Variational Discriminator Bottleneck
Download as PPTX, PDF0 likes178 views
The document proposes constraining the information flow in variational discriminators by adding an information bottleneck. This forces the discriminator to focus only on essential information needed to determine if a sample is real or generated, rather than nonessential details. Experiments show this stabilized training for generative adversarial networks (GANs), improved imitation learning, and helped with inverse reinforcement learning tasks by focusing the discriminator on high-level features.
1 of 13
Download to read offline
Variational Discriminator Bottleneck
- 1. VARIATIONAL DISCRIMINATOR BOTTLENECK: IMPROVING IMITATION LEARNING, INVERSE RL, AND GANS BY CONSTRAINING INFORMATION FLOW Yawei Luo
- 2. Notoriously D & G D can always find out the nonessential information from G(z) to make a judgement “fake”. -> Uninformative gradients -> Unstable training! How to force D to focus on essential information of G(z)?
- 3. Preliminaries • Mutual Information • Object function in information theoretic view
- 7. Back to GANs
- 8. Back to GANs Vanilla GAN: GAN with VIB:
- 9. Training I(X, Z) > Ic -> beta ++ I(X, Z) < Ic -> beta --
- 10. Experiments - IMITATION LEARNING
- 11. Experiments - IMITATION LEARNING
- 12. Experiments - INVERSE REINFORCEMENT LEARNING
- 13. Experiments – image generation