Temporal pattern recognition

A network is described that takes as its input individual incoming events. Sequences of these events (letters, phonemes, or, more abstractly, object sightings in a vision system), received by the system over time are categorized as specific sequences by the temporal system. The temporal system produces Gaussian classifications that represent the statistics of the temporal data, self-developed classes. The system recognizes sequences in a noisy environment, giving as output a Gaussian distance from the stored sequence, thus providing an analog measure of closeness of fit to currently known patterns. The system can recognize sequences with missing or extraneous elements, as well as out-of-order sequences. In addition, a desirable prediction property-that the system realizes it may be in a particular sequence long before the entire sequence has been introduced- is a consequence of the multidimensional Gaussian distance calculation.<>