I know there are multi-scale or patch-specific GANs, but the losses there still seem to be averaged out. A D is 'looking' at all pixels already, so it doesn't necessarily seem like that much extra work to produce additional labels/losses?
-
Show this thread
-
Like, when I look at GAN samples, usually there's some specific region which looks bad, typically much less than 1/4th of the total pixels, and the rest are fine. GANs aren't *really* RL so why provide them impoverished RL-style global losses?
1 reply 0 retweets 3 likesShow this thread -
Replying to @gwern
I must not understand this proposal. Why wouldn't you get the same probability at all pixels? What distinguishes one pixel from another in the loss function for D?
1 reply 0 retweets 1 like -
Replying to @ESYudkowsky
The D usually emits a single scalar loss for the whole image, eg 0-3 for a WGAN. My suggestion is that it could instead emit for, say, a 64x64px image, a 64x64 array filled with 0-3 scalars, explaining how each pixel contributes to the global average loss of 2.2234.
2 replies 0 retweets 0 likes -
Replying to @gwern
Why isn't that already automatically implicit in doing backprop from D's output to D's input pixels / G's output? How would you train D to do that in any other way?
1 reply 0 retweets 0 likes -
Replying to @ESYudkowsky
It is implicit but that doesn't mean equivalent. The signal could be a lot noisier. You can train AG with MCTS supervision, all of which feedback is 'implicit' in win/loss feedback, but the latter is still much slower/unstabler.
1 reply 0 retweets 0 likes -
Replying to @gwern
Did you have a particular proposal in mind? What came to my mind after 30sec was training G to complete arbitrary fractions of images including 100%, with both contiguous and discontiguous regions replaced, and then training D to judge pixel probabilities.
1 reply 0 retweets 0 likes -
Replying to @ESYudkowsky
I hadn't thought about how exactly you'd train the D itself (rather than using it to train the G). Yes, inpainting regions of a real image would be one way... or noise/shuffle random pixels in just real images. Or assign all pixels in fake/real images arbitrary low/high constants
1 reply 0 retweets 0 likes -
Replying to @gwern
The last idea gets you constant predictions for all pixels, I expect. Training D this way seems like the obvious hard part of the problem, requiring cleverness and probably not working.
1 reply 0 retweets 0 likes -
Replying to @ESYudkowsky
Possibly, yeah. On the other hand, REINFORCE and GANs don't seem like they should work either with crude average global feedback (and often don't). But it seems like getting richer supervision out of Ds is something obvious which someone should've tried but no one has...
2 replies 0 retweets 0 likes
I've seen "train D to output probabilities, but train G to match activations in pre-final layer(s) of D". Played myself with training D to estimate distance from image to reconstructed image in a VAEGAN, didn't help (when I clumsily tried) even though it's a more pixelwise idiom.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.