Semi-Supervised Deep Representation Learning

Abstract

Deep neural networks need a lot of data to show their full potential in modeling and solving problems. However, in many real-world applications labeling data is expensive or not feasible while abundant unlabeled data is available. Semi-supervised learning has shown to be a successful solution for such scenarios. In this thesis, we introduce semi-supervised algorithms based on neural networks to tackle a couple of problems and applications. We proposed semi-supervised algorithms for three categories of machine learning problems: verification problem, multi-view problems, and fairness. First, we propose two semi-supervised algorithms for verification problem. One of those benefits from auto-encoders and the other one benefits from adversarial training to exploit the unlabeled data and improve the performance of the verification task. Then, we present a multi-view learning algorithm capable of benefiting the cross-view correlation to exploit the structural information of the unlabeled data. In the last work, we propose using unlabeled data, which usually contain less bias than labeled data, to improve the fairness of neural network classifiers.