Reproduce “PathNet: Evolution Channels Gradient Descent in Super Neural Networks” 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

Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, Daan Wierstra

PathNet searches routes through a shared modular network, training selected modules with gradient descent and freezing a completed route’s weights before the next task. In noisy two-class MNIST tasks with 50% salt-and-pepper noise, it reached 0.998 training-set accuracy in a mean of 167 generations, versus 195 for independent fixed-path learning. These counts do not measure wall-clock time or total compute.

Key insights

  • A hybrid approach combining global architectural search (evolution) with local parameter optimization (gradient descent) is effective for adaptive resource allocation and transfer.
  • Fixing optimal pathways for learned tasks is a robust mechanism to prevent catastrophic forgetting while still allowing for knowledge reuse.
  • Reinitializing unused parameters (those not in the fixed pathway) is crucial for positive transfer, as it prevents negative interference from irrelevant pre-trained weights.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.