This tweet is unavailable
The discriminator does provide per-pixel correction signals for the generator. If some regions look generated, D will output "fake" more confidently. Updating G to fool D will correct these regions (also taking the context into account to the extent that D recognizes it).
2
0
0
0
Replying to @mere_mortise @gwern
Assuming a CNN, many spatial predictions by D would basically implement a bagging scheme: You would not only look once at the input image, but multiple times from shifted positions. This would make the D stronger. Not sure that would help since D is prone to overfitting anyway.

11:48 PM · Nov 26, 2017

1
0
0
0
You’re unable to view this Tweet because this account owner limits who can view their Tweets.
You’re unable to view this Tweet because this account owner limits who can view their Tweets.