Stochastic State Switching in Attractor Neural Networks

Hendriks, Ruben (2025) Stochastic State Switching in Attractor Neural Networks. Master's Thesis / Essay, Physics.

Preview	Text mPHYS2025HendriksRT-2.pdf Download (11MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (184kB)

Abstract

The CogniGron institute at the University of Groningen aims to mimic the computational abilities and energy efficiency of the brain in hardware. Converging evidence from neuroscience suggests that during computation and decision making, metastable stochastic switching between attractor states occurs in the brain. In this thesis, a minimal generalization of a stochastic two-pattern Hopfield network and a stochastic sparse block attractor network are shown to support controllable metastable stochastic attractor state switching. These networks are then used as building blocks to construct a network that can emulate an arbitrary embedded two-state Markov chain, and as a consequence, any N-state Markov chain, with its attractor state dynamics. The Hopfield implementation results in a malfunctioning embedded Markov chain due to the dense activity of Hopfield states. The sparse block network supports a successful embedding of a Markov chain, effectively demonstrating robust, controllable, multi-timescale state-dependent computing in sparse attractor networks. The generalized Hopfield network is rigorously analyzed with statistical mechanics theory. A transformation between Hopfield- and multi-group Curie-Weiss networks has been developed, which is used to analytically derive the phase diagram of the generalized Hopfield network and its key metastable properties.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Szabo, R. and Chicca, E. and Cotteret, M.A.
Degree programme:	Physics
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	18 Jul 2025 11:41
Last Modified:	18 Jul 2025 11:41
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/36403

Actions (login required)

View Item