Structured neural networks for signal classification

Abstract

In this paper, artificial neural networks are considered as an emergent alternative to the classical `model-based approach' to the design of signal-processing algorithms. After briefly examining the pros and cons of the neural-network approach, we propose the application of structured neural networks (SNNs) for the classification of signals characterized by different `information sources', such as multisensor signals or signals described by features computed in different domains. The main purpose of such neural networks is to overcome the drawbacks of classical neural classifiers due to the lack of general criteria for `architecture definition' and to the difficulty with interpreting the `network behaviour'. Our structured neural networks are based on multilayer perceptrons with hierarchical sparse architectures that take into account explicitly the `multisource' characteristics of input signals and make it possible to understand and validate the operation of the implemented classification algorithm. In particular, the interpretation of the SNN operation can be used to identify which information sources and which related components are negligible in the classification process. SNNs are compared with both commonly used fully connected multilayer perceptrons and the k-nearest neighbour statistical classifier. Experiments on two multisource data sets related to magnetic-resonance and remote-sensing images are reported and discussed.