Generalized Energy-Based Models
summary
Paper argues that GAN critics contain reusable density information and turns them into a generalized energy-based model that learns an energy E and normalizer C via a smooth variational lower bound (a KL-smoothie implemented in an RKHS). Training alternates critic updates and generator updates; sampling shifts the latent Z distribution with MCMC/HMC instead of naive importance weights. Empirical results report consistent FID improvements over the generator-alone baseline; code and a concurrent Jensen–Shannon variant are released.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:38Training generative models
- 2:08Outline
- 5:24The Integral Probability Metrics
- 6:20Maximum mean discrepancy
- 8:23A variational lower bound
- 11:15Case of the KL
- 12:04Empirical properties of KALE
- 17:04Topological properties of KALE (2)
- 23:24Generalized Energy-Based Models - the idea
- 29:19KALE and the energy function
- 31:18Training the base measure (generator)
- 34:49Sampling from the model
- 37:15Examples: sampling at modes
- 37:55Examples: moving between modes
- 42:51Summary
- 43:59Questions?
related talks
01h 00mTutorial on Deep Learning IV
11mNonparametric Density Estimation and Convergence of GANs under Besov IPMs
44mRobust Estimation and Generative Adversarial Nets
40mTraining generative models from locally privatized data via entropic optimal transport
48mUnveiling the secrets of a machine learning model
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)
52mDeep Generative Models And Unsupervised Methods For Inverse Problems
44mChoose the right neural generative model for your problem