Hacker News Reader

Comment by CShorten | Hacker News Reader

I made a video explaining the details of this paper if you are interested! https://youtu.be/H8F6J7mYyz0

Very interesting video! How does the GameGAN image discriminator work? I don't know so much about GANs but it seem to me that classifying PacMan images as real or fake would be a very hard problem because of the low resolution.

Suppose PacMan is in a corridor and there is a dot to the left of him and three to the right: ".P..." That is plausible because PacMan hasn't eaten any dots yet. But if the second dot to the right is missing: ".P._." something is wrong because the first dot to the right is still there and PacMan can't jump over dots. It must be very hard to get the discriminator to understand that the second image is fake.

Thanks for watching bjourne! There are three different discriminators, the first does realism for the game, the second does action-conditioned discrimination (making sure the generator takes in the action for its next frame prediction), and the third is a temporal realism with a 3D CNN.

Interesting scenario, it would be a cool experiment to walk through the latent space and try to find that kind of frame in the GAN latent space. I'm sure you could find it, but the question might be how far off the generated data's manifold is it.

Thanks for doing this. I stumbled upon one of your previous videos a while ago and really liked it. I'm super indebted to YouTubers like yourself that take the time to distill and explain complex topics, papers, and new research.

Thank you so much!

Reply on news.ycombinator.com