Efficient Continual Learning
: Approaches and Measures

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)

Abstract

In the real world, we often encounter situations where data distributions are changing over
time, and we would like to timely update our models by the new data, with bounded growth
in system size and computational cost. Continual learning is a research topic for dealing
with such scenarios. The main challenge of continual learning is catastrophic forgetting: when
training a model by sequential tasks, the model tends to forget previously learned tasks after
learning new ones. Here different tasks mean different training and testing sets.
In Bayesian continual learning, we would prefer the posteriors of a new task being as close as
possible to the previous ones. We propose Gaussian Natural Gradients for Bayesian continual
learning, which uses natural gradients instead of conventional gradients as natural gradients
prefer the smallest change in terms of distributions rather than parameters. We also propose
Stein Gradient-based Episodic Memories that can construct compact episodic memories using the
information in posteriors in Bayesian continual learning. In addition, we propose Discriminative
Representation Loss for general continual learning, which decreases the diversity of gradients
between new and old tasks through optimizing representations instead of re-projecting the
gradients. It effectively improves performance with low computational cost compared with related
work.
Moreover, we propose β
3
-Item Response Theory model for evaluating classifiers in continual
learning. The ability inferred by β
3
-IRT is weighted by difficulty of individual samples, which can
provide more sensible evaluations than typically used average accuracy. We furthermore propose
Continual Density Ratio Estimation for evaluating generative models in continual learning
without storing any data from previous tasks. To the best of our knowledge, this is the first
measure for generative models that satisfies the restriction of continual learning.
In summary, we propose several efficient approaches and measures for continual learning in
this thesis.
Date of Award28 Sept 2021
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorPeter A Flach (Supervisor) & Tom Diethe (Supervisor)

Keywords

  • Continual Learning
  • Machine Learning

Cite this

'