II Machine Learning Methods for Complex and Quantum Systems

Study of the criticality on RBMs using AIS

Not scheduled

20m

Aula Magna (Facultat de Física, Universitat de Barcelona)

Poster

Speaker

Arnau Prat Pou (Universitat Politècnica de Catalunya)

Description

We have used Annealed Importance Sampling (AIS) to study the critical phenomena in Ising-like models and Restricted Boltzmann Machines (RBMs). We have proven that, using a proper initialization based on a mean-field approximation of the target distribution, we can obtain more efficient and accurate sampling of different thermodynamical quantities, and successfully detect phase transitions across different models, with a low computational cost. In contrast, standard initialization schemes can lead to biased estimates and may fail to capture transition behaviour.