Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Improved Techniques for Training GANs – OpenAI's first paper

arxiv.org

Post reply on HN
Print view View original

1–10 of 14 posts

Re: Improved Techniques for Training GANs – OpenAI's first paper

#4

Post by nl »

So this is pretty interesting.

One of the great hopes of the current deep learning boom is that somehow we will develop unsupervised or at least semi-supervised techniques which can perform close to the great results that are being seen with supervised learning.

Adversarial Networks is one of the more likely routes to semi-supervised. There is also a lot of interesting work in combining Bayesian optimization techniques with Deep Networks to develop one-shot learning[1][2]. Some of this was (very broadly) in response to the the one-shot learning paper coming out of (I've forgotten!!) where the authors are famously doubtful about the utility of Deep Learning, and showed somewhat competitive results on MNIST. (I can't remember who it was - there have been HN discussions about the group. Sorry!!)

Both OpenAI and DeepMind are following roughly similar paths here (no surprise really), and the results are looking really good.

[1] http://arxiv.org/abs/1603.05106

[2] http://arxiv.org/abs/1606.04080

Re: Improved Techniques for Training GANs – OpenAI's first paper

#5

Post by nl »

imh wrote: post #3

When they find the visual turing test results important enough to put in the abstract, it's a shame they only include tiny images in the paper :(

They are the full size images. CIFAR-10 is 32x32 colour images: https://www.cs.toronto.edu/~kriz/cifar.html

Re: Improved Techniques for Training GANs – OpenAI's first paper

#6

Post by mcadenhe »

nl wrote: post #5
imh wrote: post #3

When they find the visual turing test results important enough to put in the abstract, it's a shame they only include tiny images in the paper :(

They are the full size images. CIFAR-10 is 32x32 colour images: https://www.cs.toronto.edu/~kriz/cifar.html

If they used another dataset with images of larger dimensions could they generate larger and less blurry images?

Re: Improved Techniques for Training GANs – OpenAI's first paper

#7

Post by T-A »

nl wrote: post #4

So this is pretty interesting. One of the great hopes of the current deep learning boom is that somehow we will develop unsupervised or at least semi-supervised techniques which can perform close to the great results that are being seen with supervised learning. Adversarial Networks is one of the more likely routes to semi-supervised. There is also a lot of interesting work in combining Bayesian optimization techniqu…

> the one-shot learning paper coming out of (I've forgotten!!) where the authors are famously doubtful about the utility of Deep Learning

This? https://www.technologyreview.com/s/544376/this-ai-algorithm-...

http://cims.nyu.edu/~brenden/LakeEtAl2015Science.pdf

Re: Improved Techniques for Training GANs – OpenAI's first paper

#8

Post by romaniv »

nl wrote: post #5
imh wrote: post #3

When they find the visual turing test results important enough to put in the abstract, it's a shame they only include tiny images in the paper :(

They are the full size images. CIFAR-10 is 32x32 colour images: https://www.cs.toronto.edu/~kriz/cifar.html

This is one of many examples why claims regarding near- or super- human performance in AI papers need to be taken with a good amount of salt. CIFAR-10 is great for experimenting with your algorithms, but it's a horrible dataset to do any kinds of human-to-machine performance comparisons.

Re: Improved Techniques for Training GANs – OpenAI's first paper

#9

Post by gwern »

mcadenhe wrote: post #6
nl wrote: post #5

Earlier quoted context omitted.

They are the full size images. CIFAR-10 is 32x32 colour images: https://www.cs.toronto.edu/~kriz/cifar.html

If they used another dataset with images of larger dimensions could they generate larger and less blurry images?

A DCGAN as usually implemented (eg Soumith's Torch DCGAN implementation) can produce arbitrarily large images by upscaling. The quality won't be good, though, unsurprisingly, because it was only trained on 32px or less images. This also means that it's hard to evaluate DCGAN improvements because you're stuck squinting at 32px thumbnails trying to guess whether one blur looks more semantically meaningful than another blur.

One nice thing about this paper is that, as I've been suggested for a while, they up the input to 128px for the Imagenet thumbnails, and if you look at those, it immediately pops out that while the DCGAN has in fact successfully learned to construct vaguely dog-like images, the global structure has issues.

Post reply on HN