Message Boards Message Boards

1
|
2554 Views
|
0 Replies
|
1 Total Likes
View groups...
Share
Share this post:

NetGANOperator seemingly converging to a fully biased generator

Posted 2 years ago

Hi,

I was experimenting with NetGANOperator and the training seemed to work fine as the loss function results were evolving more or less as I would have expected.

However, when training was done, I tried to generate images from a few latent vectors and I always got the same result (visually anyway). That seemed odd to me and I thought the only way the result could not depend on the input would be that all weights had converged towards zero, leaving only the biases.

Then I wondered : isn't that a trivial solution for a GAN? The generator would always produce the same output, and the discriminator only has to learn how to recognize that particular output.

I understand this is not a question specific to Mathematica but rather to machine learning in general but I thought I could get some explanations from this community.

Here is the notebook I was working on:

POSTED BY: Lucien Grondin
Reply to this discussion
Community posts can be styled and formatted using the Markdown syntax.
Reply Preview
Attachments
Remove
or Discard

Group Abstract Group Abstract