Self-organization of feature detectors in time sequences (SOFT)-a neural network approach to multidimensional signal analysis

We present a neural network algorithm for self-organization of feature detectors in time sequences (SOFT) based on the mathematical concept of transient attractors. It evaluates local phase space volume contraction as an indicator for good short-term predictability. SOFT supports category formation and event detection in multidimensional time sequences by linking together neural function approximation and principal component analysis. Possible extensions of the algorithm including iteration and vector quantization procedures for further data analysis are discussed.