okay. And you have this network and you train it, uh, in some way via, via self play. And now we get to the part where you generalize this search procedure, right? And- [S2] Mm-hmm. [S1] ... let me see. Oh, this is here. [S2] Yes. [S1] So this search procedure, as we said, in, in alpha, uh, again, in alpha zero, you have something like, you
22.2
3.346411
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758470.mp3
[ 8232, 6201, 4227, 9164, 10732, 4022, 1641, 4708, 8802, 12394, 11827, 11289, 9975, 3965, 10126, 4008, 2976, 346, 2382, 4994, 4292, 3107, 12395, 10993, 9265, 8330, 12691, 7476, 12526, 7278, 9838, 3180, 4653, 1573, 3794, 9362, 11966, 7982, 369...
[ 0.8649237155914307, -5.67995548248291, -0.06078379973769188, 0.13706083595752716, 0.4224010407924652, 0.00014074333012104034, 0.11433828622102737, -0.0965011864900589, -0.10217013210058212, -3.73565411567688, 0.23447869718074799, 0.0815662369132042, 0.07701841741800308, 0.04024664685130119...
555
22.2
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758471
[S1] You would, you would sort of, you would make many iterations, let's say 50 iterations or something like this. In every iteration, you'd go down the tree and you find a node that you haven't expanded yet and you'd expand that node, right? [S2] [LAUGHS] [S1] In, in, in player of games, this is quite a bit more intricate, right? Uh, as, as we also have
21.56
3.133394
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758471.mp3
[ 6697, 7699, 9300, 9995, 11715, 10653, 2646, 2804, 10540, 10667, 9227, 8129, 4246, 8753, 6805, 11715, 8405, 5294, 10668, 7243, 5522, 8152, 5548, 8059, 3845, 8157, 7981, 8689, 4393, 8225, 11865, 11378, 12441, 8761, 8731, 12548, 11212, 10453, ...
[ 0.6761842966079712, -5.74947452545166, -0.0790044292807579, 0.1153910681605339, 0.46363377571105957, 0.023331545293331146, 0.1790057122707367, -0.1049560010433197, -0.19399867951869965, -3.7593936920166016, 0.2696816921234131, 0.10885990411043167, -0.02529122494161129, 0.03421109914779663,...
539
21.56
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758472
[S1] and improve the policy of both players, right? And it just does this for many, many iterations. It improves here, here, here, everywhere in the game tree until the whole game tree is, uh, approximately optimal. [S2] [LAUGHS] [S1] And the, the biggest game that has been solved so far, if you describe this in the paper, is Limit, Limit Heads Up Hold 'Em, is that correct? [S2] Yes. [S1] Fix Limit Hold 'Em. Yeah. [S2] Exactly. [S1] That's, that's actually a solved game.
26.4
3.335737
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758472.mp3
[ 6176, 6761, 10862, 11766, 9199, 7151, 7159, 7151, 6966, 8213, 8271, 8847, 9286, 5633, 8642, 4379, 10794, 7849, 9777, 7721, 11550, 12749, 9734, 8237, 5672, 10538, 9714, 1010, 2900, 2900, 10396, 7822, 6, 646, 795, 5616, 8688, 8177, 5624, 35...
[ 0.7156736850738525, -5.587573528289795, 0.002982513280585408, 0.007445030380040407, 0.5356330275535583, 0.024717509746551514, 0.13506071269512177, -0.08163948357105255, -0.12787571549415588, -3.76267147064209, 0.19723601639270782, 0.08683522790670395, 0.12096110731363297, -0.00801405124366...
660
26.4
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758475
[S1] Get the entire tree in order again. So you expand the new node and then you have to do the whole update of the whole tree for a bunch of iterations before you can expand another one, such that everything, like, stays consistent. [S2] Mm-hmm. [S1] Yeah. Okay. That's, I mean, this, this, it gives a bit of an impression of why this is, uh, much more, much more complex, right?
21.88
3.239764
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758475.mp3
[ 5737, 11346, 12251, 11251, 12781, 11792, 8337, 8793, 3609, 5322, 10685, 6444, 6516, 7340, 11863, 9240, 5248, 6633, 9597, 5565, 12130, 12099, 11685, 5469, 772, 5737, 11202, 8680, 8045, 2891, 11604, 12484, 9924, 7493, 9199, 7382, 8909, 3767, ...
[ 0.7681631445884705, -5.703742980957031, -0.10231509804725647, -0.03337116912007332, 0.5075267553329468, 0.10759980976581573, 0.26688358187675476, -0.06113842502236366, -0.13043507933616638, -3.724999189376831, 0.12453549355268478, 0.12299969047307968, 0.13823114335536957, 0.023008679971098...
547
21.88
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758476
[S1] Mm-hmm. So you collect the train. It's similar to Alpha Zero, as you say. You collect the training set as you go. So the training set for the next iteration is whatever the network had to do during this iteration. So it's not just a random sample of states. [S2] Yeah. [S1] And you train in the same manner as Alpha Zero. You train to predict your own future outputs. Is that approximate? So if, uh, let's, let's distinguish. If
28.2
3.279196
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758476.mp3
[ 8274, 11031, 8852, 8918, 4821, 6803, 7332, 9810, 12338, 12322, 7769, 6690, 9426, 10131, 4717, 4721, 4242, 3552, 3044, 10550, 4300, 12164, 10510, 11427, 10216, 5493, 10990, 2709, 5982, 11764, 11243, 7460, 10851, 9699, 4769, 8320, 2961, 3281, ...
[ 0.6245024800300598, -5.537537097930908, 0.028564058244228363, -0.09301828593015671, 0.5448161959648132, -0.008191517554223537, 0.2078540027141571, -0.03606230765581131, -0.1547589749097824, -3.8021130561828613, 0.25089919567108154, 0.1390068233013153, 0.14239872992038727, 0.075375877320766...
705
28.2
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758477
[S1] ... like one or two or three steps in the future, you actually win or lose the game, you can train on your reward of the game. But Alpha Zero also, if it doesn't win or lose the game in the next step or so, it tries to predict its own output. So it, it tries to improve that way using TD Lambda. Um, and you here have a, a TD1, right? Um, so your, your targets, what do you target? What do you- [S2] Yeah. [S1] ... give the network as labels?
28.2
3.442012
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758477.mp3
[ 9377, 12059, 11995, 5981, 8183, 5094, 7314, 12587, 3747, 2733, 2739, 7982, 11188, 8436, 11060, 12309, 8022, 10550, 5534, 843, 5225, 11729, 8684, 10443, 2693, 8999, 5495, 3028, 4962, 5984, 946, 796, 2965, 11979, 12677, 6625, 10729, 6544, 277...
[ 0.7936699986457825, -5.51356840133667, 0.056436315178871155, -0.030230702832341194, 0.5395909547805786, 0.012238852679729462, 0.10403571277856827, -0.025241291150450706, -0.1563931703567505, -3.8045663833618164, 0.25386330485343933, 0.17724861204624176, 0.07260702550411224, 0.0031833569519...
705
28.2
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758479
[S1] ... small solver where we also substitute the full solver with a small search tree. [S2] Mm-hmm. [S1] So rather than fully solving a game, we use the same method to basically do a search. [S2] Mm-hmm. [S1] And the outcome of the search, basically a small solver, is what the, the, what is the target.
17.8
2.924421
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758479.mp3
[ 5792, 11883, 12618, 9777, 6163, 4452, 6496, 12698, 6552, 3786, 8277, 10852, 8439, 10862, 10420, 8371, 7955, 8942, 5544, 5536, 3361, 2960, 8055, 8439, 5815, 8415, 5269, 676, 3374, 6007, 8084, 5460, 3675, 3749, 2867, 11765, 9141, 2837, 683, ...
[ 0.05865269526839256, -6.124532222747803, 0.05608291178941727, -0.412931352853775, 0.4433797299861908, -0.14908020198345184, 0.5093839764595032, 0.05766148120164871, -0.3578290343284607, -3.9921648502349854, 0.02870693802833557, -0.27134159207344055, -0.20437295734882355, 0.2972454130649566...
445
17.8
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758480
[S1] Maybe this has battery again. So during the inference, you, you, you make, you do these queries, you store them in this, in this buffer, and these now act as the root nodes for yet another search, which is exactly the same as the previous search, right? And so you, you sort of rely on the fact that this search procedure can give you a better output than the neural network itself, right? [S2] Yes. Yes. [S1] So, right, the, the, the query here
27.56
3.140431
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758480.mp3
[ 9867, 4763, 1053, 6382, 10142, 6861, 5800, 4437, 7436, 3785, 10167, 1818, 2841, 3984, 4925, 2230, 840, 3625, 8427, 2487, 7157, 12286, 8971, 3375, 787, 331, 4878, 9924, 2471, 2116, 9947, 7519, 11635, 11496, 8211, 6362, 8810, 4766, 4790, 99...
[ 0.7923789024353027, -5.550041198730469, 0.027606667950749397, 0.07218842953443527, 0.48033931851387024, 0.03781472146511078, 0.2419911026954651, -0.05441603809595108, -0.0684705376625061, -3.8365020751953125, 0.09331415593624115, 0.22631357610225677, 0.03045954369008541, 0.0114237563684582...
689
27.56
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758481
[S1] That's- [S2] That's, that's exactly- [S1] One would hope, though, that after a while, you know, if I do this again and again and again, at the end, I wouldn't even have to ask the neural network anymore. I, sorry, I wouldn't even have to do search anymore during inference. Is that something you-
18.68
3.288989
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758481.mp3
[ 5704, 942, 6527, 6078, 11928, 9304, 5378, 914, 6060, 3995, 6873, 5976, 9946, 9270, 4140, 3621, 7291, 4190, 6326, 9404, 4244, 4187, 7260, 5977, 10737, 11555, 9043, 5592, 11612, 9364, 9708, 8164, 6555, 10739, 9131, 9435, 9506, 876, 8923, 64...
[ 0.8008184432983398, -5.787570476531982, -0.0611325278878212, -0.019659439101815224, 0.5070059895515442, 0.10248430073261261, 0.19144339859485626, -0.0783533975481987, 0.0350622832775116, -3.7344932556152344, 0.14348526298999786, 0.05675715208053589, 0.10968799144029617, 0.09621210396289825...
