A neural network approach to large dimensional spectral pattern processing

The authors present a multiple neural network system that extracts and interprets spatiotemporal features from two-dimensional spectral images. The system uses interconnected multiple networks where the first network extracts spatial features and successive networks label and classify the features. The labeling network uses a priori knowledge on its connection weights, thereby eliminating the need for extensive learning. The model was applied to speech spectral images to extract morphological properties of speech sound corresponding to certain phonetic cues. This approach enabled extraction of spatio-temporal features from large images using neural networks and also provided a mechanism to use a priori knowledge in the connection weights of the network