Osaulenko V. Models of biological neural networks for spatio-temporal association memory

Osaulenko V.M. Models of biological neural networks for spatio-temporal association memory. - The manuscript. Thesis for a candidate degree in specialty 05.13.23 - Systems and applications of artificial intelligence. - National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute", Kyiv, 2019. This work uses the results of biological neural network research to refine previous and build new models of spatial-temporal associative memory. A new model of spatial associative memory based on the sigma pi neuron is proposed, which takes into account dendritic calculations and has a high memory capacity, which exceeds the classic Willows model by some indicators. It is shown that the use of a sigma-pi neuron and sparse activation improves the order of the hierarchical temporal memory (HTM) memory capacity of the transitions with the same number of connections. A new model of sequence representation in a binary sparse distributed representation is proposed, which incorporates the neurons excitability and reproduces qualitatively biological effects such as preservation of similarity, sensitivity to order, sequence completion, and temporal similarity. The presented models of spatial-temporal associative memory show worse image recognition results for robotics tasks compared to the deep neural network approach, but they are biologically plausible, have attractive computational properties and are therefore worth for further research. Keywords: associative memory, dendritic computation, sequence prediction, sequence recognition, sigma-pi neuron, sparse distributed representation