467
18.68
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758483
[S1] Um, yeah, so I think we're, we're, we're almost getting already to the sort of results, uh, weight. Would you, would you maybe summarize the results a little bit? I think if people are super interested, they, they may go into, uh- [S2] Okay. [S1] ... into the paper and into the tables, but maybe you can just summarize a little bit of the results. You, you compared against Alpha Zero in
24.16
3.287836
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758483.mp3
[ 8809, 6248, 11495, 6647, 10678, 8558, 10333, 10806, 3677, 8737, 8353, 9746, 8307, 9992, 9802, 12456, 5145, 6193, 9870, 12547, 10651, 11701, 1982, 4385, 8162, 3032, 11125, 7859, 10341, 3286, 4706, 9762, 9249, 9761, 8237, 6526, 7513, 5992, 86...
[ 0.8647300004959106, -5.66499137878418, -0.02568153478205204, 0.04067225381731987, 0.41001150012016296, 0.006983225233852863, 0.13845588266849518, -0.037692584097385406, -0.07274401932954788, -3.7482709884643555, 0.24121645092964172, 0.21376483142375946, 0.035175975412130356, 0.079560078680...
604
24.16
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758487
[S1] Do, do, do you mean for a Slumbot or for Scotland Yard? Are we, are we still talking about poker? [S2] Oh, sorry. Yeah. Let's, let's talk about poker for a while. So the, the, the player of games here gains, what is this, seven millibit blinds per, per hand. [S1] Yeah. [S2] Over Slumbot.
17.28
3.149203
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758487.mp3
[ 7395, 9889, 1043, 3474, 11595, 11586, 8532, 9042, 11602, 3868, 3549, 7157, 7378, 8684, 11043, 8483, 11484, 8924, 5916, 8996, 6802, 5683, 6197, 6761, 4705, 9266, 6313, 6776, 3669, 1086, 803, 715, 11585, 8156, 10987, 11044, 10980, 11045, 8477...
[ 0.6516265273094177, -5.895227909088135, 0.009245111607015133, -0.08588244020938873, 0.6077544689178467, -0.09641308337450027, 0.36244603991508484, -0.013638182543218136, 0.09601058065891266, -3.8351714611053467, 0.07469741255044937, -0.026080967858433723, -0.14910459518432617, 0.1630379110...
432
17.28
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758489
[S1] Yeah. Okay. So, and here on the right side, you have a Pym Bot, which is, uh, the, a Scotland Yard-specific bot. For people, maybe people don't, does anyone not know what Scotland Yard is? Um, maybe you can describe 10 seconds what Scotland Yard even is as a game. It's somewhere, right? [S2] Yeah, there's a figure maybe, right? [S1] There is this figure, right?
26.56
3.34405
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758489.mp3
[ 8816, 5161, 6276, 10124, 10725, 957, 1259, 4195, 9314, 11883, 12443, 12385, 8315, 10523, 9848, 7393, 9258, 6168, 9306, 9881, 9843, 6728, 9314, 9369, 6688, 9434, 8355, 9288, 7225, 3136, 12411, 6802, 7225, 4641, 9481, 9390, 7989, 3186, 12195,...
[ 0.7935275435447693, -5.572835445404053, 0.12411379814147949, 0.04453504458069801, 0.48048269748687744, -0.10273392498493195, 0.21431051194667816, 0.0018832404166460037, 0.014795036986470222, -3.7506208419799805, 0.18295928835868835, 0.07713478058576584, -0.05267247185111046, 0.053897190839...
664
26.56
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758490
[S1] Yeah, so this is, this would be, I guess the final win rate is here, uh, like at 55% or something like this, and that is with a huge number of, of iterations for, for PIMBot. [S2] Yes, and the player games is using only like four, 400, uh, iterations on our side. [S1] Yeah. [S2] So, yeah, as you can see, as you can see, the regardless of the scale, we, we converge to a better policy.
24.16
3.404915
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758490.mp3
[ 4641, 6264, 4226, 10052, 11204, 11190, 957, 4909, 4331, 4778, 5929, 849, 989, 10870, 8407, 8728, 5632, 8622, 12700, 736, 865, 201, 3421, 12198, 11622, 1362, 3353, 784, 4225, 5673, 3200, 12142, 12125, 3880, 841, 3746, 5302, 11038, 3619, 57...
[ 0.5384827256202698, -5.791027069091797, -0.028330879285931587, -0.2639510929584503, 0.6742650866508484, -0.13610394299030304, 0.3484649062156677, 0.017151426523923874, -0.16222161054611206, -3.872487783432007, -0.06133681908249855, 0.03603262081742287, -0.06012289226055145, 0.1543561071157...
604
24.16
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758491
[S1] ... that we, we can get you an optimal policy. Uh, other methods that are based on MCTS, they, they are not guaranteed to converge even on small games. [S2] Mm-hmm. [S1] So there's, there's, there's, there's also the limit of the, of the fact that these methods are not sound.
16.48
3.088853
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758491.mp3
[ 9654, 11280, 8793, 4361, 3164, 1195, 3423, 9413, 7493, 4486, 6933, 6291, 3684, 3365, 12614, 4747, 11873, 6637, 10023, 7710, 6754, 9562, 12783, 4527, 5712, 12176, 3366, 5990, 6686, 6739, 11517, 3255, 5217, 6313, 8530, 7463, 7781, 3327, 8731,...
[ 0.07886457443237305, -6.1377668380737305, -0.0014662544708698988, -0.36973148584365845, 0.5468883514404297, -0.1706390380859375, 0.5816513299942017, 0.0765233039855957, -0.2312539666891098, -4.021181583404541, 0.061468906700611115, -0.22643156349658966, -0.16595059633255005, 0.272782802581...
412
16.48
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758492
[S1] I see. So I think the easiest to, for us, was, uh, poker. [S2] Mm-hmm. [S1] That, like, people probably can train on a few, few GPUs. Uh, the, the by far the hardest, uh, is, uh, is Go, where we used, uh, a, a lot of, uh, a lot of TPUs, but that was simply because we, we had them, uh, available.
24.8
2.994558
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758492.mp3
[ 5672, 8728, 8812, 11262, 12748, 5009, 5520, 4048, 12678, 7557, 8849, 6370, 9379, 6684, 4134, 4126, 5221, 7197, 9247, 4188, 6257, 4307, 7593, 11216, 3024, 11052, 10986, 10396, 10453, 6291, 8353, 9882, 11930, 12443, 12450, 6257, 9313, 8813, 7...
[ 0.12942062318325043, -5.924623012542725, 0.06003768742084503, -0.3600447475910187, 0.423245906829834, -0.054726727306842804, 0.47788262367248535, 0.016716040670871735, -0.18660074472427368, -3.9063918590545654, 0.14512431621551514, -0.1722191870212555, -0.21258500218391418, 0.3432011306285...
620
24.8
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758494
[S1] You know, no human, no, like, no, not all humans together even will ever match Alpha Zero in, in any of these games, which is crazy. [S2] Yeah, you, you would not win a single game out of a thousand. [S1] Yeah, exactly. Um, do you, you have a bit of a demonstration ready, you told me, of, uh- [S2] Right. [S1] ... of, of, uh, the player of games playing Scotland Yard, so we can kinda see what's going on.
27.16
3.141121
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758494.mp3
[ 7337, 6746, 4619, 10118, 3724, 5030, 4340, 670, 668, 6379, 8753, 7234, 4645, 4708, 6744, 4622, 8718, 10892, 11253, 5427, 7914, 7822, 4374, 10088, 7556, 11142, 10460, 10971, 8853, 10340, 11254, 6871, 3086, 5637, 1610, 4684, 8269, 8685, 10547...
[ 0.8088822364807129, -5.557581424713135, 0.07069861143827438, -0.05184819921851158, 0.518688440322876, 0.015342357568442822, 0.19113001227378845, -0.08992741256952286, -0.05843024700880051, -3.83663272857666, 0.17919795215129852, 0.04671986773610115, 0.09885726124048233, 0.02634111046791076...
679
27.16
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758495
[S1] ... color, details. So right now, I'm in here, and say I want to go to 49, and I want to use taxi to get there. So, yeah, hopefully, like we have been talking for a while, so maybe, maybe it's not, uh, alive, uh, anymore. But, uh, yeah, probably, probably it, uh, it died, uh, it died. [S2] You have scaled to zero, pr- proper engineering. [S1] [LAUGHS] [S2] Nice.
26.52
3.007138
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758495.mp3
[ 5152, 9834, 12194, 11197, 11196, 10549, 7909, 8622, 11053, 8156, 3405, 2571, 8010, 11733, 9671, 8905, 11344, 6636, 8573, 6006, 6916, 10116, 3406, 661, 1749, 7136, 6024, 3520, 3992, 10025, 8930, 9889, 9833, 5152, 3609, 9931, 7330, 3602, 3971...
[ 0.4854830801486969, -5.77764892578125, 0.045557454228401184, -0.1976032555103302, 0.3517671227455139, -0.033793605864048004, 0.39570799469947815, -0.026476843282580376, -0.16026295721530914, -4.013639450073242, 0.1912401020526886, -0.21041753888130188, -0.21932384371757507, 0.2058086246252...
663
26.52
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758496
[S1] Uh, we did. It's, I don't think it's, it's turn on right now, but it's exactly what we tried to do at some, at some point. [S2] Yeah. And did you see, did you observe this? Does the longer they didn't see Mr. X, the more kind of spread out, the more unsure they become? Is that something you can clearly observe or is that something you just feel as a human? [S1] Uh, yes. And it, it, it was actually really, really fun to, to see. [S2] Yeah. [S1] See, see that.
28.44
3.01169
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758496.mp3
[ 3113, 5202, 7205, 9454, 11038, 8476, 7907, 11107, 5725, 803, 3563, 11585, 11971, 522, 5442, 11742, 11222, 7006, 8394, 9170, 7331, 8878, 11733, 9923, 8273, 4057, 3424, 856, 4385, 3968, 4896, 9377, 8949, 8702, 11733, 9795, 8082, 10544, 5794, ...
[ 0.557623565196991, -5.705898284912109, 0.10360894352197647, -0.35713663697242737, 0.547706127166748, 0.022433968260884285, 0.43861085176467896, 0.053009964525699615, -0.06325147300958633, -3.9181275367736816, 0.13343507051467896, -0.05346952751278877, -0.10590612888336182, 0.23541273176670...
711
28.44
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758497
[S1] alleviated this, uh, by sort of going to the latent space state and training everything in latent space. Is this something I could do with player of games? [S2] Uh, no, but that's, uh, that's arguably the limitation, uh, number two. I think the biggest- [S1] Yeah. [S2] ... being, uh, the biggest thing is, uh, right now the, uh, the large, uh, large belief, uh, belief space. [S1] Mm-hmm. [S2] But the sec- second one is we currently need the model of the environment.
