adversarial fisher vectors for unsupervised
play

Adversarial Fisher Vectors For Unsupervised Representation Learning - PowerPoint PPT Presentation

Adversarial Fisher Vectors For Unsupervised Representation Learning Shuangfei Zhai, Walter Talbott, Carlos Guestrin, Joshua M. Susskind Apple Inc. Questions about GANs Is the discriminator useful at test time? Do GANs learn


  1. Adversarial Fisher Vectors For Unsupervised Representation Learning Shuangfei Zhai, Walter Talbott, Carlos Guestrin, Joshua M. Susskind Apple Inc.

  2. Questions about GANs • Is the discriminator useful at test time? • Do GANs learn representations of data? • Do you need to train an additional encoder?

  3. Energy Based Model Interpretation of GANs • The WGAN formulation max G min D E x ∼ p data ( x ) [ − D ( x )] + E z ∼ p ( z ) [ D ( G ( z ))] (1) • EBM with variational training has a dual form to a WGAN e D ( x ) min D max G E x ∼ p data ( x ) [ − D ( x )] + E z ∼ p ( z ) [ D ( G ( z ))] + Entropy ( p G ), s . t . p ( x ) = (2) ∫ x e D ( x ) dx • Equation (1) and (2) can amount to the same practical implementation! Generative Adversarial Networks as Variational Training of Energy Based Models, Zhai et. al.

  4. Fisher Vectors • Fisher vectors provide a way to represent an example given a probabilistic model V x = I − 1 2 ∇ θ log p θ ( x ), s . t . , I = E x ∼ p θ ( x ) [ ∇ θ log p θ ( x ) ∇ θ log p θ ( x ) T ] • Has seen successful applications in computer vision image SIFT descriptor Gaussian Mixture Model FV Exploiting Generative Models In Discriminative Classifiers, Jaakkola and Haussler

  5. Adversarial Fisher Vectors • Step 1: train a GAN and treat it as an EBM • Step 2: compute the Adversarial Fisher Vector via: V x = ( diag (I) − 1 2 ) U x s . t . U x = ∇ θ D ( x ; θ ) − E z ∼ p ( z ) ∇ θ D ( G ( z ); θ ), I = E z ∼ p ( z ) [ U G ( z ) U T G ( z ) ] V x • Step 3: use as the representation for downstream tasks (e.g., classification)

  6. State-of-the-art Results on Linear Classification

  7. Checkout the Paper and Code!

Recommend


More recommend