TITLE: A PATTERN RECOGNITION APPROACH TO UNDERSTANDING THE MULTI-LAYER PERCEPTRON

This memorandum is concerned with the operation of a class of multi-layer associative networks commonly known as the multi-layer perceptron (MLP), Rumelhart network or back-propagation network. We describe the operation of the MLP as a pattern recognition device in terms of a feature-space representation. This allows an understanding of how structure in the training data is represented internally in the machine. Index Terms- Perceptron, neural networks, associative networks, pattern recognition, feature-space, error back-propagation. Copyright