Predropout & Inhibition: a brain-inspired method for convolutional neural network

Convolutional Neural Network (CNN) is a very popular and powerful machine learning technique for image classification. The performance of image classification depends on CNN's untangling ability of the entangled high-dimensional input data, which can be measured by the linear separability of the output vectors in the last fully-connected (fc) layer of the CNN. Inspired by the neural population coding of inferotemporal (IT) cortex in brain, we proposed a brain-inspired method named predropout & inhibition for better restricting the coding patterns and making them with lower energy and higher sparseness. The proposed predropout & inhibition method can 1) largely reduce the number of connections from the last fc layer to softmax layer, 2) restrict the energy of coding patterns at a low level significantly, and 3) improve classification performance on ImageNet-20 significantly.