Applications of binary neural networks learning to pattern classification

This paper considers the use of binary neural networks for pattern classification. An expand-and-truncate learning (ETL) algorithm is used to determine the required number of neurons as well as the connecting weights in a three-layered feedforward network for classifying input patterns. The ETL algorithm is guaranteed to find a network for any binary-to-binary mappings. The ETL algorithm´s performance in pattern classification is tested using a breast cancer database that have been used for benchmarking performance other machine learning methods. The ETL algorithm decomposes an arbitrarily linearly nonseparable function into multiple linearly separable functions, each of which is realized by a neuron in the hidden layer. ETL first finds the required hyperplanes for the given patterns, based on a geometrical analysis of the given patterns. The weights and thresholds are determined based on these identified hyperplanes. Depending on the given training patterns, the required number of neurons in the hidden layer will be determined by ETL