28.44
3.378782
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758497.mp3
[ 9257, 5673, 11054, 6647, 5781, 2772, 3212, 5963, 12102, 12549, 12675, 3867, 1003, 12549, 12548, 1951, 4957, 4941, 5005, 7297, 1878, 1831, 7307, 8289, 7937, 3920, 4906, 9762, 9761, 6177, 4115, 1040, 3028, 5023, 1955, 1694, 4008, 4064, 3488, ...
[ 0.48860248923301697, -5.790985584259033, 0.05046960338950157, -0.27827855944633484, 0.6158825755119324, -0.07995403558015823, 0.3500349223613739, -0.036542389541864395, -0.2286900281906128, -3.963975191116333, 0.0959247350692749, -0.07994077354669571, -0.14854788780212402, 0.17556807398796...
711
28.44
bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758498
[S1] So when you expand the search tree, do you need to expand once for every possible, let's say, flop combination there is? [S2] Yes. [S1] Okay. That, that is a lot of combinations, right? [S2] Or you can, or you can substitute, uh, like if you are smart about it, you, you can, again, it's a neural network.
19.92
3.151267
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p162_BV17h4y1j7aZ_p162_m4-dialogue_0758498.mp3
[ 9256, 729, 2952, 10677, 7788, 2637, 2145, 6683, 627, 11693, 12274, 11789, 4292, 12482, 10389, 2694, 7210, 12134, 4885, 10794, 9692, 6120, 8160, 6089, 6249, 10602, 12798, 12726, 12212, 11644, 9252, 9295, 7171, 2752, 3439, 1961, 7977, 6016, 3...
[ 0.49039727449417114, -5.779314994812012, -0.013721353374421597, -0.14435063302516937, 0.5896185636520386, -0.08040274679660797, 0.3361131250858307, 0.015078662894666195, -0.1620132178068161, -3.8541462421417236, 0.1130111813545227, -0.06531602889299393, -0.09758967161178589, 0.147099897265...
498
19.92
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652852
[S1] I'm not a German either, but, but I, I think that, uh, the author was called, was called Noether. Yeah, she- [S2] Yeah, yeah. So you're pronouncing it more properly than him, I think. [S1] [LAUGHS] Maybe. But essentially, could you give us maybe just first an insight, where does the name, because the name is kind of distinct, right? Because there is the Noether Theorem. [S2] Yeah. [S1] Uh, what does the Noether Theorem say in general?
24.72
3.124485
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652852.mp3
[ 3689, 6296, 4654, 6398, 8635, 11875, 10340, 5779, 7988, 10484, 9281, 423, 9591, 6723, 6977, 10114, 4046, 5935, 8028, 8074, 8019, 8228, 10613, 11430, 7949, 9079, 11062, 11134, 12244, 9358, 5472, 11637, 7970, 5443, 5509, 3113, 8041, 6007, 593...
[ 0.7102078795433044, -5.63679838180542, -0.01309791300445795, -0.03228331357240677, 0.5072011947631836, 0.10106344521045685, 0.19802667200565338, -0.023580363020300865, -0.09588833153247833, -3.8076815605163574, 0.18608111143112183, 0.08436817675828934, -0.002597721992060542, -0.00157935579...
618
24.72
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652853
[S1] of the theorem, uh, to, uh, build a new, uh, machine learning model. And the intuition is that in machine learning, symmetries are one of the core ways in which we've improved, uh, data efficiency and, and, and model performance. And so it would be very cool if we could kind of automatically learn some of these symmetries. [S2] Mm-hmm. [S1] Um, but, uh, symmetries are kind of hard to quantify and, and, and get a hold of, uh, computationally. And the intuition is that, um,
27.08
3.268367
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652853.mp3
[ 2664, 11857, 8429, 3117, 3856, 821, 9440, 3160, 4442, 11082, 10611, 6084, 8669, 11685, 7717, 8867, 11937, 9889, 6693, 6702, 7724, 6189, 738, 3680, 4253, 6014, 3902, 6014, 8126, 3510, 10030, 6282, 5777, 11818, 8361, 5153, 9419, 9960, 3120, ...
[ 0.6135091185569763, -5.736812591552734, 0.0519380047917366, -0.19224300980567932, 0.7219439744949341, -0.10584337264299393, 0.2500820755958557, 0.1610729992389679, -0.13939645886421204, -3.9237163066864014, 0.429939866065979, -0.21105913817882538, -0.28134557604789734, 0.049552276730537415...
677
27.08
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652854
[S1] Yeah, we, we've heard in, I think in the recent past even, a lot of people attempting to get more out of symmetries out of neural network with, I'm thinking of, I'm thinking of like, uh, group convolutional neural networks and so on, that try to actively build in symmetries into neural networks. [S2] Yeah. [S1] But, uh, it seems like they can only do that in situations where they know
23.8
3.387855
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652854.mp3
[ 6185, 5169, 3650, 7300, 9695, 12222, 2716, 5365, 11657, 8973, 617, 9257, 2731, 8540, 11650, 6420, 3568, 2993, 1960, 6698, 6311, 4971, 6964, 10539, 8042, 8009, 5442, 6146, 11328, 981, 6691, 9769, 9967, 11093, 11604, 10836, 10901, 11413, 8341...
[ 0.7339105606079102, -5.647878646850586, 0.06714997440576553, 0.01792934723198414, 0.44898587465286255, 0.013778088614344597, 0.1322716623544693, -0.03175373747944832, -0.10255549848079681, -3.7156288623809814, 0.2842426002025604, -0.0152367502450943, -0.02499213255941868, 0.064112715423107...
595
23.8
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652855
[S1] ... the symmetry that will appear. They already know a molecule. It doesn't matter which way I look at it, right? So I can directly build that in. But your reasoning is that because assessing conserved quantities, um, is an easier task than assessing symmetries, it might be possible to learn the conserved quantities, uh, dynamically, actually learn them from data. [S2] Yeah. [S1] Is that approximately correct?
25.4
3.406004
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652855.mp3
[ 4657, 5140, 2817, 9126, 6129, 8168, 6081, 11230, 12181, 5670, 7663, 4878, 7784, 9674, 2472, 2957, 8983, 12614, 3347, 12286, 4533, 1064, 3617, 174, 4015, 3165, 853, 1342, 3113, 8288, 10226, 10067, 12548, 12166, 9071, 2981, 395, 798, 1232, ...
[ 0.7234684824943542, -5.603499412536621, 0.10224258154630661, -0.08476816117763519, 0.39385369420051575, 0.07033456116914749, 0.07626157253980637, -0.008188603445887566, -0.0856674388051033, -3.7438881397247314, 0.2886233925819397, -0.030408453196287155, 0.10975515842437744, 0.0629710480570...
635
25.4
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652856
[S1] Yeah. [S2] Oh, right. [S1] But that, that's a, another thing that may be a bit different, uh, from our work than other works, which is that, um, some symmetries are only approximately conserved or conserved quantities are only approximately conserved. So for instance, you ha- if you have a, a dissipative system, like, uh, in, in the real world, there's friction, and so, uh, you actually lose energy. If you don't cons- if you don't consider entire system, uh, you'll, you usually have small losses.
26.8
3.236369
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652856.mp3
[ 4649, 8728, 7253, 10061, 11213, 9135, 8613, 2934, 2805, 4257, 3314, 1478, 3510, 1397, 2395, 1794, 7252, 3220, 8565, 4106, 3072, 10740, 11636, 9052, 3233, 3291, 3349, 6242, 8866, 4107, 3339, 11261, 11933, 10305, 450, 2845, 11045, 9063, 6503,...
[ 0.5145800113677979, -5.767024517059326, 0.07315463572740555, -0.2554255723953247, 0.7461598515510559, -0.14697541296482086, 0.3566681444644928, 0.1082453578710556, -0.23720677196979523, -3.967229127883911, 0.30849117040634155, -0.23702317476272583, -0.33699271082878113, 0.09816114604473114...
670
26.8
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652857
[S1] And maybe I wanna, I wanna get to sort of a little bit of an example of where, so people can imagine this a little bit more. Now, I only have a mouse here 'cause I forgot the iPad 'cause I'm, I'm stupid, but maybe we can give the small example of a, a pendulum, right? [S2] Mm-hmm. [S1] So here's a pendulum. It hangs here and it sort of
20.52
3.413876
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652857.mp3
[ 6184, 7777, 12078, 12212, 8781, 4445, 6249, 8732, 1566, 10539, 11752, 12491, 7976, 12105, 11659, 7332, 9780, 11262, 9381, 5365, 10486, 11062, 10956, 9135, 6575, 4078, 7337, 9258, 5863, 7923, 11869, 8094, 6171, 12434, 9190, 11261, 10019, 11416...
[ 0.7812725305557251, -5.752928733825684, -0.045062415301799774, -0.020360548049211502, 0.5565292239189148, -0.0051424941048026085, 0.31701356172561646, -0.050412170588970184, -0.08925320953130722, -3.732182264328003, 0.2742457687854767, -0.010137817822396755, 0.06513787806034088, 0.03976831...
513
20.52
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652858
[S1] ... the future, let's say. [S2] Yeah. [S1] Or at least from, from what I can tell. So what your model would be able to do is it would be able to predict the next time step right here, right? Then it's a bit here, here, uh, sorry. It's a, it's a little bit more up to the left, right? [S2] Mm-hmm. [S1] So it's a little bit more up and then it's, it's even more up over here and then it swings back and so on. It swings back over. Now,
23.96
3.397678
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652858.mp3
[ 5672, 7713, 9231, 3075, 5385, 11182, 4514, 6448, 4025, 7379, 10053, 5327, 5254, 4819, 10833, 12755, 10200, 2990, 3046, 403, 322, 3731, 4786, 2205, 2845, 6135, 4717, 8400, 5473, 5465, 393, 9143, 4084, 5106, 2499, 4939, 8802, 10030, 10036, ...
[ 0.8467458486557007, -5.626096725463867, -0.033399783074855804, -0.06039230525493622, 0.4727981388568878, 0.04274888336658478, 0.2260766625404358, -0.003642312018200755, -0.12002212554216385, -3.7415003776550293, 0.30607274174690247, 0.028208177536725998, 0.038150854408741, 0.02667573094367...
