Reproduce “Learning to Continually Learn” 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
Shawn Beaulieu, Lapo Frati, Thomas Miconi, Joel Lehman, Kenneth O. Stanley, Jeff Clune, Nick Cheney
An algorithm called ANML, inspired by biological neuromodulation, enables artificial neural networks to learn sequences of tasks without catastrophic forgetting. It achieved 63.8% test accuracy across 600 sequential Omniglot tasks, outperforming prior methods like OML by a wide margin and reducing the relative performance drop from i.i.d. training to sequential learning from over 70% to 10%.
Key insights
- Directly meta-learning the ability to continually learn through a meta-loss can yield more effective solutions to catastrophic forgetting than manually designed heuristics.
- Context-dependent neuromodulatory gating, inspired by biological processes, effectively enables selective activation and plasticity within a neural network.
- Modulating activations during the forward pass, rather than just learning rates, plays a crucial role in reducing interference for both computation and learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.