Feature maps based weight vectors for spatiotemporal pattern recognition with neural nets

A neural network algorithm is used to generate the spatial classes for spatiotemporal pattern recognition (SPR). This algorithm is known as Kohonen feature maps. Training vectors are presented to the network one at a time. The connection strength between the input and output nodes is adaptively updated. The adaptation process is associated with a decay of the adaptation rate as well as a shrinkage of the neighborhood for updating. The final values of connection strength represent the centroid of clusters of training patterns. The algorithm was tested with hypothetical data as well as hydrophone data. Functional forms and constants for the decay and the shrinkage were empirically determined. The algorithm performs better with broadband data than with narrowband data. Also, the algorithm works better with a smaller number of pattern classes