Asynchronous network of cellular automaton-based neurons for efficient implementation of Boltzmann machines

Abstract

Artificial neural networks with stochastic state transitions and calculations, such as Boltzmann machines, have excelled over other machine learning approaches in various benchmark tasks. The networks often achieve better results than deterministic neural networks of similar sizes, but they require implementation of nonlinear continuous functions for probabilistic density functions, thus resulting in an increase in computational effort. The architecture size of cutting-edge artificial neural networks are ever-growing; therefore, they require dedicated hardware. Conversely, asynchronous cellular automaton-based neuron models have been investigated to model the highly nonlinear dynamics of biological neurons. They are special types of cellular automata and are implemented as small asynchronous sequential logic circuits. In this study, we propose a new type of asynchronous network of cellular automaton-based neuron for the efficient implementation of Boltzmann machines. Experimental comparisons demonstrate that the proposed approach achieves comparable or better performances in such benchmark tasks as image classification and generation while it requiring much less computational resources than traditional implementation approaches.