Invariant Information Learning for Image Recognition

Abstract

Neural network is difficult to understand the invariance of input data, which is one of the causes of weak neural network generalization. So the researchers usually carry out data augmentation method on the training set, which makes the neural network remember different deformation patterns. We propose an invariant information learning framework:original CNN+Spatial information Function Zone(SFZ). This framework uses correlation matrix method instead of data augmentation method to make the neural network have the ability to learn the invariance of input data. Finally, our experiment shows that CNN+SFZ can effectively help improve generalization ability without data augmentation. In the absence of data augmentation for the training set, the network with SFZ reduced the error rate by 9.01% over the original network.