Reinforcing feature distributions of hidden units of Boltzmann machine using correlations

Abstract

This paper introduces and analyses the method of applying Neuroscience methods to Boltzmann Machine, involving a combination of cognitive psychology, information theory, and dynamical systems. We utilized the emergent property of the probability of hidden layers to find the pattern of how units are behaving when stimulated by the visual layer and research into enhancing the predictive encoding capability of the encoding layer. We measure the connections and links between the units of the encoding layer by approximating it with the probability distribution of two units' activation behaviours. For example, the portion of the Auditory cortex responsible for processing auditory information, such as music, differs from the sections responsible for processing visual information, although they can still be linked and active concurrently. Besides, Neurons can modify their connections by learning new information and reinforcing the connections that have been utilized more frequently, and forgetting the connections if the probability distributions of two units diverge much. The Boltzmann machine is the probabilistic inference machine for ground truth using the free energy principle. The latter has stepped further from the concept to interpret cortical responses as a fundamental of intelligent agency. With simple and random interactions of each neuron, this 'intelligent agency' could achieve sophisticated functions in a specific area of a brain. Randomness is also a vital aspect of learning since it may achieve balance and embrace regularities according to Ramsey's Theory.