MixFeat: Mix Feature in Latent Space Learns Discriminative Space

Deep learning methods perform well in various tasks. However, the over-fitting problem, which causes the performance to decrease for unknown data, remains. We hence propose a method named MixFeat that directly creates latent spaces in a network that can distinguish classes. MixFeat mixes two feature maps in each latent space in the network and uses unmixed labels for learning. We discuss the difference between a method that mixes only features (MixFeat) and a method that mixes both features and labels (mixup and its family). Mixing features repeatedly is effective in expanding feature diversity, but mixing labels repeatedly makes learning difficult. MixFeat makes it possible to obtain the advantages of repeated mixing by mixing only features. We report improved results obtained using existing network models with MixFeat on CIFAR-10/100 datasets. In addition, we show that MixFeat effectively reduces the over-fitting problem even when the training dataset is small or contains errors. MixFeat is easy to implement and can be added to various network models without additional computational cost in the inference phase.