Markov Chain Monte Carlo Sampling Algorithms in Neural Circuitry

Abstract

There is growing evidence that the nervous system is tuned and adapted to the sta- tistical properties of the sensory information from the surrounding environment and that it is possible that the brain performs some form of probabilistic inference. From this perspective, Markov Chain Monte Carlo (MCMC) simulations could serve as a possible and intriguing explanation for how neural activity could compute probabilis- tic inference. This paper derives both continuous and spiking neural network models that implement MCMC sampling algorithms with the hope of finding a fast, efficient, and biologically plausible algorithm for inference. Both sets of models show promise in performing probabilstic inference, but the spiking models show decreasing per- formance in high-dimensional data. Extending these algorithms to multi-chain sim- ulations or strict excitatory/inhibitory enforcement on the network could provide a solution to the shortcomings of the spiking network models.