Self-Organization of Linear Pattern Classifiers. From "One-Shot" to Statistical Learning

We studied competitive learning dynamics in different time scale regimes of learning. By first assuming that learning is fast, we develop a self-organizing linear pattern classifier which can categorize patterns based on the presentation of single examples. The system can recognize patterns as "unknown" and autonomously instantiate new categories. Competition is governed by neural field dynamics, the properties of which enable us to consider the plastic behavior also in the limit of statistical learning. We can show that in this regime the dynamic equations can be approximated by the well known Kohonen algorithm, including a shrinking neighborhood function and decreasing learning rate. Our results show that biological systems might solve different learning tasks with identical rules by varying relative time scales among neural activity, learning and environmental dynamics. 1 Introduction In theories on biological as well as artificial neural networks the notion of "learning" is usuall..