Random Features for Discrete Amortized Samplers

Tiago Silva, Amauri H Souza, Salem Lahlou

Keywords: Bayesian Inference, GFlowNets, GNNs, RNI

TL;DR: We characterize the use of random features in Markov amortized samplers as a mixture model, showing they improve the sampler's expressivity and goodness-of-fit

Abstract: Sequential discrete amortized samplers, such as GFlowNets, learn a stochastic decision process (SDP) that generates each state in proportion to a given unnormalized probability mass function. While successfully applied to wide-ranging tasks, they suffer from slow convergence, underdetermination, and limited expressiveness. Notably, we show these issues can be mitigated by adding random inputs to the SDP's policy network. This approach naturally provides a mixture semantics to the learning algorithm, which can be directly adapted to any existing SDP-based discrete samplers. In view of this, we refer to our method as Mixture Model Augmentation (MMA). Our empirical analysis on standard benchmark problems indicates MMA % improves distributional accuracy and accelerates learning convergence and improves distributional accuracy.

Submission Category: Extended Abstract

Overaged Verification: Yes

Latin American Hispanic Heritage: Yes

Icml Proceedings Status: No

Submission Number: 27