599
23.96
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652859
[S1] It will not actually make use of the fact that something is conserved. [S2] Exactly. [S1] So you, you go ahead and you say, "Since this is a dynamical system, we know more about the system. We can impose additional constraints." And the additional constraints right here, if I see this correctly, essentially at every time step you say,
19
3.446182
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652859.mp3
[ 6769, 3084, 12526, 12245, 602, 2733, 11565, 5332, 8780, 10557, 10418, 4270, 4136, 8815, 8189, 10036, 11555, 12762, 3984, 7965, 12612, 6157, 10916, 12790, 12181, 7308, 11298, 7578, 6796, 7434, 9605, 10451, 7756, 2566, 199, 5019, 8672, 2970, ...
[ 0.7644492387771606, -5.7555670738220215, -0.04248243570327759, -0.016025247052311897, 0.3863508403301239, 0.07892829179763794, 0.1953110694885254, -0.03826412558555603, -0.15935736894607544, -3.7887473106384277, 0.26067420840263367, 0.019033467397093773, 0.02837173454463482, 0.038677737116...
475
19
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652860
[S1] And it always needs to predict the same thing, right? Since, since it needs to, uh, figure out a quantity that is conserved. [S2] Exactly. [S1] And now it is, it is, if I just train a neural network to always predict the same number right here, I would just end up with a neural network that is predicting some, some kind of a constant, right? [S2] Yeah. [S1] So your method figures out how do I need to build
28.4
3.270911
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652860.mp3
[ 5672, 3633, 11300, 11773, 12061, 9282, 9119, 9931, 3036, 7931, 10418, 3170, 10557, 12180, 6552, 651, 10901, 12683, 12492, 5905, 2897, 3265, 4060, 3204, 10930, 4016, 3037, 3853, 9033, 11182, 12189, 9258, 12179, 4121, 194, 5046, 8681, 9689, 6...
[ 0.7935916781425476, -5.611802101135254, 0.016887884587049484, 0.00680740037932992, 0.3808141350746155, 0.017207946628332138, 0.19656731188297272, -0.03630748391151428, -0.1571626514196396, -3.7609992027282715, 0.24260857701301575, 0.08283078670501709, 0.028543855994939804, 0.07491514831781...
710
28.4
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652861
[S1] Um, so this is the physics 101 data set, um, from Josh Tenenbaum's group. I think Jesuine was the first author. And they have a collection of videos, and in this case it's a, they have a hand dropping an object passively, like it, it just lets it drop down, and the object falls down, and there's a second object at the end of the ramp. They collide, and then, uh, the other one, sometimes depending on the masses and the friction and whatnot, uh, the dynamics are kind of, can, can change. [S2] Mm-hmm. [S1] Uh, but that's the data set.
26.48
3.312711
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652861.mp3
[ 11604, 6809, 9829, 7284, 10069, 8882, 9865, 4839, 11766, 10670, 11190, 10741, 8190, 6886, 8403, 3293, 6371, 6880, 2824, 8116, 10542, 6725, 2696, 12188, 11587, 3025, 11091, 5961, 3204, 2835, 11188, 11686, 6496, 5432, 9180, 11652, 3008, 6092, ...
[ 0.5998235940933228, -5.807607650756836, -0.08215015381574631, -0.02447568066418171, 0.748042106628418, -0.04305308312177658, 0.3246605396270752, 0.08104247599840164, -0.15597166121006012, -3.897087335586548, 0.31464555859565735, -0.20380322635173798, -0.21485930681228638, -0.00103789870627...
662
26.48
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652862
[S1] And does, so that there are multiple videos- [S2] Yes. [S1] ... and it's always different objects or- [S2] Uh, like some objects could be, uh, common between, uh, videos, but- [S1] Mm-hmm. [S2] ... there's lots of objects, so, uh, it's not always the same object and that's- [S1] Yeah. [S2] ... can point to the fact that it can vary. Um, so- [S1] Mm-hmm. [S2] ... one nice thing about the, in other networks is that, um, they can deal with, uh, with raw video. So some, usually conserve quantities,
27.08
3.275151
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652862.mp3
[ 8777, 10139, 7284, 3764, 4962, 6769, 10989, 11260, 12203, 10315, 11406, 8715, 8149, 4445, 8801, 9762, 6805, 9755, 11865, 6193, 8281, 10496, 11221, 10677, 7957, 5577, 5570, 3010, 6866, 9250, 9242, 11298, 5154, 8348, 7204, 4379, 10712, 5504, ...
[ 0.6509643793106079, -5.726364612579346, -0.018672071397304535, -0.0627375990152359, 0.591183066368103, -0.032542627304792404, 0.30112096667289734, 0.05806456133723259, -0.19579486548900604, -3.8861513137817383, 0.22996042668819427, -0.07245976477861404, -0.16312433779239655, 0.012229015119...
677
27.08
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652863
[S1] Yeah. So here, the, the, the diagram shows a little bit of, of what you're, what you, um, are trying to do, but also what you're trying to avoid. So the bottom path right here, if I see this correctly, that would be if I did nothing else except the bottom path, I would build this neural network to just predict sort of the, the, the future time steps. And that often turns out poorly. [S2] Mm-hmm. [S1] I don't know. This is a, quite a pixel-ish,
29.08
3.469985
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652863.mp3
[ 7274, 9265, 1603, 10117, 10742, 1836, 553, 3741, 11946, 9306, 11932, 6756, 12403, 6747, 6813, 4710, 7211, 4134, 4206, 7205, 4142, 2597, 2076, 1626, 6177, 5737, 5145, 9793, 4681, 3731, 7849, 3609, 8833, 7217, 4674, 6705, 4629, 722, 6112, 3...
[ 0.8140379786491394, -5.603418350219727, -0.007100473158061504, 0.008310975506901741, 0.4742708206176758, 0.0031012094113975763, 0.19011451303958893, -0.08267756551504135, -0.13734076917171478, -3.7368741035461426, 0.27166295051574707, 0.019436826929450035, 0.10032743215560913, 0.0496099404...
727
29.08
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652864
[S1] Um, mess, but it, it, it sort of, it sort of, all of a sudden there are, like, three objects instead of two, and th- the one is, the one is kind of gone or split up. [S2] Yeah. [S1] And it's a, it's a bit of a, a mess. And you attribute this to the fact that it's just a video prediction or-
19
3.301306
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652864.mp3
[ 4129, 7713, 12391, 9199, 4079, 3575, 3566, 3575, 4015, 6221, 5678, 3110, 11303, 7387, 11305, 8360, 7729, 2052, 5718, 11294, 11365, 8742, 10908, 12775, 12790, 11755, 12652, 7666, 3552, 3474, 2961, 3465, 464, 7322, 12517, 9974, 4394, 7184, 53...
[ 0.8282726407051086, -5.7625956535339355, -0.1088097095489502, -0.04559221863746643, 0.3750263750553131, 0.06801831722259521, 0.267947793006897, -0.039654675871133804, -0.0733712837100029, -3.7556073665618896, 0.2285650670528412, 0.11248553544282913, 0.08990537375211716, 0.0297853983938694,...
475
19
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652865
[S1] At the very few, like at the beginning of the frames, like the first few frames, there was not that much mistakes, but when you go very far into the future, then it's, it's much harder. Um, so- [S2] Yeah. [S1] ... those two problems, lack of data and the fact that you go a lot of, into the future. [S2] Your method is, and you also have an algorithm described somewhere. It's a bit of a, it's a, it's a algorithm that is, oh, right here. It's an algorithm that has multiple steps in it, and one special part is that you have this sort of inner optimization loop right here. [S1] Yeah. [S2] Now,
29.8
2.94873
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652865.mp3
[ 4203, 8282, 6884, 5502, 9974, 5121, 3216, 7437, 9414, 9990, 7557, 4363, 9380, 1181, 3877, 3805, 3155, 3107, 6177, 6185, 3594, 3594, 3155, 1371, 12781, 9069, 341, 8590, 10124, 3440, 7593, 9987, 11020, 5206, 1759, 12654, 8862, 5541, 10094, ...
[ 0.670535147190094, -5.663235664367676, 0.008839651942253113, 0.011367485858500004, 0.5341405868530273, 0.04107864573597908, 0.20862184464931488, -0.023860398679971695, -0.16035950183868408, -3.809791088104248, 0.2830492854118347, 0.03774743154644966, 0.03392321988940239, 0.0837969407439231...
745
29.8
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652866
[S1] I wanna maybe go back to the diagram and let's go, let's walk through it once. Before we, before we, you know, take a look at the formulas and all, we can walk through it once. So the first thing that happens, if I understand correctly, is you take your first input and you do exactly what we just said. You run it through a forward prediction neural network that just tries to predict the future, um, just plain by itself. [S2] Yeah. [S1] Right.
27.72
3.344877
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652866.mp3
[ 5744, 11299, 12222, 8125, 9446, 2805, 7925, 12022, 8453, 9071, 12143, 9127, 8687, 2469, 6828, 11314, 6705, 2062, 8421, 12267, 12163, 7776, 4516, 12549, 10054, 7558, 9542, 4420, 602, 1401, 6819, 6258, 11987, 5935, 10470, 1058, 10984, 11767, ...
[ 0.7751098871231079, -5.5993523597717285, 0.008594769053161144, -0.00978932436555624, 0.36663269996643066, 0.04201892018318176, 0.1823028326034546, -0.044867321848869324, -0.17145252227783203, -3.7749855518341064, 0.3244750201702118, 0.029104329645633698, 0.09248568117618561, 0.093528315424...
693
27.72
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652867
[S1] So this has, this has a bit of a, of a default thing. But now you try to improve that. And this is all, this is the entire thing we're describing right now. That is one forward pass through your system. [S2] Mm-hmm. [S1] So ev- you would take every single prediction that you made and you would feed it through this G-network right here. And this G-network is, you call it an embedding network. [S2] Mm-hmm. [S1] That is the thing ultimately that's trying to predict a conserved quantity.
28.96
3.486659
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652867.mp3
[ 2088, 8849, 1232, 860, 1629, 2560, 6116, 12629, 5601, 5569, 7668, 9727, 8567, 3472, 2961, 329, 6169, 4137, 2825, 12268, 3427, 5513, 3554, 12222, 11198, 6485, 458, 9135, 4122, 10665, 11749, 7245, 5580, 10551, 3491, 6128, 3502, 4015, 1462, ...
[ 0.7404557466506958, -5.631862163543701, 0.03038349375128746, 0.022720934823155403, 0.44495266675949097, 0.06365137547254562, 0.15796950459480286, -0.04675755277276039, -0.14977319538593292, -3.754385471343994, 0.2632155120372772, 0.03537017107009888, 0.047918953001499176, 0.073940545320510...
