HABITUATION BASED NEURAL NETWORKS FOR SPATIOTEMPORAL CLASSIFICATION

A new class of neural networks is proposed for the dynamic classificat ion of spatio-temporal signals, These networks are designed to classif y signals of different durations, taking into account correlations amo ng different signal segments, Such networks are applicable to SONAR an d speech signal classification problems, among others, Network paramet ers are adapted based on the biologically observed habituation mechani sm. This allows the storage of contextual information, without a subst antial increase in network complexity. We introduce the concept of a c omplete memory. We then prove mathematically that a network with a com plete memory temporal encoding stage followed by a sufficiently powerf ul feedforward network is capable of approximating arbitrarily well an y continuous, causal, time-invariant discrete-time system with a unifo rmly bounded input domain, The memory mechanisms of the habituation ba sed networks are complete memories, and involve nonlinear transformati ons of the input signal, In networks such as the time delay neural net work (TDNN) [35] and focused gamma networks [8], nonlinearities are pr esent in the feedforward stage only. This distinction is made importan t by recent theoretical results concerning the limitations of structur es with linear temporal encoding stages, Results are reported on class ification of high dimensional feature vectors obtained from Banzhaf so nograms.