Temporal knowledge: Recognition and learning of time-based patterns

Abstract : A self-organizing, distributed, massively parallel network anatomy for the recognition of input stimuli and the learning of temporal patterns is proposed. The network adapts itself to recognize individual incoming events in the first, or static, subsystem. These recognized events, received by the system over time, are simultaneously categorized as specific sequences by the temporal subsystem. Separate attentional mechanisms allow for the recognition of events with a low signal-to-noise ratio while simultaneously allowing attention in the temporal subsystem to be focused only on sequences that meet some minimum length criterion. The static subsystem is based on the adaptive resonance paradigm of S. Grossberg. The temporal subsystem, a gaussian classifier, processes the static information produced by the first subsystem. These gaussian classifications represent the statistics of the temporal data and use a scheme of moving mean and moving covariance to update the classes. Via supervised learning these self-developed classes are then combined into an overall probability estimate using a bayesian probability scheme.