Reproduce “Eidetic Learning: an Efficient and Provable Solution to Catastrophic Forgetting” 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
Nicholas Dronen, Randall Balestriero
Eidetic Learning presents a method that provably prevents catastrophic forgetting in neural networks by strategically leveraging their excess capacity through structured pruning and freezing of task-specific neurons. It achieves competitive accuracy on benchmarks like Permuted MNIST (98.31%) and Sequential CIFAR100 (ResNet50, 75-80%) while requiring no rehearsal and supporting automatic task routing at inference time.
Key insights
- Catastrophic forgetting can be provably solved by strictly separating and protecting task-specific knowledge within network capacity through specific 'persistence' and 'resistance' conditions.
- Overparameterization in modern neural networks provides sufficient excess capacity that can be strategically exploited for continual learning without requiring rehearsal.
- Careful handling of components like Batch Normalization layer statistics and structured neuron pruning are crucial for preventing interference and preserving learned features across sequential tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.