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

Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

youtube.com

Post reply on HN
Print view View original

51–60 of 68 posts

Re: Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

#51

Post by letlambda »

yorwba wrote: post #49
IshKebab wrote: post #30

Earlier quoted context omitted.

I would say that it probably is. If you trained a network using illegal data (e.g. cp images) then not only did you have to have that data once, which is of course illegal, but the data itself is at least partially encoded in the network weights which I think should make it illegal.

I guess you might be able to get around this if you train only on legal content and can interpolate into content that would be illegal if a real recording. However, I'm not sure whether there are any other applications for this specific interpolation scenario that would lead to it being developed, as the effort required to make it work is likely much higher.

Having the model produce realistic interpolations through areas of the latent space that had no associated training data is surely something that people will be trying to make happen.

Re: Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

#52

Post by pault »

duskwuff wrote: post #12

I love how using training data from the Internet has resulted in the GANs believing that a picture of a "cat" often contains text -- or, at least, shapes resembling text -- at the top and bottom of the image. (Visible on the right side of the screen starting around 4:20.)

The interpolating animals look so much like my dreams it's unsettling. I often wake up when things get too "unrealistic" and people or animals are changing shape too rapidly. Whenever I see something come out of a NN that looks like something that came out of my brain I get a little future shock.

Re: Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

#53

Post by platz »

pault wrote: post #52
duskwuff wrote: post #12

I love how using training data from the Internet has resulted in the GANs believing that a picture of a "cat" often contains text -- or, at least, shapes resembling text -- at the top and bottom of the image. (Visible on the right side of the screen starting around 4:20.)

The interpolating animals look so much like my dreams it's unsettling. I often wake up when things get too "unrealistic" and people or animals are changing shape too rapidly. Whenever I see something come out of a NN that looks like something that came out of my brain I get a little future shock.

Did you see the the horror faces it generated for humans? Good for Halloween!

Re: Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

#55

Post by superfx »

jaakkol wrote: post #50
argonaut wrote: post #29

Worth pointing out the main criticism of GANs, which is that right now researchers don't really have a way to tell if a GAN is just copying and pasting the training data or not (there is no "test set" unlike in supervised learning). And in fact an ideal GAN could just learn to output the training set. One example someone found in the generated images for this model: https://twitter.com/nalkalchbrenner/status/92401333…

An GAN that just made verbatim copies from the training set would not be able to smoothly interpolate in latent space. Also, see Fig. 10 in the paper at http://research.nvidia.com/publication/2017-10_Progressive-G... .

But they seemed to have picked closest neighbors in pixel space instead of z-space, which is not the best idea, no?

Re: Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

#56

Post by ericjang »

argonaut wrote: post #29

Worth pointing out the main criticism of GANs, which is that right now researchers don't really have a way to tell if a GAN is just copying and pasting the training data or not (there is no "test set" unlike in supervised learning). And in fact an ideal GAN could just learn to output the training set. One example someone found in the generated images for this model: https://twitter.com/nalkalchbrenner/status/92401333…

This is a general problem for generative models, regardless of whether they are explicit latent density (variational autoencoders) or implicit density (GANs) or fully-observed (PixelCNN). To date no consensus has been reached in the research community on what the "right" way to evaluate a generative model is: e.g. Parzen Windows estimators, log-probability, ELBO, etc.

Yann Lecun himself (as of NIPS2016) was pretty critical of probability-based metrics, as those have strong dependence to the choice of model (e.g. if the model is poor the log-probability is meaningless).

In GANs, the critic and the generator are trained w.r.t. each other, reaching some kind of equilibrium. A recent proposal that seems to be "ok" for evaluating GANs was proposed by https://arxiv.org/abs/1705.05263, which is to train a separate critic on the generator, for use in evaluation (the generator never sees gradient information from this critic). This evaluation critic approximates the Wasserstein distance. One could imagine actually training the independent critic on a validation set of images not seen by the training set.

Re: Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

#57

Post by subcosmos »

subcosmos wrote: post #24
avaer wrote: post #13

Wonder how long until this is co-opted by the porn industry and what the law will have to say about it. Is it illegal to own a digital brain that can think up illegal porn?

Note to all: Please don't do this. Lots of missing kids out there and the agents working in this area don't need anymore distractions.

Love how this was downvoted. Way to stay sane HN... I don't think that any conversations focused at further promoting child abuse imagery should be encouraged in this community, particularly when such misapplication of these technologies would profoundly derail existing investigations.

But fuck it. Downvote away!

Re: Progressive Growing of GANs for Improved Quality, Stability, Variation [video]

#58

Post by subcosmos »

has2k1 wrote: post #44
goombastic wrote: post #6

Interesting. In another 300-500 years, I am pretty sure we will start simulating sensory experience and ultimately the past. I am not sure if I am a toy simulation of the past from the future, right now.

> In another 300-500 years instead of "In the future", triggers me. Do you know something we do not? I am resigned to the notion that everything is up for grabs. > I am not sure if I am a toy simulation We need "reality discriminants". The trippy thing is if they could exist and their output is not necessarily boolean. There would be a threshold point at which beings can exist along the simulated-real spectrum, were…

The way these GAN algorithms work is precisely by building a network that discriminates reality from fake. That's why they become so good at this!
Post reply on HN