Reproduce “MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking 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

Ruokai Yin, Yuhang Li, Abhishek Moitra, Priyadarshini Panda

Yale University's Intelligent Computing Lab developed MINT, a quantization scheme that achieves average 93% memory, 90% computation energy, and up to 96% hardware area reductions for Spiking Neural Networks while maintaining competitive accuracy by quantizing both weights and membrane potentials and eliminating the need for hardware multipliers during inference.

Key insights

  • Sharing a single learnable scaling factor across weights, membrane potentials, and thresholds enables the transformation of SNN inference into purely integer operations, eliminating the need for costly hardware multipliers.
  • Quantizing membrane potentials alongside weights is critical for comprehensive memory reduction, as they become the dominant memory overhead, especially with increasing batch sizes.
  • Careful training strategies, including gradient scaling for the learnable shared scaling factor, are essential to maintain high accuracy under aggressive low-bit quantization.

Summary by alphaXiv.

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