Unsupervised learning: model-based clustering and learned compression

This thesis addresses two central tasks prevalent in many modern data processing, storage, and transmission pipelines: Clustering and compression. Specifically, in the first and the second part of this thesis, we study the problems of subspace clustering and random process clustering, respectively. While clustering problems are arguably among the most archetypal problems in unsupervised learning, compression methods are traditionally hand designed. In the third and fourth part of this thesis, we leverage machine learning techniques for compression, a trend that only emerged recently. In more detail, we propose a deep generative model-based framework for lossy data compression on one hand, and we study compression of neural network models for inference on resource-constrained hardware on the other hand.