AUTHOR=Tsukada Hiromichi , Tsukada Minoru 

TITLE=Comparison of Pattern Discrimination Mechanisms of Hebbian and Spatiotemporal Learning Rules in Self-Organization

JOURNAL=Frontiers in Systems Neuroscience

VOLUME=Volume 15 - 2021

YEAR=2021

URL=https://www.frontiersin.org/journals/systems-neuroscience/articles/10.3389/fnsys.2021.624353

DOI=10.3389/fnsys.2021.624353

ISSN=1662-5137

ABSTRACT=The Hebbian learning rule (HEB) and spatiotemporal learning rule (STLR) differ in the mechanism of self-organization. In particular, the HEB and STLR achieve self-organization of information from the external world by firing and not firing output neurons, respectively. In the HEB, the neuron group having the most similar weight to the input vector is selected, and its synaptic weight is strengthened to ensure that input vectors in regions close to each other produce the same output. Consequntly, after learning the training vector, the synapse weight vector exhibits a unimodal distribution near the training vector. For similar input vector sequences, although the distribution changes to a steep distribution whenever the input vector is trained, unfired neurons do not fire for subsequent inputs. Therefore, the output is a constant output vector. This function shows the feature of pattern completion of the HEB rule. In contrast, in the STLR, when the training vector is in the same vector series, then it does not change. Therefore, the synapse weight vector has a unimodal distribution centered on the training vector. Although the variation for each training changes to a steep distribution similar to the HEB, the output is not constant because the threshold value of the neuron reads the change in the synapse weight vector distribution and expresses it in the output vector sequence. For similar input sequences, the synapse weight vector develops a new steep unimodal near a slightly different training vector, and exhibits a multimodal distribution depending on the sequence. The threshold value of the neuron reads this change and expresses it in the output vector series. This feature demonstrates the high pattern-discrimination ability of the STLR. From these results, it can be observed that the difference between the self-organization of the HEB and STLR yields a difference in the dynamic change of the distribution of synapse weight vectors constructed dynamically through training.
This difference adequately explains well the features of pattern completion in the HEB and the features of high pattern discrimination in the STLR.