Towards Flexible Active And Online Learning With Neural Networks

Abstract

Deep learning has elicited breakthrough successes on a wide array of machine learning tasks. Outside of the fully-supervised regime, however, many deep learning algorithms are brittle and unable to reliably perform across model architectures, dataset types, and optimization parameters. As a consequence, these algorithms are not easily usable by non-machine-learning experts, limiting their ability to meaningfully impact science and society.

This thesis addresses some nuanced pathologies around the use of deep learning for active and passive online learning. We propose a practical active learning approach for neural networks that is robust to environmental variables: Batch Active learning by Diverse Gradient Embeddings (BADGE). We also discuss the deleterious generalization effects of warm-starting the optimization of neural networks in sequential environments, and why this is a major problem for deep learning. We introduce a simple method that remedies this problem, and discuss some important ramifications of its application.