Reproduce “Continual Learning through Control Minimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sander de Haan, Yassine Taoudi-Benchekroun, Pau Vilimelis Aceituno, Benjamin F. Grewe
Researchers at the Institute of Neuroinformatics at the University of Zurich and ETH Zurich developed Equilibrium Fisher Control (EFC), a control-theoretic framework that reformulates continual learning as a control minimization problem within neural activity dynamics. EFC autonomously improves performance in class-incremental learning by dynamically coupling learning and preservation signals, achieving substantial gains over prior regularization methods in Class-IL benchmarks like Split-MNIST (51.4% accuracy vs. ~20% for EWC/oEWC/SI).
Key insights
- The framework's learning dynamics inherently lead to a "continual-natural gradient" property, implicitly encoding an approximation of the full prior-task Fisher Information Matrix from only diagonal storage.
- EFC's dynamic and sample-specific modulation effectively filters gradient interference that plagues traditional regularization methods, enabling successful task discrimination and convergence in Class-IL.
- The approach provides tighter forgetting bounds compared to standard backpropagation and EWC by capturing off-diagonal curvature and dynamically adapting to parameter movement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.