Methods are developed for extending the unsupervised projection pursuit learning algorithm of Bienenstock, Cooper and Munro (BCM) (1982) to time-dependent classification problems. Recurrent and differential models of BCM which look for temporal structure in the evolution of high-dimensional inputs are described. Ordinary BCM obtains a 10db improvement in a noise tolerance study when compared with backward propagation (BP) for a database of simulated inverse synthetic aperature radar (ISAR) presentations. The recurrent and differential BCM models address the problem of classification from sequences of multiple presentations.
展开▼
