Dynamic Mathematics for Automated Machine Learning Techniques

Abstract

Machine Learning and Neural Networks have been gaining popularity and are widely considered as the driving force of the Fourth Industrial Revolution. However, modern machine learning techniques such as backpropagation training was firmly established in 1986 while computer vision was revolutionised in 2012 with the introduction of AlexNet. Given all these accomplishments, why are neural networks still not an integral part of our society? ``Because they are difficult to implement in practice.'' I'd like to use machine learning, but I can't invest much time. The concept of Automated Machine Learning (AutoML) was first proposed by Professor Frank Hutter of the University of Freiburg. Machine learning is not simple; it requires a practitioner to have thorough understanding on the attributes of their data and the components which their model entails. AutoML is the effort to automate all tedious aspects of machine learning to form a clean data analysis pipeline. This thesis is our effort to develop and to understand ways to automate machine learning. Specifically, we focused on Recurrent Neural Networks (RNNs), Meta-Learning, and Continual Learning. We studied continual learning to enable a network to sequentially acquire skills in a dynamic environment; we studied meta-learning to understand how a network can be configured efficiently; and we studied RNNs to understand the consequences of consecutive actions. Our RNN-study focused on mathematical interpretability. We described a large variety of RNNs as one mathematical class to understand their core network mechanism. This enabled us to extend meta-learning beyond network configuration for network pruning and continual learning. This also provided insights for us to understand how a single network should be consecutively configured and led us to the creation of a simple generic patch that is compatible to several existing continual learning archetypes. This patch enhanced the robustness of continual learning techniques and allowed them to generalise data better. By and large, this thesis presented a series of extensions to enable AutoML to be made simple, efficient, and robust. More importantly, all of our methods are motivated with mathematical understandings through the lens of dynamical systems. Thus, we also increased the interpretability of AutoML concepts.