A fast modified constructive-covering algorithm for binary multi-layer neural networks

For binary neural networks (BNNs), constructive covering frameworks have been investigated recently. While these frameworks are fast, they have limitations of generalization and accurate classification for learning from limited number of samples. In this paper, we propose modified constructive-covering algorithm (MCCA), which consists of two processes: generalization process and modification process. Errors introduced in the generalization process are revised in the modification process by adding modification neurons. In our approach, we visualize hidden neurons in terms of hypershperes. The learning process is the geometrical expansion process of these hypershperes. Through our experimental work in Section 5, we conclude that, MCCA is not sensitive to the order in which the input sequence is given. In addition, MCCA results in simple neural network structures by less training time.