Highly misleading descriptions in some cases. The most recent one on embedding, the advances are real... they just come from deeper better NNs, not the fancy loss functions. Embarrassing for the researchers, yet, calling them 'not real' is false. And while http://papers.nips.cc/paper/7350-are-gans-created-equal-a-la... was interesting, it's extremely obsolete: the most recent NN considered is WGAN and the hardest dataset is CIFAR-10; that DCGAN can - with enough sheer brute force and retrying enough times - sometimes match WGAN on 28px (Fashion-MNIST) or 32px (CIFAR-10) images is mildly interesting, but pretty much irrelevant now, as stability is very important and no one in their right mind would claim that DCGAN could, say, outperform StyleGAN 2 on 1024px FFHQ or BigGAN on 512px ImageNet if only you tried DCGAN hard enough.
This article is biased as it only cites those couple of fields or problems where neural nets have not made significant improvements (even that part is shady as I am not familar with that much).
There have been significant improvements in the architecture of NNs in the last decade unlike what this article claims. This article doesn't mention anything about GANs or Autoencoders or VAEs or Flow or recent NLP advancements which are clearly way way better than what we have seen 5 years ago. Seems like the authors are living in 2015.
You are complaining how the article doesn't mention GANs while replying to a comment about article's treatment of GANs. Revealing you are biased against even reading the biased article, or comprehending the parent comment.
>> This article doesn't mention anything about GANs or Autoencoders or VAEs or Flow or recent NLP advancements which are clearly way way better than what we have seen 5 years ago.
The whole point of Bousquet study was comparability, so it had to be the lowest common denominator, especially in a field that defines SotA differently year to year in each wave of research. Your objection wouldn't embarrass anyone.
As I said, that is not what anyone reading this article would take away. Anyone reading an article littered with misleading statements like "and some of the gains may not exist at all" (repeated twice!) or "Eye-catching advances in some AI fields are not real" or "phantom progress" is not going to conclude, "actually, systems do perform much better than years ago and are increasingly powerful and relevant to the real world, it's just the performance gains do not come from the claimed sources". A reader is going to take away the conclusion "NNs are just as dumb and useless as they were 5 or 10 years ago in many fields, lol AI hype" - which is the opposite of reality.
8 comments
[ 3.2 ms ] story [ 35.8 ms ] threadA result about GANs is prominent in the article.