How to Leverage Predictive Uncertainty Estimates for Reducing Catastrophic Forgetting in Online Continual Learning
Abstract: Many real-world applications require machine-learning models to be able to deal with non-stationary data distributions and thus learn autonomously over an extended period of time, often in an online setting. One of the main challenges in this scenario is the so-called catastrophic forgetting (CF) for which the learning model tends to focus on the most recent tasks while experiencing predictive degradation on older ones. In the online setting, the most effective solutions employ a fixed-size memory buffer to store old samples used for replay when training on new tasks. Many approaches have been presented to tackle this problem and conflicting strategies are proposed to populate the memory. Are the easiest-to-forget or the easiest-to-remember samples more effective in combating CF? Furthermore, it is not clear how predictive uncertainty information for memory management can be leveraged in the most effective manner. Starting from the intuition that predictive uncertainty provides an idea of the samples' location in the decision space, this work presents an in-depth analysis of different uncertainty estimates and strategies for populating the memory. The investigation provides a better understanding of the characteristics data points should have for alleviating CF. Then, we propose an alternative method for estimating predictive uncertainty via the generalised variance induced by the negative log-likelihood. Finally, we demonstrate that the use of predictive uncertainty measures helps in reducing CF in different settings.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission:
Changes from the original submission are highlighted in blue.
- Introduction: in response to reviewers pj6v and 3Hfb, we reorganized the introduction to make sure to explain the reasoning and hypothesis behind our investigation and proposed methodology, and to make sure each term is followed by its description. Furthermore, we improved the overall readability of the section and included a clarification about the definition of representative samples.
- Additional experiments and analysis: as requested by reviewer PuAH, we included 1) additional experiments with A-GEM, 2) an analysis of the uncertainty spectrum of the samples stored in the memory buffer, and 3) a time complexity analysis between ER and the proposed method.
- Improved notation and descriptions: as pointed out by reviewers 3Hfb and pj6v, we improved the notation throughout the whole manuscript and the description of terms and datasets.
- Improved overall presentation: we rephrased and restructured different sections of the manuscript to improve the presentation. To facilitate a more straightforward comparison of strategies (least uncertain/most uncertain) in Tables 1 and 2, we now highlight the columns corresponding to each strategy (top, step, and bottom) with different colors.
Supplementary Material: zip
Assigned Action Editor: Emmanuel Bengio
Submission Number: 3663
Loading