Deep Clustering Based on Implicit Likelihood Maximization

Georgios Vardakas, M.Sc. in Data and Computer Systems Engineering, Department of Computer Science and Engineering, School of Engineering, University of Ioannina, Greece, October 2021. Deep Clustering Based on Implicit Likelihood Maximization. Advisor: Aristidis Likas, Professor. Deep Learning is a type of machine learning and artificial intelligence that imitates how the human brain can learn. It is one of the essential elements of data science, which includes statistics and predictive modeling. Although DL started mainly for supervised tasks, lately, it has found success in several unsupervised learning fields, like clustering, dimensionality reduction, etc. Clustering belongs to unsupervised machine learning and is defined as a process of assigning objects to groups so that the data share common characteristics. Therefore, the main goal of clustering is for objects belonging to the same group to be similar (or related) to each other and differ (or not be related) to objects in different groups. This way, clustering explores the data and aims to find (hidden) structures in them. At the same time, clustering is one of the most challenging problems in the field of machine learning. In this master’s thesis, we study Deep Clustering methods. Deep clustering is a new promising area of clustering algorithms that emerged in recent years. The main goal of Deep Clustering is to create clustering algorithms merged with Deep Learning methods to exploit their representational power. Τherefore, in this thesis, we will clearly describe the new machine learning area of Deep Clustering and why it is considered promising. Afterward, we will present two Deep Clustering algorithms that were studied. The first Deep Clustering algorithm that we will discuss is the ClusterGan, which makes use of a modified Generative Adversarial Networks’ architecture in order to cluster the data in latent space Z. The second Deep Clustering method that we will present is our contribution, and it is based on a generative Deep Neural Network model that is trained by Implicit Likelihood Maximization (IMLE). IMLE provides an effective way of maximizing the likelihood of the model indirectly. The Deep Clustering methodology that is based on IMLE also clusters the data in the latent space. Finally, we will analyze the experiments that took place and present the experimental results.