CortexNet: Convolutional Neural Network with Visual Cortex in human brain

摘要

The current Convolutional Neural Networks (CNN) [1-4] aim to extend the depth of the network for processing huge data sets effectively. However, they only partially imitate brain function, such as delegating weights by convolution operation. We propose a new CNN architecture by introducing a Cortex block, which mimics the human visual connectome [5]. Extracting features from binocular information is repeated until the end of top-level cell, Inferior Temporal. With the learning process, a human brain activates some neurons by following the results of predicting the future input data. Based on reflecting these brain functions, we design a Cortex block with general methods in CNN, such as convolutional layer and subsampling layer. Cortex block reduces the number of learnable parameters as well as observes two main functions of human visual system. Cortex blocks stacked up to make CortexNet, showed performance parity with ResNet and SENet on CIFAR-10 and Tiny-ImageNet.