Hacker News new | past | comments | ask | show | jobs | submit login

Are GAN's pretty hot these days or is that just a coincidence of the authors preferences? I just received Ian Goodfellow's Deep Learning textbook [1] and I know he pretty much invented the technique, so I'm wonder how influential / important GANs are in the field.

[1] http://www.deeplearningbook.org/

GANs are for graphics, not machine learning. They have no test scores as they cannot run on the test set. Therefore, its hard to tell how much they overfit.

But they have uses, like for lossy compression of images or texture generation. Stuff which focuses on the graphical side of things where it outperforms machine learning methods with its crisper samples.

For an overview of this argument: https://arxiv.org/abs/1511.01844

The idea of adversarial training is important and relevant in ML though! It allows for setting up losses which are hard to formulate otherwise.

Huh? GANs are for graphics, not machine learning? I mean, first of all, how are GANs not machine learning? They are definitely "machine learning", and I don't think anyone would disagree.

Second of all, GANs are most definitely not just for graphics. They've been applied to text generation, to generating adversarial examples, to data preprocessing, etc.

Third, I have no idea what you even mean by "test set" in the context of GANs. It is true that it's hard to tell their performance, but that's irrespective of whatever you're talking about. It's hard to evaluate performance because we're usually judging the quality of the generated images, and we don't have any good ways of evaluating "perceptual loss", or how real an image looks.

As for the OP, GANs have been a very hot topic. Not as hot as this blog post makes them look perhaps (with nearly every paper about them...), but I wouldn't really disagree with any of the papers posted. Only one I'm not familiar with is the "most useful" one, but the rest were all pretty great papers imo. As for Ian Goodfellow, he's a very smart guy who seems to do a pretty good job explaining things. I saw a couple YouTube videos from him at a meetup covering his DL book, and he did a great job teaching.

Although I would agree that GANs are part of machine learning, some people definitely do disagree, and their concerns are valid. It's definitely an area of open research.

Your third point is actually the point of those who disagree. It's the same reason why we have the principle of unfalsifiability in science.

I'm a bit confused about the point that you're stating. I've never seen anybody not group GANs under machine learning.

Machine learning is typically split into supervised learning, unsupervised learning, and reinforcement learning, and GANs are usually considered part of unsupervised learning. I guess the part I don't understand is what you mean by "their concerns are valid"? What are their concerns about? Whether GANs are a promising path of research? And if GANs aren't part of machine learning what are they?

The difficulty of evaluating GANs is nowhere near the level of unfalsifiability, and it is not caused by GANs themselves being a bad technique, but by the problem space they are applied in.

When you are trying to generate "realistic" samples of human concepts, the ultimate measure of evaluation is whether humans think that the output is realistic. So you have no choice but to ask humans to judge the quality of your results. That's a standard thing to do e.g. in text-to-speech generation, whether GANs are used or not.

I don't think GANs are being used to do anything very serious at the moment (I would love to be corrected on this), but they're very exciting in terms of their capabilities and promise. Some good progress has been made over the last year in making them easier to train and making them so they produce varied outputs

WaveNet2 is using a variant of GANs to produce the most high fidelity speech (text->to->voice) of any computer system out there - in realtime. WaveNet from last year took 50 sec (i think) to generate 1 sec of audio. By using a variant of a GAN to re-engineer WaveNet, they can now produce real-time high quality text-to-voice: https://deepmind.com/documents/131/Distilling_WaveNet.pdf

WaveNet2 is not a GAN, and describing it as a "variant of GANs" is like calling Python a variant of Java: it's highly misleading (WaveNet2 and GANs are both generative models, Python and Java are both programming languages).

Also WaveNet2 makes no improvements in the actual quality of the model, only the run-time performance.

Yep, from the paper:

> It is worth noting the parallels to Generative Adversarial Networks (GANs [7]), with the student playing the role of generator, and the teacher playing the role of discriminator. As opposed to GANs, however, the student is not attempting to fool the teacher in an adversarial manner; rather it cooperates by attempting to match the teacher’s probabilities. Furthermore the teacher is held constant, rather than being trained in tandem with the student, and both models yield tractable normalised distributions.

WaveNet2's main resemblance to a GAN is that it uses another neural network for the loss function.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact