Deep convolutional nets have brought about breakthroughs in processing images, video, speech and audio.
Unresolvable The public record could not settle it.
Factual: Statements about the present or the past, checked against the record: did they hold up?
What would count as true or false
True if, by May 2015, the literature shows deep convolutional nets producing breakthrough (large, state-of-the-art) advances in all four areas: images (for example ImageNet 2012), video (for example action recognition), speech (for example CNN acoustic models) and audio. Mostly true if the breakthroughs are well documented for images and speech but the advances in video or audio, while real, were more modest. Mixed if only about half of the four areas had breakthroughs attributable to ConvNets. Mostly false if a breakthrough is documented only for images. False if there were no ConvNet-driven breakthroughs.
Set before the research began.
Evidence (9)
-
E1 Two-Stream Convolutional Networks for Action Recognition in Videos Proceedings of NeurIPS (Karen Simonyan and Andrew Zisserman, Visual Geometry Group, University of Oxford)
“It also exceeds by a large margin previous attempts to use deep nets for video classification.”
-
E2 Two-Stream Convolutional Networks for Action Recognition in Videos Proceedings of NeurIPS (Karen Simonyan and Andrew Zisserman, Visual Geometry Group, University of Oxford)
“Our architecture is trained and evaluated on the standard video actions benchmarks of UCF-101 and HMDB-51, where it is competitive with the state of the art.”
-
E3 Two-Stream Convolutional Networks for Action Recognition in Videos Proceedings of NeurIPS (Karen Simonyan and Andrew Zisserman, Visual Geometry Group, University of Oxford)
“The learnt representation, finetuned on the UCF-101 dataset, turned out to be 20% less accurate than hand-crafted state-of-the-art trajectory-based representation [20, 27].”
-
E4 Two-Stream Convolutional Networks for Action Recognition in Videos Not shown on page; NeurIPS proceedings per the URL
“Our architecture is trained and evaluated on the standard video actions benchmarks of UCF-101 and HMDB-51, where it is competitive with the state of the art. It also exceeds by a large margin previous attempts to use deep nets for video classification.”
-
E5 Two-Stream Convolutional Networks for Action Recognition in Videos Not shown on page; NeurIPS proceedings per the URL
“The combination of the two nets further improves the results (in line with the single-split experiments above), and is comparable to the very recent state-of-the-art hand-crafted models [20, 21, 26].”
-
E6 Two-Stream Convolutional Networks for Action Recognition in Videos Not shown on page; NeurIPS proceedings per the URL
“but the results were significantly worse than those of the best hand-crafted shallow representations [20, 26].”
-
E7 Two-Stream Convolutional Networks for Action Recognition in Videos Not shown on page; NeurIPS proceedings per the URL
“The learnt representation, finetuned on the UCF-101 dataset, turned out to be 20% less accurate than hand-crafted state-of-the-art trajectory-based representation [20, 27].”
-
E8 Convolutional Neural Networks for Speech Recognition IEEE/ACM Transactions on Audio, Speech, and Language Processing (authors from Microsoft Research and others) · October 2014
“Experimental results show that CNNs reduce the error rate by 6%-10% compared with DNNs on the TIMIT phone recognition and the voice search large vocabulary speech recognition tasks.”
-
E9 End-to-end learning for music audio Ghent University library (UGent biblio), recording the ICASSP 2014 paper by Sander Dieleman and Benjamin Schrauwen · 2014
“Although they do not outperform a spectrogram-based approach, the networks are able to autonomously discover frequency decompositions from raw audio, as well as phase-and translation-invariant feature representations.”
Reasoning
For speech, E8 reports that CNNs cut error rates by 6%-10% compared with DNNs on TIMIT and voice search, which is a real advance. For video, the two-stream ConvNet was only 'competitive with' or 'comparable to' the hand-crafted state of the art (E2, E4, E5), and earlier deep video models did worse (E6, E7). For audio, raw-audio ConvNets did not beat a spectrogram-based approach (E9). That rules out 'true', but there is no evidence on images, and the choice between 'mostly true', 'mixed' and 'mostly false' depends on that missing area, so the claim can't be settled.
Cite this claim
Yann LeCun, “Deep convolutional nets have brought about breakthroughs in processing images, video, speech and audio.” (said 27 May 2015). Verdict: unresolvable. Reliability API, claim c_87061d7ab10f, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_87061d7ab10f
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Yann LeCun
-
Mostly true
“For starters, transformers are equivariant to permutations.”
Transformers are equivariant to permutations, as ConvNets are equivariant to translations.
-
Mostly true
“If it were the case, we would have AI systems that could teach themselves to drive a car in 20 hours of practice, like any 17 year-old.”
-
Mostly true
“The architecture is good, the architecture of the encoder is good, but the fact that you train the system to reconstruct images does not lead it to produce to long, good generic features of images.”
-
Mostly true
“LLMs can do none of those or they can only do them in a very primitive way and they don’t really understand the physical world.”
-
Mostly true
“No, there’s one thing that autoregressive LLMs or that LLMs in general, not just the autoregressive one, but including the bird style bidirectional ones, are exploiting and its self supervised…”
Verdicts are made automatically; how verdicts are made.