Convolutional Neural Networks Considering Robustness Improvement and Its Application to Face Recognition

This paper proposes a novel activation function to promote robustness to the outliers of the training samples. Data samples in the decision boundaries are weighted more by adding the derivatives of the sigmoid function outputs to avoid drastic update of the network weights. Therefore, the network becomes more robust to outliers and noisy patterns. We also present appropriate backpropagation learning algorithm for the convolutional neural networks. We evaluate the performance improvement by the proposed method on a face recognition task, and proved that it outperformed the state of art face recognition methods.