724
28.96
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652868
[S1] But it's not, it's not necessarily just outputting one number. It's outputting an entire vector. [S2] Mm-hmm, mm-hmm, yeah. [S1] So it's an, outputting an embedding vector and the, the goal obviously is that for all of these inputs, it should output the same embedding vector.
15.8
3.217804
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652868.mp3
[ 6184, 7729, 10985, 8639, 12382, 462, 12188, 9810, 5523, 5592, 3331, 10388, 7990, 10868, 10620, 9889, 5664, 2953, 11297, 9258, 8749, 12757, 9745, 2898, 714, 10460, 7994, 11382, 1568, 4163, 10964, 12221, 2912, 5568, 3813, 5600, 6611, 11195, 5...
[ 0.796582818031311, -5.8233723640441895, -0.140767902135849, 0.0009273095056414604, 0.4723527133464813, 0.08299762010574341, 0.21464170515537262, -0.07483884692192078, -0.09227418899536133, -3.7269985675811768, 0.2423694133758545, 0.0036838860251009464, -0.01649666018784046, 0.0425464510917...
395
15.8
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652869
[S1] Exactly. [S2] But all the same across the, across the, uh, video sequence, okay? So this is how we can imagine you train this G-network to sort of predict whatever is special about this particular data point, but inside of the data point, conserved among all the frames. [S1] Exactly. Because if it was the same A for everyone, then you would have the issue that you mentioned at the beginning, then it's a useless conserved point.
25.44
3.396624
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652869.mp3
[ 11938, 6354, 8876, 9877, 7572, 3019, 8638, 4901, 9947, 2204, 10593, 8063, 6391, 10944, 4452, 8233, 10870, 2806, 3177, 8746, 6193, 11375, 8502, 10421, 10421, 11059, 7957, 3777, 3447, 1435, 10721, 8657, 3539, 11764, 11749, 2436, 2445, 5669, 6...
[ 0.7086426615715027, -5.662154674530029, 0.014759296551346779, -0.09785746783018112, 0.5294415950775146, 0.029721437022089958, 0.1695239245891571, 0.007338250055909157, -0.2629934847354889, -3.8269009590148926, 0.3107975721359253, -0.004091132432222366, 0.024529557675123215, 0.0559459961950...
636
25.44
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652870
[S1] Yeah. Um- [S2] It, it restricts, it restricts what you can output, right? Because ideally, the F-network should only output whatever the G-network says is, is the same, right? [S1] Yeah. [S2] If the F-network can only output things that the G-network will embed to the same place in the embedding space or a similar place.
21.16
3.433089
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652870.mp3
[ 3793, 1747, 7061, 1341, 1128, 6697, 8754, 7891, 6761, 5161, 9811, 7021, 3886, 3887, 3951, 1391, 1334, 3822, 6237, 6819, 11741, 9858, 11408, 5665, 8962, 8745, 7204, 11677, 11156, 2258, 3741, 2844, 5524, 7004, 9192, 5601, 3336, 3785, 9618, ...
[ 0.823232889175415, -5.681219577789307, -0.03168993070721626, -0.006517555098980665, 0.5076174736022949, 0.03293566778302193, 0.14903941750526428, -0.02575136534869671, -0.13338777422904968, -3.7200796604156494, 0.3030100464820862, 0.028910284861922264, 0.07075834274291992, 0.04021979868412...
529
21.16
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652871
[S1] ... of this information that the G-network output out of the initial sequence again. And you do this in a very special way in that you actually take the parameters of F and you update them on the fly. [S2] Yes. [S1] Uh, you update them on the... So this is within a forward pass. You actually update the parameters
19.36
3.43071
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652871.mp3
[ 6192, 7712, 11503, 5943, 3190, 5879, 5221, 489, 5552, 1834, 5760, 10151, 10071, 1427, 5513, 2962, 5593, 457, 3778, 11825, 5225, 12450, 6769, 8218, 12710, 12108, 10766, 3931, 6009, 443, 5525, 5782, 3118, 10972, 11767, 12181, 12676, 5082, 970...
[ 0.721305787563324, -5.728343963623047, -0.037755001336336136, -0.044394638389348984, 0.4791141450405121, 0.08710592240095139, 0.17870290577411652, -0.06644228100776672, -0.1278018057346344, -3.764582872390747, 0.2792128622531891, -0.05600711703300476, 0.10150863230228424, 0.066768683493137...
484
19.36
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652872
[S1] into the direction of the gradient of G. [S2] Exactly. Yes. So- [S1] So, yeah. Sorry. This is, I think that, that, it takes it, yeah. So here you have this neutralos. [S2] Yes, exactly. [S1] Which you maybe want to talk about this briefly.
17.72
3.271819
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652872.mp3
[ 6193, 7712, 11365, 11723, 11585, 11277, 12371, 5225, 12233, 6571, 10957, 3206, 8256, 9695, 4742, 8577, 10087, 6438, 788, 845, 9647, 11253, 7466, 8754, 12196, 9697, 10192, 3045, 8342, 8207, 6422, 8809, 8310, 7998, 5932, 4081, 8385, 3959, 925...
[ 0.8298799991607666, -5.770203113555908, -0.0668395534157753, -0.07657423615455627, 0.44547709822654724, 0.00853731483221054, 0.17481829226016998, -0.06453493982553482, -0.22571659088134766, -3.7357938289642334, 0.28373637795448303, -0.04596327617764473, -0.00012848200276494026, -0.01433231...
443
17.72
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652873
[S1] and now we update theta, and theta are the parameters of F, right? [S2] Exactly. [S1] Theta are the parameters of F. We update these on the fly, and I, I, I suppose that we just do this, you know, in the moment and for the next data point we, we go back to the sort of original parameters. [S2] Yes. [S1] Uh, and do this again. So this is sort of an on-the-fly update for, you know, a temporary update of these parameters into the direction
25.92
3.239306
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652873.mp3
[ 3113, 7720, 11638, 7374, 5637, 8773, 8718, 11222, 12725, 12603, 10403, 2716, 2683, 11101, 12547, 5061, 6379, 8745, 8745, 5745, 11375, 5566, 10619, 9258, 10289, 5737, 9194, 9824, 3625, 8002, 11180, 9711, 2518, 5061, 7557, 7557, 11928, 6177, ...
[ 0.7682434320449829, -5.552507400512695, 0.08033601194620132, -0.12870219349861145, 0.45626768469810486, 0.04527881368994713, 0.19887854158878326, 0.01992158405482769, -0.09041404724121094, -3.7750139236450195, 0.3482648432254791, 0.020271461457014084, 0.040589723736047745, 0.06202369183301...
648
25.92
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652874
[S1] Exactly, exactly. So this is, uh, some, uh, previous work of ours, uh, which we call tailoring. And the idea of- [S2] Mm-hmm. [S1] ... of tailoring is just because of what you said, that the fact that the adaptation is customized for each individual data point. [S2] Mm-hmm. [S1] Uh, and the idea there was, uh, a general way of encoding inductive biases, uh, with, uh, unsupervised auxiliary losses. So
19.72
3.062619
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652874.mp3
[ 8744, 11622, 7309, 10012, 833, 3935, 9079, 7453, 11946, 1884, 3934, 7378, 3305, 6193, 5160, 12315, 4455, 12444, 1857, 3047, 4598, 7332, 10147, 3660, 1820, 9000, 5440, 11125, 10413, 6350, 3152, 3361, 12180, 5962, 3009, 10650, 11532, 4554, 58...
[ 0.496900349855423, -5.9109930992126465, 0.004077565390616655, -0.20128045976161957, 0.7276714444160461, -0.1661660224199295, 0.4107508957386017, 0.10646949708461761, -0.12711673974990845, -3.8591232299804688, 0.3728911876678467, -0.25075677037239075, -0.3427937626838684, 0.0831920430064201...
493
19.72
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652875
[S1] ... loss is in general, you say, for instance, one thing we could say is, "Oh, why not we add energy conservation when, when we train?" Sometimes, auxiliary losses would say, "Okay, I train for good predictions and I train for energy conservation at training time." But if you do that, you're not going to enforce energy conservation at test time. [S2] Yeah. [S1] Because at test time, you're going to have a generalization gap, uh, in energy conservation.
21.08
3.318252
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652875.mp3
[ 4204, 4256, 4172, 10046, 3374, 6308, 4995, 3531, 11091, 5981, 5845, 5550, 10494, 6626, 8064, 9573, 5824, 979, 7183, 6212, 9600, 11247, 12524, 5908, 7981, 8502, 5762, 4875, 9541, 2839, 4955, 8592, 3036, 12775, 7621, 4192, 5497, 5405, 6557, ...
[ 0.547906756401062, -5.831568717956543, -0.0893828272819519, -0.13150782883167267, 0.8349735736846924, -0.18422837555408478, 0.40756598114967346, 0.11023510992527008, -0.15759624540805817, -3.870570421218872, 0.36208376288414, -0.2018779069185257, -0.2628067433834076, 0.005724901333451271, ...
527
21.08
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652876
[S1] But- [S2] Yeah. [S1] Uh, energy, uh, conservation, or any type of conservation, or any auxiliary laws, can be checked before making the prediction at test time or at training time. Inside the prediction function, I can first make my prediction and see, okay, uh, do I like it? Does my auxiliary laws, does my unsupervised laws like this prediction?
18.64
3.094048
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652876.mp3
[ 7186, 8224, 9301, 3618, 5620, 10673, 10675, 6365, 5776, 3088, 3473, 7474, 5792, 2068, 1057, 8026, 12758, 11923, 4084, 3212, 3617, 9266, 5160, 4810, 7721, 5665, 1683, 7493, 2998, 3900, 4266, 3681, 9243, 8913, 6369, 9843, 3666, 9819, 9647, ...
[ 0.5740753412246704, -5.960941791534424, -0.14256669580936432, -0.047546517103910446, 0.7567527890205383, -0.1291540414094925, 0.39323049783706665, 0.05375494435429573, -0.045079074800014496, -3.8543436527252197, 0.3468329906463623, -0.2539527714252472, -0.3055499494075775, 0.00299502094276...
466
18.64
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652877
[S1] Yeah, maybe it's, it's also important to highlight that the, the, the parameter here, this theta that we start with, and also the parameters of G, those are the ones that will be learned during the training procedure across the entire training data set. And then the parameters here, those are always constructed in the moment, data point by data point- [S2] Exactly. [S1] ... to,
24.2
3.378997
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652877.mp3
[ 6696, 5144, 4748, 12164, 10677, 1843, 1057, 4135, 11109, 11745, 7181, 8040, 12100, 10644, 11182, 9729, 2970, 5537, 3024, 6793, 6249, 9383, 11678, 12110, 5648, 3027, 464, 11631, 7799, 7918, 1954, 5513, 3044, 8431, 1438, 7446, 5668, 8297, 140...
