New TTFS SNN training framework achieves SOTA accuracy on 5 benchmarks
Single-spike neural networks now rival traditional models with 99.48% on MNIST and stable training.
Spiking Neural Networks (SNNs) promise energy-efficient neuromorphic hardware thanks to their event-driven mechanisms. Among neural coding schemes, Time-to-First-Spike (TTFS) coding uses only one spike per neuron, offering extreme sparsity and ultra-low power consumption — but training TTFS networks from scratch has historically been unstable and inaccurate due to vanishing gradients and diminished signals. Researchers from Peking University, Intel Labs, and CNRS now introduce a comprehensive framework to overcome these barriers, achieving best-in-class accuracy across five standard benchmarks.
The proposed method tackles training instability head-on with tailored parameter initialization and normalization layers that preserve gradient flow. A novel temporal output decoder encourages earlier firing, reducing latency, while a critical insight shows that max-pooling violates the single-spike constraint — average-pooling must be used instead. The result is a step-by-step TTFS SNN that trains stably and quickly. On MNIST it hits 99.48% accuracy, on CIFAR10 90.56%, and on neuromorphic DVS Gesture 95.83%. This opens the door for deploying ultra-efficient, spike-based AI on edge devices without sacrificing accuracy.
- Single-spike TTFS SNNs trained from scratch with stable convergence using new initialization and normalization methods
- Achieves SOTA accuracy: 99.48% on MNIST, 90.56% on CIFAR10, 95.83% on DVS Gesture
- Identifies that average-pooling preserves single-spike constraints while max-pooling violates them, reducing latency with temporal output decoder
Why It Matters
Enables practical, ultra-low-power neuromorphic AI that rivals traditional accuracy, perfect for edge and IoT devices.