0
$\begingroup$

In pix2pix GAN paper( https://arxiv.org/abs/1611.07004), authors found that the noise vector and the dropout are not efficient in grasping the full entropy of the data distribution we want to model. The model learns during the training in a manner to ignore the noise provided by the noise vector and the dropout adds only minor stochasticity in the output.

What is the right method to grasp the full entropy of the data distribution we want to model?

$\endgroup$
1
$\begingroup$

This is the mode-collapse problem. There are a few ways to tackle this

  1. Use mode-collapse avoiding techniques for GAN models. For example Diversity-Sensitive Conditional Generative Adversarial Networks

  2. Use conditional generative models which explicitly maximize likelihood (VAEs, flows, autoregressive, etc), because they tend to be less vulnerable to this problem.

$\endgroup$
  • $\begingroup$ Could your answer be more detailed so that I tick it as the right answer $\endgroup$ – ProEns08 Oct 21 at 11:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.