CLASSIFICATION OF TEMPORAL TRAJECTORIES BY CONTINUOUS-TIME RECURRENT NETS

We introduce an algorithm that is able to classify temporal patterns. It is based on an extension of Pearlmutter's algorithm for the generat ion of temporal trajectories, and relies on the use of variational met hods. The classification is based on the time trajectory generating th e pattern instead of the static pattern itself. The algorithm is appli ed to a classification problem, that is, the recognition of five tempo ral trajectories. For simple problems, like the one considered in this paper, the algorithm is operational Indeed, a 10-unit network is able to correctly classify the patterns, after a training time of about 1 000 epochs. Simulation results show that the net can cope with noisy o r distorted patterns (in space). The effect of pure time distortions i s more problematic but, in theory, this effect can be abolished by con sidering the curve length as a parameter, instead of time. We intend t o use this method for the purpose of shape recognition; this will allo w us to make comparisons with more classical techniques.