Work

Improving annealed importance sampling's evaluation of the partition function on restricted Boltzmann machines

Year: 2024

Type: conference-paper

Abstract: The calculation of the partition function Z is fundamental in probabilistic models such as Restricted Boltzmann Machines (RBMs). Annealed Importance Sampling (AIS) is an algorithm that gives a stochastic estimation of the value of Z. In this work we show that, with a proper initialization, one can significantly reduce the number of sampling steps needed to converge. In order to find the best strategy, several initial probability distributions used to start the AIS algorithm are tested. We found two successful strategies that perform well in the problems studied. We conclude that those strategies are a good starting point to estimate the partition function through AIS with a relatively low computational cost.

Source: UPCommons institutional repository (Universitat Politècnica de Catalunya)

Authors: Prat Pou, Arnau, Romero Merino, Enrique, Martí Rabassa, Jordi, Mazzanti Castrillejo, Fernando Pablo

Language: English

Topic: Markov Chains and Monte Carlo Methods

Subfield: Statistics and Probability

Field: Mathematics

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7139804034

FWCI:

Cites:

Cited by:

Related to: 10