Abstract:
Deep Learning has risen recently as the de-facto machine learning approach for
complex and rich-data domains. However, this has been established so far mostly
in the setting of supervised learning. Supervised learning requires a tremendous
amount of labeled data that is expensive, time-consuming, and not scalable. To
overcome those limitations, researchers have invested recently in unsupervised
deep learning approaches which make use of the huge amount and easily available
unlabeled data. In this dissertation, our work in the domain of unsupervised
deep learning was two-fold. We rst apply a novel model interpretation technique
to study the learned concepts of di erent unsupervised deep learning approaches
and how they compare to each other. Second, we considered one of the proposed
approaches, Deep Cluster, proposing an explanation on why it works, supporting
it with experiments, and following it by a series of ablation experiments to
understand some parameters-performance dynamics.