The “capture effect”: a new self-organizing network for “adaptive resolution” clustering in changing environments

In this paper two open questions in pattern recognition are addressed: learning data clusters appearing naturally at various scales (or resolutions); and online learning (or learning in changing environments). These problems are faced using self-organizing neural networks. In particular, a new mechanism is presented, called “capture effect”, concerning an adaptive recruitment of neurons and local modulation of the neural receptive fields. The network is able, as shown in the experiments, to discriminate and code data clusters at a scale adapted to local data density, and to perform it online, accepting new input information without damaging previously encoded patterns. However, the network is also able to “forget”, by releasing previously recruited neurons which are no longer “refreshed” by input patterns