Reproduce “Contraction Properties of the Global Workspace Primitive” 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

Michaela Ennis, Leo Kozachkov, Jean-Jacques Slotine

Researchers at MIT and Harvard advanced the "RNNs of RNNs" framework, achieving 68.91% accuracy on sequential CIFAR10, a new state-of-the-art for provably stable recurrent neural networks. This work also derived new theoretical conditions for stable modular architectures with nonlinear inter-area connections, demonstrating enhanced robustness and interpretability, particularly with Global Workspace topologies.

Key insights

  • Sparse inter-area connectivity is crucial for scaling up multi-area RNNs, enabling performance gains with an increased number of modules without a prohibitive increase in dense connections.
  • The Global Workspace (GW) topology provides a highly effective and parameter-efficient architecture for stable RNNs, exhibiting superior robustness to module ablation and clearer interpretability through specialized subnetwork roles.
  • Applying advanced control theory tools like contraction analysis can relax stability constraints in RNNs, enabling the design of more complex, neurobiologically-inspired architectures with provable guarantees.

Summary by alphaXiv.

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