Understanding SyncMap: Analyzing the components of Its Dynamical Equation

SycnMap has been recently proposed as an unsupervised approach to perform chunking. This model, which falls under the paradigm of self-organizing dynamical equations, can achieve learning merely using the principle of self-organization without any objective function. However, it is still poorly understood due to its novelty. Here, we provide a comprehensive analysis of the underlying dynamical equation that governed the learning of SyncMap. We first introduce several components of the dynamical equation: (1) Learning rate, (2) Dynamic noise, and (3) Coefficient of attraction force; As well as model-specific variables: (4) Input signal noise and (5) Dimension of weight space. With that, we examine their effect on the performance of SyncMap. Our study shows that the dynamic noise and dimension of weight space play an important role in the dynamical equation; By solely tuning them, the enhanced model can outperform the baseline methods as well as the original SyncMap in 6 out of 7 environments.