Weighted Contrastive Divergence

TL;DR

This paper introduces Weighted Contrastive Divergence (WCD), a new algorithm for training energy-based models like RBMs, which improves upon standard CD and persistent CD with minimal additional computational cost.

Contribution

The paper proposes WCD, a modified version of CD that enhances gradient approximation accuracy with small changes and demonstrates superior performance experimentally.

Findings

WCD outperforms standard CD and persistent CD in experiments.

WCD achieves better gradient estimates with minimal extra computation.

Experimental results show improved learning efficiency.

Abstract

Learning algorithms for energy based Boltzmann architectures that rely on gradient descent are in general computationally prohibitive, typically due to the exponential number of terms involved in computing the partition function. In this way one has to resort to approximation schemes for the evaluation of the gradient. This is the case of Restricted Boltzmann Machines (RBM) and its learning algorithm Contrastive Divergence (CD). It is well-known that CD has a number of shortcomings, and its approximation to the gradient has several drawbacks. Overcoming these defects has been the basis of much research and new algorithms have been devised, such as persistent CD. In this manuscript we propose a new algorithm that we call Weighted CD (WCD), built from small modifications of the negative phase in standard CD. However small these modifications may be, experimental work reported in this paper suggest that WCD provides a significant improvement over standard CD and persistent CD at a small additional computational cost.