Understanding SyncMap’s Dynamics and Its Self-organization Properties: A Space-time Analysis

Abstract

Human are shown able to rapidly recognize patterns in sequences by detecting and chunking together the patterns found, without supervised signals. Recently, inspired by how neuron groups act in quickly switching behaviors, SyncMap was proposed to solve chunking problems based solely on self-organization. The idea is to create dynamical equations that maintain an equilibrium state by dynamically updating with positive and negative feedback loops. When the underlying structure changes, the system can quickly adapt to the new structure. Although SyncMap can solve chunking problems effectively, the properties of its dynamics during training, is still underexplored. Here, we give a detailed investigation of SyncMap’s dynamics by using several experiments to demonstrate the behaviors of SyncMap from the perspectives of space and time, in which a problem that causes imprecise results in the original work was identified. We then propose a solution call SyncMap with moving average (i.e., SyncMap-MA), which surpasses the original work and the baselines in all experiments, suggesting that the modification here is effective and can be integrated in the future version of the algorithm.