Unsupervised Progressive Learning for Agents in Online and Dynamic Environments

Deep learning systems have demonstrated remarkable capabilities in learning useful representations from complex data. However, these capabilities typically rely on assumptions that the training data is fully available in advance, that some or all of it is labeled, and that its distribution is fixed and matches future test data. These assumptions stand in stark contrast to the conditions under which biological learners—such as humans or animals—learn from their environments. To bridge this gap, this dissertation introduces Online Unsupervised Continual Learning (O-UCL), a new learning setting where data arrives as an unlabeled, non-stationary stream. The thesis first formalizes the O-UCL problem and presents an initial solution inspired by neuroscience and classical machine learning. It then introduces a second approach that leverages contrastive learning to handle more challenging natural image streams. Finally, it explores additional challenges unique to O-UCL and outlines the properties necessary for robust, adaptive solutions to those challenges. This work lays a foundation for the development of learning algorithms that can operate in complex, ever-changing real-world environments.

Date

2025-04-25

Resource Type

Text

Resource Subtype

Dissertation

Full item page