Reproduce “Hebbian Learning with Global Direction” 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

Wenjia Hua, Kejie Zhao, Luziwei Leng, Ran Cheng, Yuxin Ma, Qinghai Guo

A Global-guided Hebbian Learning (GHL) framework combines local synaptic plasticity with a global directional signal derived from task loss, allowing bio-inspired learning to achieve performance competitive with backpropagation on datasets such as ImageNet and scale to very deep architectures.

Key insights

  • Incorporating a global directional signal (inspired by neuromodulation) into local Hebbian plasticity bridges the gap between biological plausibility and the ability to optimize towards global task objectives in deep networks.
  • The proposed framework demonstrates that dynamically adjusted update magnitudes from local Hebbian rules, combined with a simple sign-based global directional signal, are sufficient for competitive performance.
  • GHL's model-agnostic design reduces its sensitivity to specific network architectures and activation functions, enhancing its generalization compared to prior Hebbian methods.

Summary by alphaXiv.

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