[ 0.7895370125770569, -5.644230842590332, -0.04177823290228844, -0.0060137175023555756, 0.42354488372802734, -0.0021596341393887997, 0.18649786710739136, -0.07277274131774902, -0.17062503099441528, -3.722886085510254, 0.29839184880256653, 0.015071652829647064, 0.1124911978840828, 0.046819902...
605
24.2
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652878
[S1] ... As you say, tailor the inductive bias. And the inductive bias in this case would sort of be the, this entire term right here. [S2] Exactly. [S1] Essentially says, you know, what do, how do I need to change my predictor in order to conserve the particular thing that G decides is the common quantity for this data point. [S2] Yeah, yeah.
20.36
3.449466
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652878.mp3
[ 6681, 10385, 10004, 7276, 12526, 12278, 8637, 6420, 960, 9861, 8966, 2514, 11753, 5584, 11774, 2454, 4997, 10848, 10768, 11226, 7666, 10226, 12772, 12675, 11668, 2708, 8109, 11107, 2893, 8525, 3280, 7014, 11971, 4055, 9246, 8513, 6455, 8951, ...
[ 0.893925130367279, -5.672552108764648, -0.03166671097278595, -0.014118524268269539, 0.4624546766281128, -0.04575183615088463, 0.1679815948009491, -0.04169321805238724, -0.08559892326593399, -3.7523529529571533, 0.21940895915031433, -0.040457434952259064, 0.03477838262915611, 0.011410579085...
509
20.36
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652879
[S1] and the L task here, that would just be the, what do you call the task loss. This would be the video prediction loss. [S2] Exactly. [S1] Or something like this. Okay. So my, I have, I have, I have a lot of questions. Um, first, first of all, this, it seems, it seems quite, um, intricate, right? Because if I think, okay, these outer gradients right here, especially this gradient right here, this is, how do I need to change theta? Now, okay,
28.16
3.311375
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652879.mp3
[ 2672, 4131, 8950, 2549, 2421, 4630, 9733, 5760, 1305, 3088, 1939, 9924, 4939, 5681, 10861, 12782, 12787, 10925, 5341, 5453, 7457, 8392, 6632, 5044, 10175, 9661, 8637, 9598, 8617, 5585, 4891, 10147, 7600, 4877, 9990, 12135, 8020, 2964, 3734,...
[ 0.7504711151123047, -5.548388481140137, 0.012188289314508438, -0.03246091678738594, 0.564893901348114, 0.04579634591937065, 0.16643042862415314, -0.03624039143323898, -0.14696210622787476, -3.7966809272766113, 0.19664233922958374, 0.003792242845520377, 0.1527264565229416, 0.050970606505870...
704
28.16
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652880
[S1] How do we need to change theta? This depends on these predictions right here. These predictions right here have one forward pass using theta, then have a gradient with respect to theta right here inside of them, right? [S2] Yeah. [S1] And, and the, all of those come from this quantity, which is already a forward pass using theta, right? Is, is this,
24.56
3.491502
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652880.mp3
[ 4649, 7713, 6980, 7468, 9725, 8054, 9794, 2910, 7966, 11781, 12694, 9989, 10952, 1942, 8321, 12689, 6639, 5085, 9806, 7554, 5056, 1808, 5832, 8634, 11552, 12178, 8962, 8567, 9455, 2752, 11701, 7004, 8088, 1153, 5248, 12758, 6184, 6192, 7154...
[ 0.7064188718795776, -5.677851676940918, -0.014043397270143032, -0.039112769067287445, 0.4936772584915161, 0.11409427970647812, 0.24087347090244293, -0.04528265818953514, -0.044007617980241776, -3.7510719299316406, 0.2546812891960144, -0.009845473803579807, 0.11130528151988983, 0.0409582592...
614
24.56
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652881
[S1] Uh, is stable, then it works fine. Um, but if it, if the overall thing is already unstable, then it's ex- it's extremely tricky to, to, uh, add things there. So for instance, uh, one thing we realized was that, um, because video prediction is very expensive, uh, and, and basically we couldn't fit that many examples on a GPU, literally, I think, two or four. [S2] Right. [S1] Uh, so, um, we were initially using batch normalization.
26.84
3.245497
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652881.mp3
[ 3624, 6658, 9902, 6591, 6007, 6063, 3503, 5998, 3438, 3958, 3958, 3446, 3446, 3965, 3884, 7467, 8233, 8289, 11300, 12613, 9988, 3916, 712, 2834, 5528, 2953, 4442, 9770, 8809, 5289, 5664, 7198, 5027, 5912, 3088, 3338, 12781, 11723, 6163, 5...
[ 0.6157985925674438, -5.811166763305664, 0.04294733703136444, -0.11421522498130798, 0.6925292611122131, -0.1748984009027481, 0.352013498544693, 0.08588189631700516, -0.18237292766571045, -3.95110821723938, 0.4015984535217285, -0.2507287263870239, -0.35030320286750793, 0.02351865917444229, ...
671
26.84
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652882
[S1] If the grad- of the, the original gradients, because of the batch normalization, if you compute the batch statistic with a very small batch, it's already very crazy and unstable. [S2] Mm-hmm. [S1] And then we could have it. When the or- the other thing is already stable, then it seems, um, for us, it worked pretty out of the box when we, when we swapped the linear normalization. [S2] Okay. That sounds good. Yeah, I would, I would expect so, um-
22.08
3.207426
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652882.mp3
[ 6192, 9241, 12565, 11651, 5916, 3738, 2840, 9053, 11524, 12037, 12166, 12677, 4485, 6809, 11929, 11873, 11322, 9256, 8745, 6185, 5673, 6193, 12369, 8790, 5723, 11371, 804, 3532, 11230, 5012, 6298, 10085, 10021, 7919, 3291, 667, 5121, 5614, ...
[ 0.7831315398216248, -5.824700832366943, -0.048473093658685684, -0.1159050464630127, 0.6492183208465576, -0.050268951803445816, 0.43590280413627625, 0.051971226930618286, -0.11161254346370697, -3.9607834815979004, 0.19581933319568634, -0.13610436022281647, -0.30627191066741943, 0.0105286445...
552
22.08
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652883
[S1] was when we were using batch normalization and we were wondering, "Oh, is- is our another network unstable?" Or- [S2] Yeah. [S1] And then we realized, okay, no, it's the- it's the- the vanilla network that was unstable. Um, but, uh, that was part of our concern because, uh, there is some papers that mention that when there's a- when you're back propagating through a very deep, um, graph, uh, then the gradients are sometimes not very informative. [S2] Mm-hmm. [S1] Um, in our case, we found that, uh, when the- our thing is pretty stable,
27.24
3.025206
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652883.mp3
[ 4202, 9240, 4190, 3109, 5226, 5300, 5302, 3375, 8951, 8502, 11126, 11702, 6609, 5464, 5449, 4949, 8878, 2675, 5366, 12278, 9942, 5782, 2789, 12102, 165, 7990, 5525, 3853, 10051, 5959, 4033, 11669, 11988, 4638, 8189, 10166, 11585, 4032, 8853...
[ 0.7478934526443481, -5.684506416320801, -0.013155665248632431, -0.14443309605121613, 0.7434341907501221, -0.15126082301139832, 0.322050005197525, 0.1281297206878662, -0.04854438453912735, -3.9509739875793457, 0.42470863461494446, -0.22037771344184875, -0.3357049822807312, -0.08138347417116...
681
27.24
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652885
[S1] That, that, that might also be something where stability or computational graph size. First of all, you just do a gradient step. Many things would be possible, right? You could do an Ad- an Ad-grad step, you could do an Adam step, you could do a line search or a Newton step or anything like this, but you have chosen to do like the most simple thing, which is a single gradient step, right? [S2] I think, I think the, the- [S1] Yeah.
26.32
3.392578
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652885.mp3
[ 5648, 5681, 4621, 5450, 11260, 10683, 4380, 3096, 1042, 2963, 11709, 12590, 5515, 9446, 1688, 4116, 3282, 10683, 7477, 33, 8807, 11126, 11637, 9998, 7768, 8754, 6176, 9240, 8503, 10429, 8934, 5537, 3537, 3892, 3166, 10416, 5015, 9030, 4931,...
[ 0.7932926416397095, -5.604663848876953, 0.02702830731868744, -0.014306903816759586, 0.38908687233924866, 0.021092094480991364, 0.1211545467376709, -0.039863090962171555, -0.17423300445079803, -3.7924094200134277, 0.34380316734313965, 0.00015696670743636787, 0.08745371550321579, 0.073340982...
658
26.32
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652886
[S1] Uh, when you force the conserve G, all these features say, okay, no, you should conserve G and, therefore, it's kind of, uh, projecting one dimension. And so, um, in particular for conserve quantities, applying the same laws over and over, uh, it's kind of stable because, uh, it, you will just keep going closer to this, uh, manifold of- [S2] Mm-hmm. [S1] ... predictions that conserve, uh, G.
21.4
3.29015
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652886.mp3
[ 6184, 7712, 11998, 12269, 11630, 11759, 8686, 3365, 107, 5738, 8297, 8424, 9236, 2731, 11702, 12358, 9859, 9038, 5214, 1195, 5360, 2801, 10354, 8033, 6547, 1296, 7961, 8415, 12506, 5367, 11412, 6552, 3043, 5542, 3348, 682, 8913, 9153, 11089...
[ 0.5323896408081055, -5.891219139099121, -0.08418269455432892, -0.07803566753864288, 0.8425347208976746, -0.09039291739463806, 0.44781380891799927, 0.15543779730796814, -0.1104183942079544, -3.890838384628296, 0.27585387229919434, -0.14703437685966492, -0.2515547573566437, 0.039770547300577...
535
21.4
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652887
[S1] So there's no, no, let's say, danger of over, overdoing. I mean, there's a little bit, but it, as I said, it hits after like 100 steps, which- [S2] Yeah, yeah. [S1] ... which is quite a bit, right? Given that you train with one.
12.2
3.534303
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652887.mp3
[ 6827, 5681, 4428, 3024, 11118, 3787, 9206, 8021, 5534, 5001, 339, 8214, 6053, 10934, 10934, 10405, 7894, 10470, 7966, 6749, 9827, 11379, 9826, 9998, 7387, 9793, 10890, 10011, 7286, 3236, 7217, 3138, 6799, 10901, 10613, 10412, 10918, 7911, 6...
