Nonlocal optimization of binary neural networks

TL;DR

This paper formulates the training of Binary Neural Networks as a discrete inference problem and introduces stochastic message passing algorithms that outperform traditional gradient methods in finding optimal parameters.

Contribution

It proposes stochastic versions of Belief Propagation and Survey Propagation for BNN training, improving over existing gradient-based methods.

Findings

Stochastic BP and SP find better BNN configurations.

The methods outperform traditional gradient approaches.

The approach is effective in under-parameterized BNN settings.

Abstract

We explore training Binary Neural Networks (BNNs) as a discrete variable inference problem over a factor graph. We study the behaviour of this conversion in an under-parameterized BNN setting and propose stochastic versions of Belief Propagation (BP) and Survey Propagation (SP) message passing algorithms to overcome the intractability of their current formulation. Compared to traditional gradient methods for BNNs, our results indicate that both stochastic BP and SP find better configurations of the parameters in the BNN.