An unsupervised learning algorithm for the sequential classification of patterns using static nets

A flexible training algorithm which would enable static nets to perform sequential classification of input patterns is proposed. This algorithm allows the net to iteratively formulate a suitable target output sequence for each input sequence without any supervision. The algorithm is tested on a neural net specially designed to classify time-sensitive input patterns. Test results (for speaker independent word recognition) using a spatiotemporal network suggest that the method works well with feedforward nets trained using backpropagation