[ 0.990776002407074, -5.788281440734863, -0.13488072156906128, 0.05750149488449097, 0.500278890132904, -0.0030178504530340433, 0.2911088764667511, -0.07028695195913315, -0.13602083921432495, -3.7283763885498047, 0.2363615483045578, -0.04533426836133003, -0.015723034739494324, -0.007587935309...
305
12.2
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652888
[S1] But because it's a neural network, then suddenly I think you've tra- you're going, uh, outside. It's kind of a distribution shift. You train G to be useful for one or two or three- [S2] Yeah. [S1] ... green steps. Now you're using it for 100. It doesn't make you any promises. Um- [S2] Yeah.
14.44
3.254836
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652888.mp3
[ 7266, 11273, 5727, 5668, 8937, 8631, 5494, 1015, 7535, 11344, 5192, 1755, 3619, 6498, 10117, 11929, 2934, 10550, 4544, 12573, 10946, 3026, 9463, 7693, 10006, 11614, 2901, 3502, 9318, 8100, 12715, 5273, 5427, 7955, 478, 9941, 11882, 2395, 30...
[ 0.61126708984375, -6.0702619552612305, -0.09284675866365433, 0.016791410744190216, 0.6812424659729004, -0.09810420125722885, 0.4329046905040741, 0.07113566994667053, -0.20854023098945618, -3.883117437362671, 0.19127848744392395, -0.1606038212776184, -0.2587840259075165, 0.05229330062866211...
361
14.44
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652889
[S1] ... forward predicts. For example, you could have optimized the predictions themselves at runtime- [S2] Yes. [S1] ... uh, to make both of them happy. You could have, um, I, I don't know, you could have, uh, just learned it as, as one thing and not even bothered with runtime optimization. Why did you
19.64
3.340468
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652889.mp3
[ 3106, 5723, 3496, 4088, 2734, 10414, 8493, 2892, 685, 11118, 5517, 10889, 6761, 10609, 3501, 2822, 8515, 12207, 9558, 10915, 3434, 5522, 4440, 12005, 12662, 11885, 6821, 6809, 4458, 3064, 10492, 8467, 5588, 12054, 12572, 400, 8677, 12278, 1...
[ 0.8185270428657532, -5.7657470703125, -0.06964588165283203, 0.055917996913194656, 0.4193628430366516, 0.03934613615274429, 0.2387813925743103, -0.013472999446094036, -0.022922717034816742, -3.714689254760742, 0.2835234999656677, -0.022177862003445625, 0.14194782078266144, 0.084327965974807...
491
19.64
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652890
[S1] and this is going to lead me to, uh, good predictions. But, uh, this is only happens, you only can look at the effect at the very end of training, and then you're going to use that on validation. [S2] Mm-hmm. [S1] And so, uh, you could do that, and I think there's papers that do that using implicit gradients, um, but the signal is much, much, uh, more cumbersome. Um, instead, if you use, if you say, "Okay, no, uh, the way I'm optimizing this is inside the prediction function," then you can literally compute the grain, the, the, the computation graph and optimize it, and to it.
29.68
2.82559
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652890.mp3
[ 3632, 8784, 4702, 10173, 5039, 9302, 7246, 3200, 3933, 12180, 6476, 977, 3918, 3906, 1749, 8996, 5815, 3358, 7382, 7188, 10888, 2980, 3732, 5900, 4998, 9304, 4244, 10902, 4996, 6176, 11009, 2869, 22, 87, 599, 534, 1179, 8817, 11301, 4031,...
[ 0.679276704788208, -5.700066566467285, -0.016779586672782898, -0.08445495367050171, 0.8063709139823914, -0.06250432878732681, 0.313103586435318, 0.15463517606258392, -0.1054757684469223, -3.950892448425293, 0.4176570773124695, -0.16354911029338837, -0.32921481132507324, -0.0107185160741209...
742
29.68
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652891
[S1] ... the ways to change the output conditioned on the input that, uh, kind of still do not, um, deviate too much from what it has learned. [S2] Mm-hmm. [S1] Uh, so theta captures the dynamics and says, "Okay, I probably got it a bit wrong because I'm not conserving G." Uh, so, but- [S2] Mm-hmm. [S1] ... but I don't want to deviate too much from what I've learned. So, uh, optimizing theta, still make sure that you're satisfied what you've learned so far, and then, and then it leads to much, much larger, uh, improvements. [S2] Yes.
25.2
3.105202
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652891.mp3
[ 4137, 8280, 6732, 523, 5595, 9079, 2667, 5308, 11123, 11600, 11659, 6040, 3145, 8464, 5412, 4172, 12177, 9682, 12252, 11981, 9740, 2818, 281, 7364, 9174, 3453, 5950, 11062, 8907, 5728, 10537, 237, 3303, 4626, 5897, 9386, 5675, 9643, 5430, ...
[ 0.6061731576919556, -5.809365749359131, -0.14145398139953613, -0.1265614926815033, 0.7608469128608704, -0.11323311924934387, 0.38745638728141785, 0.1073879525065422, -0.11600266396999359, -3.897937059402466, 0.37421494722366333, -0.1293376237154007, -0.2509937286376953, -0.0410993658006191...
630
25.2
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652892
[S1] It, it's, it's, I think there's something like what you said that, that, that, that going to be, uh, there. Uh, in particular, it, it, I, I think it, G has a feeling like, uh, like this adversarial discriminator because it's telling you, "Oh, if you're not satisfying G conservation- [S2] Yeah. [S1] ... then most likely you are wrong." Especially if you don't satisfy it by a large amount. Because, again, they're, they're approximately conserved, so, um, that's one. Um, so one thing, uh, I'm,
25.12
2.912322
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652892.mp3
[ 8232, 11356, 12180, 8203, 5825, 4244, 9044, 8908, 2569, 5570, 5512, 2957, 11102, 3998, 4956, 11875, 12069, 12172, 8202, 2753, 265, 4300, 11932, 12598, 11252, 11756, 9938, 8408, 12186, 11858, 6289, 8116, 8085, 4554, 3401, 3410, 5465, 8072, 5...
[ 0.4954885244369507, -5.853196620941162, -0.07474453002214432, -0.08247414231300354, 0.8160539269447327, -0.015793561935424805, 0.374176949262619, 0.10357244312763214, -0.1906626671552658, -3.941352128982544, 0.26861968636512756, -0.16720768809318542, -0.26535215973854065, -0.00352293974719...
628
25.12
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652893
[S1] ... interested in, uh, going forward, and, and I think that, that could be a, a venue of, of many future works, is that we focused a lot on when we were trying to make predictions on kind of generative, uh, networks. So the fact that you, sorry, generative not in the sense of self-supervised learning, but- [S2] Yeah. [S1] ... but more in, like, you predict the next input given the, the pre- the, sorry, the, the, the output given the input. So you have to generate the, the, the thing.
24.44
3.111304
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652893.mp3
[ 5672, 5681, 12307, 12692, 11979, 11524, 4577, 5534, 5019, 3856, 5394, 12639, 5633, 8256, 3549, 11662, 11734, 9805, 8269, 8781, 5710, 4293, 6950, 6062, 5998, 3446, 3958, 4406, 4406, 3894, 4462, 1846, 9886, 9763, 3437, 7791, 3934, 9940, 7261,...
[ 0.6688110828399658, -5.806269645690918, -0.1100572794675827, -0.035200610756874084, 0.7366737127304077, -0.1497323364019394, 0.36055871844291687, 0.09014707058668137, -0.165283665060997, -3.9384846687316895, 0.42196738719940186, -0.16906088590621948, -0.4072304666042328, -0.079797074198722...
611
24.44
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652894
[S1] ... and code and, and construct maybe architecturally different from, from the F-networks. [S2] Mm-hmm. [S1] Uh, and maybe combining this, uh, proposal networks with this, uh, checking networks, uh, may, may make a different architecture classes that could be useful.
14
3.190112
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652894.mp3
[ 8792, 12435, 8934, 8270, 9870, 6358, 3726, 4181, 2125, 4110, 9770, 8260, 7267, 9940, 12086, 9550, 9246, 8810, 10089, 8502, 10482, 7766, 3351, 9739, 8320, 2991, 7598, 7318, 8772, 4830, 7606, 8270, 12459, 3508, 9319, 7571, 3920, 9553, 5477, ...
[ 0.9038700461387634, -5.875949382781982, -0.24904504418373108, 0.11993444710969925, 0.7119415402412415, -0.1272292286157608, 0.4260317385196686, 0.14173874258995056, -0.1981106996536255, -3.9633781909942627, 0.34821978211402893, -0.10712137818336487, -0.447752445936203, -0.10954824835062027...
350
14
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652895
[S1] Yeah, I wanted to get a little bit more into, so you have, you have experimental results where you compare to various baselines, like, you know, without, um, and, and, and obviously, obviously you're better than them, which is what we've come to expect from machine learning papers. [S2] [LAUGHS] [S1] I wanna, I wanna focus a little bit on also, here you have an investigation into what the,
26.16
3.419096
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652895.mp3
[ 8745, 5201, 7370, 11213, 10674, 6519, 1461, 3688, 9266, 8747, 6526, 12205, 6349, 106, 4194, 11866, 8755, 9143, 2596, 7862, 7995, 11946, 10379, 11124, 9232, 10776, 8729, 8208, 8656, 5089, 11158, 7958, 7958, 5398, 1107, 6177, 9770, 5674, 1217...
[ 0.8947421908378601, -5.5941572189331055, -0.006880355533212423, -0.05736515298485756, 0.4877748191356659, 0.023617416620254517, 0.1917615830898285, -0.0539386086165905, -0.09207727015018463, -3.766387701034546, 0.31037652492523193, 0.037830036133527756, 0.13041190803050995, 0.0654689371585...
654
26.16
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652896
[S1] We don't know. And the sixth one, we, uh, we found that it was following, uh, blue objects very closely. So here, of course, we only show, uh, one example over time. [S2] Mm-hmm. [S1] So this is a time sequence as we track the object. On, on the appendix, we, we show that they're, it basically didn't matter. The example didn't matter. It, it reproduced very nicely. And that also gave us confidence that the G-network was learning something meaningful.
20.56
3.229515
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652896.mp3
[ 6696, 8216, 3243, 2854, 12623, 5393, 7935, 3310, 8270, 9045, 7988, 7917, 11988, 8865, 12385, 11827, 10329, 4684, 10038, 12511, 6211, 3053, 8495, 8980, 3561, 2768, 6105, 12621, 8802, 4314, 3937, 180, 7990, 11262, 9951, 11334, 9362, 4195, 109...
