Defense Date

2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Bartosz Krawczyk

Abstract

Continual learning aims at developing models that are capable of working on constantly evolving problems over a long-time horizon. In such environments, we can distinguish three essential aspects of training and maintaining machine learning models - incorporating new knowledge, retaining it and reacting to changes. Each of them poses its own challenges, constituting a compound problem with multiple goals.

Remembering previously incorporated concepts is the main property of a model that is required when dealing with stationary distributions. In non-stationary environments, models should be capable of selectively forgetting outdated decision boundaries and adapting to new concepts. Finally, a significant difficulty can be found in combining these two abilities within a single learning algorithm, since, in such scenarios, we have to balance remembering and forgetting instead of focusing only on one aspect.

The presented dissertation addressed these problems in an exploratory way. Its main goal was to grasp the continual learning paradigm as a whole, analyze its different branches and tackle identified issues covering various aspects of learning from sequentially incoming data. By doing so, this work not only filled several gaps in the current continual learning research but also emphasized the complexity and diversity of challenges existing in this domain. Comprehensive experiments conducted for all of the presented contributions have demonstrated their effectiveness and substantiated the validity of the stated claims.

Rights

© The Author

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission

5-5-2022

Share

COinS