Deep Generative Models for Semi-Supervised Machine Learning

The reintroduction of deep neural networks has a large impact on the modeling capabilities of modern machine learning. This reignites the general public’s dream of achieving artificial intelligence, and spawns rapid progress in large scale industrial machine learning development, such as autonomous driving. However, the leaps in development are still confined to a rather limited learning domain, in which labeled data is required. Labeled data is hard and costly to acquire, due to the amount needed to efficiently learn a modern machine learning model, and that many data sources are not directly interpretable. Consequently, research in different learning paradigms that utilize vast amounts of unlabeled data is getting more and more attention. Albeit possessing intriguing theoretical properties, machine learning models that learn from unlabeled data are still an unsolved research topic. The thesis comprises methods that utilize the power of deep neural networks to learn from both labeled and unlabeled data. A background for the theoretical foundation of the proposed methods are described and empirical results showing their capabilities within generation and classification tasks are presented. Finally, a real-life application within condition monitoring for sustainable energy is demonstrated, proving that the proposed methods have the expected impact and are applicable.