New method boosts Spiking Neural Networks by 2x accuracy
Spiking Neural Networks just got 2x more accurate at tiny time steps
A new ANN-SNN conversion method reduces conversion error by aligning residual membrane potentials. It combines dynamic initial potential tuning with feature enhancement, introduces a regularization loss to correct truncation bias, and adds a grouped-convolution SCR-Conv2d layer that sharpens feature discrimination and stabilizes encoding at tiny timesteps. When integrated with the state-of-the-art QCFS baseline, the approach delivers consistent low-latency performance gains, generalizes to ReLU CNNs, ANN Transformers, and multi-threshold SNN variants, and verifies prominent accuracy improvements on CIFAR-10, CIFAR-100, and ImageNet at T=2, 4, and 8—with negligible extra computation overhead.
- Accuracy improvements up to 2x at T=2 timesteps on CIFAR-10/100 and ImageNet
- Introduces L_RMPD loss and SCR-Conv2d layer for residual membrane potential alignment
- Works with existing QCFS baseline with <1% additional computation overhead
Why It Matters
Enables real-time, ultra-low-power AI on neuromorphic chips with high accuracy at minimal latency.