Hierarchical pattern extraction for machine perception
A. Coward, Sudhendu Kumar, A. Hung, GRAHAM A. JULLIEN · 2002
The authors present a new architecture for machine perception of objects using a hierarchical pattern extraction technique. The resulting architecture is a neural network with ordinary logic gates as the neurons and simple heuristic pattern association techniques as the training algorithm. The architecture consists of a multilayer network of neurons and a final layer with a single neuron. The interconnections between the different layers are determined on the fly during the training process. Most of the data that is to be processed during training can be represented as binary values; likewise, all synapse values are binary. The application area is in the recognition of a single object type from a field of object types, a common problem in machine perception. The authors introduce the architecture and training algorithm, and present initial results using statistically defined objects.