Reproduce “UltraLIF: Fully Differentiable Spiking Neural Networks via Ultradiscretization and Max-Plus Algebra” 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
Jose Marie Antonio Miñoza
A framework called UltraLIF creates fully differentiable Spiking Neural Networks using ultradiscretization and max-plus algebra, eliminating the forward-backward mismatch in SNN training. This approach yielded notable accuracy improvements, such as an 11.22% gain on the SHD dataset and a 7.96% increase on DVS-Gesture for single-timestep tasks, alongside flexible energy efficiency management.
Key insights
- The framework establishes a pioneering link between SNN dynamics and tropical algebraic geometry, offering a mathematically principled and consistent approach to SNN differentiability.
- Employing identical smooth functions (LSE and soft sigmoid) for both forward and backward passes resolves the fundamental forward-backward mismatch of surrogate gradient methods, enabling more stable gradient-based optimization.
- A learnable temperature parameter "," in the LSE and sigmoid functions facilitates automatic curriculum learning, allowing the model to adapt its activation sharpness during training for optimal performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.