Reproduce “Metalearned Neural Memory” 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

Tsendsuren Munkhdalai, Alessandro Sordoni, Tong Wang, Adam Trischler

Microsoft Research developed Metalearned Neural Memory (MNM), an external memory mechanism for recurrent neural networks that models memory as a rapidly adaptable deep neural network function. This approach enables one-shot information storage and achieves state-of-the-art performance on complex reasoning tasks like bAbI question answering, solving 19-20 tasks with near-zero error and demonstrating improved sample efficiency in reinforcement learning.

Key insights

  • Modeling external memory as a deep neural network function allows for inherent compression and generalization beyond simple recall, enabling the memory to infer and adapt.
  • Meta-learning principles are applied to enable rapid, one-shot memory updates by training the controller to learn an effective 'writing' strategy, allowing the neural memory to quickly encode new key-value bindings.
  • A learned local update rule (MNM-p) significantly improves computational efficiency by enabling gradient-free, parallelizable memory parameter updates, overcoming the cost of higher-order gradients and sequential backpropagation.

Summary by alphaXiv.

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