UltraLIF: Fully Differentiable Spiking Neural Networks via Ultradiscretization and Max-Plus Algebra

Researchers have introduced UltraLIF, a new framework that makes Spiking Neural Networks (SNNs) fully differentiable and trainable with standard backpropagation. It replaces problematic surrogate gradients with a mathematical technique called ultradiscretization, derived from tropical geometry. This eliminates the forward-backward mismatch during training. Experiments on six benchmarks, including neuromorphic vision and audio, show performance improvements over existing methods, with gains especially strong in single-timestep and temporal data settings.