Auto-learning by dynamical recognition networks

Demonstrates that a new type of hybrid networks can be useful for detecting "unknown" patterns and auto-learning of them. An essential point of the mechanism is a dynamical recognition based on chaotic EEG activities of mammalian brains. The chaotic activities are generated by designing recurrent connection weights in the hybrid networks with feedforward and recurrent connections. Harnessing the chaotic dynamics of recurrent networks, the networks can recognize "known" patterns and their neighbors as conventional recognition methods are possible. We present some simulation results illustrating the networks ability for deciding whether input patterns are "known" or "unknown" by observing temporal stability of output patterns. Finally, it is shown that recognition of "unknown" patterns makes it possible for the networks to learn new patterns automatically.