A learning algorithm for the classification of dynamic events using a neuron-like dynamic tree

Abstract

The use of neuron-like networks (NN) for pattern recognition has a well-established history and numerous current applications. Most such applications are to static patterns while the theory developed for temporally changing visual patterns usually assumes rigid objects with well-defined boundaries. In applications such as analysis of cardiac movement, however, the object is flexible and the images are often imperfect. The authors current model for NN activity captures the dynamic nature of the signal processing of the neural dendritic tree, allowing both faster learning of dynamic patterns and a very reduced number of receptors required for distinguishing diverse types of motion or changes. The design of the NN model is presented and a training algorithm which exhibits in practice extremely fast convergence (as few as I5 iterations) to near optimal recognition behavior is introduced.