Semi-Supervised Learning via New Deep Network Inversion
This provides a simple and efficient semi-supervised learning method that works across various systems without architectural changes, though it appears incremental in approach.
The authors tackled the problem of semi-supervised learning with very few labeled examples by developing a new framework based on deep network inversion, achieving 99.14% test accuracy on MNIST with only 5 labeled examples per class.
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 outperforms current state-of-the-art methods on MNIST reaching $99.14\%$ of test set accuracy while using $5$ labeled examples per class. 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.