[ 0.7469408512115479, -5.853476524353027, -0.14090795814990997, -0.05596029385924339, 0.751301109790802, -0.1364717334508896, 0.3977026045322418, 0.12452472746372223, -0.13454671204090118, -3.9666528701782227, 0.40852949023246765, -0.1679067760705948, -0.29268166422843933, -0.037393745034933...
514
20.56
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652897
[S1] Everything is physics. If you're in the real world, um, like cars or people moving around. But, but they also, like, they also have some intrinsic mov- movement that not, doesn't follow. Passive physics loss, but, um, there's other- [S2] Do you have, do you have, like, something in mind? Like, except, except cuts between, you know, scenes. [S1] Yeah, cut, that cut, you, you'll get goodbye. [S2] Yeah. [S1] [LAUGHS] [S2] Do, do, do, do you have, do you have anything other in, is there, like, a prominent example- [S1] Yeah. [S2] ... where this type of model would, would fail?
29.24
3.006374
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652897.mp3
[ 2608, 8792, 2142, 10166, 5934, 297, 9175, 7434, 3424, 4950, 7254, 9940, 2399, 3928, 4178, 6512, 4940, 7365, 264, 853, 4868, 6812, 1289, 723, 7595, 9194, 9638, 6882, 4184, 7006, 9492, 8866, 9958, 12468, 6893, 7387, 8786, 4309, 12622, 666, ...
[ 0.6893365979194641, -5.614926815032959, 0.042998090386390686, -0.06723102182149887, 0.6423336267471313, 0.040243588387966156, 0.25403547286987305, 0.02138715796172619, -0.1199512928724289, -3.930209159851074, 0.27294161915779114, -0.09119095653295517, -0.05804353207349777, 0.06012815609574...
731
29.24
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652899
[S1] Uh, so, go ahead. [S2] One, one easy example of something that would fail is you have a video and you, uh, often have things that entered the video that were not in the video. [S1] Yeah. [S2] Um, then here you get into trouble because there's a, something that was not observed. It's the same thing that we were talking energy dissipation before. If you con- and consider the entire system, then maybe there's something that's going to get conserved if you consider heat and whatnot. [S1] Yeah. [S2] But anything that you cannot observe, then forces to- some things that are not getting observed. Um, so, yeah, extra objects, um,
29.24
3.159605
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652899.mp3
[ 3958, 4342, 6492, 10748, 10742, 7022, 8093, 2870, 3892, 6964, 6443, 3817, 11354, 1455, 3373, 4380, 4331, 7925, 7989, 12725, 12579, 10827, 11412, 10894, 8917, 4316, 9902, 2669, 8054, 11260, 8913, 10898, 8788, 4235, 11370, 12325, 12620, 11660, ...
[ 0.4923214316368103, -5.820369720458984, 0.04222998023033142, -0.08360499143600464, 0.8046947717666626, -0.1394883096218109, 0.3521970510482788, 0.12299881130456924, -0.287580668926239, -3.9160878658294678, 0.2971951961517334, -0.1567407250404358, -0.3573246896266937, 0.04400205612182617, ...
731
29.24
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652900
[S1] Yes, yes, exactly. So, yeah, things, and one other thing, I think, conversely, it could be that there's a lot of work that will need to be done if the camera is, uh, is, uh, moving a lot, um, because- [S2] And then- [S1] ... all of these objects will for sure appear that were not there, because you're looking at stuff that was not there. Uh, so if you look at the videos, this video is static.
21.72
3.152182
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652900.mp3
[ 6747, 11283, 11819, 12315, 8738, 7180, 6859, 11724, 10677, 6583, 1428, 5456, 833, 1090, 4372, 11726, 4023, 4372, 1481, 4463, 12444, 1410, 8054, 2422, 12058, 4451, 5844, 5020, 8528, 5604, 10540, 10414, 5342, 7903, 661, 8801, 9249, 8361, 7261...
[ 0.6294112205505371, -5.844748497009277, -0.09993685036897659, -0.04443734511733055, 0.6771799325942993, -0.14522503316402435, 0.34605643153190613, 0.07604585587978363, -0.1143597960472107, -3.979724884033203, 0.4002266228199005, -0.15539821982383728, -0.2601327896118164, -0.046978581696748...
543
21.72
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652901
[S1] But it's, I mean, just, just out of intuition, it seems more likely that the network detects something like, you know, there's, there's a blue bunch of pixels and, and, uh, an orange bunch of pixels, and these pixels sort of move together as objects. [S2] Yeah. [S1] Rather than the network from video somehow determining, "Aha, there's laws of physics and there's gravity and there's friction and there's sliding." It, the first situation seems a bit more likely here, right? [S2] Yes, yes.
28.48
3.357961
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652901.mp3
[ 8233, 5162, 12115, 6811, 1571, 10537, 10685, 11502, 5136, 9824, 10927, 12245, 9225, 2962, 3041, 832, 9379, 8370, 9257, 6704, 10927, 6142, 6686, 11677, 12675, 9286, 5764, 4268, 9958, 9974, 8820, 6886, 4314, 7200, 8273, 12673, 10717, 10619, 9...
[ 0.6319044828414917, -5.533860206604004, 0.08225808292627335, -0.13259358704090118, 0.4269329607486725, -0.0019388390937820077, 0.2652531862258911, -0.02685299701988697, -0.13872471451759338, -3.815322160720825, 0.3156309425830841, 0.01258399710059166, 0.06842222064733505, 0.098022297024726...
712
28.48
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652902
[S1] Sure. Uh, I didn't know exactly how. And then, um, the, Ross DeDray gave a talk at MIT, uh, it's online on the YouTube, uh, seminar. Uh, and he was talk- telling us how, um, it's very hard to encode inductive biases in neural networks. And in their case, basically, they were predicting how a robot was pushing a bunch of carrot, and the carrot was moving around, and they trained it- [S2] Mm-hmm. [S1] ... they trained a carrot predictor, uh,
27
3.271325
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652902.mp3
[ 6184, 4170, 4825, 7521, 10073, 10201, 5558, 2982, 460, 778, 6306, 9825, 7209, 9876, 8878, 3519, 886, 886, 886, 4966, 12534, 6142, 12686, 5826, 11615, 9365, 5406, 5718, 11342, 3447, 5414, 4822, 10012, 970, 8638, 4910, 12570, 3868, 4878, 39...
[ 0.5178981423377991, -5.885233402252197, 0.004738064482808113, -0.12374090403318405, 0.675106942653656, -0.06685514748096466, 0.3644954562187195, 0.1357923001050949, -0.12221521139144897, -3.9031918048858643, 0.28747910261154175, -0.2730565071105957, -0.31297799944877625, 0.0156395714730024...
675
27
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652903
[S1] Cool. Is there anything you, else you want to say about the, the experimental results? We touched on sort of upping the inner steps and the, and the, uh, the grad chem, but is there anything- [S2] We- [S1] ... special you want to say about sort of your, your tests on, uh, for example, the pendulums or-
17.52
3.381934
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652903.mp3
[ 11922, 7345, 11476, 5033, 5221, 7819, 10466, 5221, 3154, 7211, 8801, 6805, 12534, 3181, 4716, 3306, 7200, 7729, 5633, 4646, 12701, 3529, 3009, 10739, 8092, 8164, 11996, 11611, 9539, 6568, 2992, 12180, 11797, 9940, 11716, 7837, 7894, 3295, 3...
[ 0.8053982853889465, -5.765412330627441, -0.01784892939031124, 0.016501566395163536, 0.41396841406822205, 0.025999844074249268, 0.16099964082241058, -0.09079333394765854, -0.05853462964296341, -3.7169065475463867, 0.26951244473457336, -0.068867526948452, 0.030464189127087593, 0.008943463675...
438
17.52
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652904
[S1] Yeah, I think some of the experiments, uh, depends on the, how much time we have, but on the, on the pendulum, there was a symbolic component, so the, the G doesn't have to be fully neural. [S2] Yeah. [S1] Uh, so in, in the origin, in the first, I think those are the first experiment, the G is kind of a program with some parameter. [S2] Yeah. [S1] Uh, like a formula, and there we search over formulas, uh, because it's a state information, the pendulum that you draw, like the angle and the momentum.
26.16
3.166793
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652904.mp3
[ 9810, 3589, 6402, 11219, 8114, 11244, 10004, 6560, 2856, 11617, 11600, 11355, 11870, 4749, 6225, 9298, 7371, 3611, 1170, 4064, 3545, 7990, 11437, 7779, 8500, 5219, 81, 2980, 11110, 11110, 5991, 5998, 5999, 5999, 5487, 6070, 1820, 6818, 9377...
[ 0.5962957143783569, -5.826827049255371, -0.009019976481795311, -0.054559897631406784, 0.6371845006942749, -0.06325756013393402, 0.35271942615509033, 0.1084279865026474, -0.170784592628479, -3.8984322547912598, 0.32036352157592773, -0.21627023816108704, -0.346279501914978, 0.017285477370023...
654
26.16
bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652905
[S1] And there you, we search over formulas, um, and then there's some parameters as well that get trained over, uh, with gray in the center. [S2] Yeah. [S1] And there we saw that, okay, we, we are able to recover the true formulas of the energy and it leads to better prediction than a vanilla MLP that does not learn about- [S2] Yeah. [S1] ... observations. Um- [S2] And there also you can see that, that actually you can, you can even handle these approximate constraints where you have real data which then the networks that have the hard-coded constraints can't handle as well.
29.88
3.230198
24,000
audio/en/bilibili_data_1624174493_BV17h4y1j7aZ_p164_BV17h4y1j7aZ_p164_m4-dialogue_0652905.mp3
[ 3691, 9361, 4837, 4975, 5647, 3650, 9709, 11251, 5533, 6859, 9484, 3212, 2740, 2676, 10118, 5088, 8528, 3547, 7982, 2885, 8330, 12689, 8038, 7782, 2675, 8494, 3459, 7408, 5498, 756, 8311, 5523, 8284, 5236, 8932, 8467, 11122, 11562, 4522, ...
[ 0.6508334875106812, -5.683277130126953, 0.07212638109922409, -0.19087669253349304, 0.5142034292221069, -0.032329898327589035, 0.2462475746870041, 0.02678028494119644, -0.190087229013443, -3.8605897426605225, 0.28105399012565613, -0.15421684086322784, -0.07265264540910721, 0.103970505297183...
747
29.88