Reproduce “Working memory facilitates reward-modulated Hebbian learning in recurrent 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
Roman Pogodin, Dane Corneil, Alexander Seeholzer, Joseph Heng, Wulfram Gerstner
The authors combine a two-dimensional bump-attractor network with a recurrent reservoir and train a linear readout using reward-modulated Hebbian learning. The attractor supplies evolving internal activity, while updates are applied after each trial. On ten-second signals, the system reaches normalized cross-correlation 1 after 200 trials; without attractor input it fails, showing dependence on added temporal structure rather than animal use.
Key insights
- A dynamic working memory component can significantly enhance the efficiency and robustness of biologically plausible reward-modulated Hebbian learning rules in recurrent neural networks.
- Structured internal representations, like those from working memory, allow learning to proceed effectively even with delayed feedback and over long temporal dependencies.
- The integrated architecture facilitates multi-task learning within a single network by leveraging unique neural trajectories generated by the working memory for different tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.