[Paper Review] f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization

Sebastian Nowozin, Botond Cseke+1 more|arXiv (Cornell University)|Jun 2, 2016

Adversarial Robustness in Machine Learning29 references636 citations

TL;DR

The paper shows that GAN training is a special case of variational divergence minimization over f-divergences, and generalizes generative neural samplers to be trained with any f-divergence using a variational discriminator.

ABSTRACT

Generative neural samplers are probabilistic models that implement sampling using feedforward neural networks: they take a random input vector and produce a sample from a probability distribution defined by the network weights. These models are expressive and allow efficient computation of samples and derivatives, but cannot be used for computing likelihoods or for marginalization. The generative-adversarial training method allows to train such models through the use of an auxiliary discriminative neural network. We show that the generative-adversarial approach is a special case of an existing more general variational divergence estimation approach. We show that any f-divergence can be used for training generative neural samplers. We discuss the benefits of various choices of divergence functions on training complexity and the quality of the obtained generative models.

Motivation & Objective

• Motivate and formalize training generative neural samplers via divergences beyond Jensen-Shannon/GAN.
• Introduce variational divergence minimization (VDM) as a general framework for learning Q_theta to approximate P.
• Derive f-GAN objectives for a broad family of f-divergences and show how to implement them with neural nets.
• Clarify practical optimization methods and activation designs to handle different f-divergences.
• Demonstrate empirical effects of divergence choices on image data and model performance.

Proposed method

• Review f-divergences and their variational representations via convex conjugates f*.
• Formulate the f-GAN saddle-point objective F(theta, omega) = E_{x~P}[T_omega(x)] - E_{x~Q_theta}[f*(T_omega(x))].
• Represent the variational function as T_omega(x) = g_f(V_omega(x)) with appropriate output activations to match dom(f*).
• Show that GAN/JS is a special case of the general F(θ, ω) objective with a specific activation and T*.
• Propose a practical single-step gradient method to optimize the saddle point, updating both generator and variational function in one pass.
• Discuss training tips such as alternative generator updates, real/fake statistics, and optimization aids (Adam, gradient clipping).

Experimental results

Research questions

• RQ1 Can any f-divergence be used to train generative neural samplers via a variational objective?
• RQ2 How does the choice of f-divergence affect training dynamics and the quality of learned distributions?
• RQ3 What are the practical considerations for implementing the variational function across divergences?
• RQ4 Does a single-step optimization approach converge to a saddle point under reasonable conditions?
• RQ5 How do divergences influence results on real image datasets like MNIST and LSUN?

Key findings

• Any f-divergence yields a valid training objective for generative samplers when coupled with a variational discriminator.
• GAN training is recoverable as a special case of the broader f-GAN/VDM framework.
• A practical single-step gradient method can converge to a saddle point under mild smoothness and convexity assumptions.
• Different divergences lead to different learned distributions when the model is misspecified, highlighting divergence impact on model fitting.
• Empirical studies on MNIST and LSUN show divergences produce varying sample quality and likelihood behavior, with KL-related divergences sometimes yielding higher holdout likelihood in MNIST experiments.
• The framework provides guidance on output activations and T* corresponding to each f-divergence.