The Cramér-InfoGAN and Partial Inverse Filter System for Unsupervised Image Classification

The scarcity of labelled data and the abundance of raw data have endowed unsupervised approaches to image classification with great significance. To enable learning in domains where labelled data are few, an unsupervised image classification system consisting of Cramer-InfoGAN and Partial Inverse Filter is proposed in this article. The former learns disentangled representation of data and builds a mapping from it to images, which is the main functionality of InfoGAN. Moreover, it combines Cramer GAN to improve the stability and convergence of training, also to monitor the reliability of learned representation by observing the energy distance of Cramer GAN. The latter classifies images by mapping it back to its representation. Furthermore, it is regularized by gradient penalty to suppress noise and by independence constraint to reduce entanglement of different dimensions. Experiments on MNIST and CIFAR-10 datasets show improved convergence and respectable classification accuracy of the system.