Reproduce “Neural Arithmetic Units” 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

Andreas Madsen, Alexander Rosenberg Johansen

Researchers from Technical University of Denmark (DTU) introduce Neural Addition Units (NAU) and Neural Multiplication Units (NMU), which are more stable and interpretable components for neural networks to perform exact arithmetic. These units improve upon the existing NALU architecture by achieving significantly higher success rates on arithmetic tasks, such as 98% for multiplication compared to 31% for NAC*, and demonstrate superior convergence speed and robust extrapolation on sequential data.

Key insights

  • Constraining weight spaces (e.g., [-1, 1] for NAU, [0, 1] for NMU) and applying sparsity regularization can effectively bias neural networks to learn exact arithmetic operations.
  • The original NALU's multiplicative unit (NAC*) and its gating mechanism are inherently unstable and problematic, especially when handling inputs near zero or selecting operations.
  • Reliably learning division within a neural network structure is exceptionally difficult due to singularity issues, leading to its exclusion from the proposed NMU.

Summary by alphaXiv.

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