Research map / Probabilistic and scientific modeling / Dynamical systems
UltraLIF: Fully Differentiable Spiking Neural Networks via Ultradiscretization and Max-Plus Algebra
Jose Marie Antonio Miñoza
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace surrogate-gradient spikes with ultradiscretized (LSE) neurons | UltraLIF family · SHD (T=1) | accuracy (%): 40.02 -> 51.24 | Improved |
| replace surrogate-gradient spikes with ultradiscretized neurons | UltraLIF family · MNIST (T=1) | accuracy (%): 95.58 -> 95.67 | No clear effect |
| increase timesteps to T>=10 | UltraLIF family vs surrogate baselines · SHD, N-MNIST | accuracy (%) | Hurt |
| add sparsity penalty lambda=0.1 (ablation) | UltraPLIF · CIFAR-10 (T=1) | accuracy (%): 43.27 -> 43.60 (spike rate 0.458 -> 0.240) | Improved |
| add sparsity penalty lambda=0.1 (ablation) | UltraPLIF · CIFAR-10 (T=30) | accuracy (%): 46.58 -> 46.98 | Improved |
| add sparsity penalty lambda=0.1 (ablation) | UltraDPLIF · CIFAR-10 (T=10) | accuracy (%): 45.75 -> 45.32 | Hurt |
| use hard binary spikes at inference (ablation) | UltraLIF family · MNIST (T=1) | accuracy (%): 95.67 -> 93.72 | Hurt |
| use hard binary spikes at inference (ablation) | UltraLIF family · CIFAR-10 (T=1) | accuracy (%): 43.27 -> 43.35 | No clear effect |
| learnable temperature epsilon replaces fixed surrogate sharpness | UltraPLIF on ResNet50 spiking head · CIFAR-10 (ResNet50, T=1) | accuracy (%): 31.83 -> 92.78 | Improved |
| use ultradiscretized neurons in a 2-layer conv SNN | UltraDPLIF · CIFAR-10 (Conv 2L, T=1) | accuracy (%): 74.37 -> 70.54 | Hurt |
| use ultradiscretized neurons in ResNet18 spiking head | UltraDPLIF · CIFAR-10 (ResNet18, T=1) | accuracy (%): 93.12 -> 93.37 | No clear effect |
| use ultradiscretized neurons in fully connected nets of depth 1 to 3 | UltraPLIF/UltraDPLIF · CIFAR-10 (FC 1L) | accuracy (%): 41.79 -> 46.40 | Improved |
Closest papers by meaning
- P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.89)
- A$^2$SG: Adaptive and Asymmetric Surrogate Gradients for Training Deep Spiking Neural Networks (ICML 2026 · similarity 0.89)
- Neural Dynamics Self-Attention for Spiking Transformers (ICLR 2026 · similarity 0.88)
- NDOT: Neuronal Dynamics-based Online Training for Spiking Neural Networks (ICML 2024 · similarity 0.88)
- Sign Gradient Descent-based Neuronal Dynamics: ANN-to-SNN Conversion Beyond ReLU Network (ICML 2024 · similarity 0.88)
- SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
