MLP@P seminars : Marco Baiesi (Padova)

Quand

11/12/2026    
10:30 - 11:30

Où

LISN, bat 660 salle 2014 (2° étage)
660 Av. des Sciences, 91190 Gif-sur-Yvette, France, Gif-sur-Yvette, 91190

Type d’évènement

Carte non disponible

The emergence of latent cycles in nonequilibrium Markov machines

Marco Baiesi

Università di Padova

Seminar of the Series MLP@P (Machine Learning Physics @ Plateau), joint with LISN and IPhT.
Where: LISN, bat 660 salle 2014 (2° étage)
More info at: https://vale1925.wixsite.com/mlpatp

Latent-variable models such as restricted Boltzmann machines learn a data distribution through an alternation between visible and hidden variables, with a sampling dynamics that satisfies detailed balance. We introduce a nonequilibrium Markov machine (NEMM), a simple two-layer generative model in which visible and hidden variables interact through two independently parametrized transition matrices, defining a Markov chain whose steady state is out of equilibrium. Training relies on an approximate log-likelihood gradient that depends on the last two steps of the chain. The application of a NEMM to MNIST digits shows that likelihood maximization never selects the reversible limit and that the hidden dynamics self-organizes into cycles through the latent states. Interestingly, models learning more irreversible dynamics reproduce the data class distribution more faithfully and generate sequences that alternate between digits. In comparison, a standard restricted Boltzmann machine tends to remain trapped in one mode, generating the same digit for many consecutive samples. To conclude, we discuss how controlling and enhancing irreversibility may improve generative models.

Retour en haut