Semi-Supervised Learning via New Deep Network Inversion

We exploit a recently derived inversion scheme for arbitrary deep neural networks to develop a new semi-supervised learning framework that applies to a wide range of systems and problems. The approach reaches current state-of-the-art methods on MNIST and provides reasonable performances on SVHN and CIFAR10. Through the introduced method, residual networks are for the first time applied to semi-supervised tasks. Experiments with one-dimensional signals highlight the generality of the method. Importantly, our approach is simple, efficient, and requires no change in the deep network architecture.