Semi-Supervised Learning via New Deep Network Inversion

TL;DR

This paper introduces a novel semi-supervised learning method leveraging deep network inversion, achieving high accuracy with minimal labeled data across various systems without altering network architecture.

Contribution

It presents a new inversion-based semi-supervised learning framework that is broadly applicable, simple, and efficient, outperforming existing methods on standard benchmarks.

Findings

Achieved 99.14% accuracy on MNIST with only 5 labels per class

Demonstrated method's effectiveness on one-dimensional signals

No modifications needed to existing deep network architectures

Abstract

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.