Core Concepts
Efficient learning requires memorizing class-specific information and accumulating knowledge about general mechanisms.
Abstract
Introduction to the challenges of continual learning.
Introduction of Infinite dSprites as a benchmark tool.
Evaluation of major types of continual learning methods on the benchmark.
Proposal of Disentangled Continual Learning approach.
Summary of contributions and findings.
Comparison of different methods and experiments.
Limitations, societal impact, and reproducibility considerations.
Stats
Continual learning methods break down on Infinite dSprites benchmark over time.
Disentangled Continual Learning separates memorization from generalization.
Contrastive learning baseline decays over time but not as rapidly as naive fine-tuning.
Quotes
"Efficient learning requires memorizing class-specific information and accumulating knowledge about general mechanisms."