Interactive reservoir computing for chunking information streams

Toshitake Asabuki, Naoki Hiratani, Tomoki Fukai

Abstract: Author summary Varieties of information processing require chunking, but chunking arbitrary complex sequences as flexibly as the brain does remains a challenge. In this study, we solved this important but difficult problem of chunking by "reservoir computing" inferred from brain computation. In the method, chunking occurs automatically when a pair of such modules supervising one another agree on the recurring response patterns to remember. Our model revealed a unique dynamical mechanism for embedding the temporal community structure of inputs into low-dimensional neural trajectories. From a biological viewpoint, our model based on pairwise computing suggests the computational implications of brain's bi-hemispheric information streams. Owing to recent technological advances, the implementation of this model by electronic devices should be straightforward.