Anefficient learning algorithm for binary feedforward neural networks

This paper proposes a novel learning algorithm of Binary Feedforward Neural Networks (BFNNs) by combining the self-adaptations of both architecture and weight. Similar to the learning algorithm of Extreme Learning Machine (ELM), our algorithm only adapts the number of hidden neurons and output weights to effectively train BFNNs with a single hidden layer for classification problems. The algorithm consists of two steps including the expanding and pruning phases. During the expanding phase, the algorithm increases the hidden neurons and also searches the weight of the output neuronusing the Perceptron Learning Rule to increase the learning accuracy. In the pruning phase, the least relevant hidden neurons measured by a proposed binary neuron´s sensitivity are pruned to reduce the complexity of the model and increase the generalization ability. Experimental results confirmed the feasibility and effectiveness of the proposed algorithm.