XOResNet uses XOR meta-residuals to boost deep SNN learning
Novel OR-ADD shortcut and XOR residual selection beat existing deep SNNs on four benchmarks.
Spiking neural networks (SNNs) are gaining traction for their energy efficiency and biological plausibility, but training deep SNNs remains challenging due to spike redundancy and information loss in residual connections. A new paper by Jianfang Wu and Junsong Wang introduces XOResNet, a deep SNN architecture that tackles these issues with two key innovations: an OR-ADD (OA) shortcut connection that merges output spikes and currents from two branches, and XOR meta-residuals that select only the most useful pre-learning residuals using an Exclusive-OR operation. This combination reduces redundant learning in the backbone branch while maintaining strong gradient flow.
Tested on Fashion-MNIST, CIFAR-10, CIFAR-100, and miniImageNet, XOResNet consistently outperforms existing state-of-the-art deep SNNs trained via gradient descent. The results validate that the OA shortcut prevents information loss in non-identity mappings and the XOR meta-residuals eliminate spike redundancy in identity mappings. The paper provides 33 pages of analysis with 12 figures and 7 tables, offering new architectural insights for building high-performance neuromorphic systems. This work could accelerate the adoption of SNNs in low-power edge computing and real-time AI applications.
- XOResNet introduces OR-ADD shortcut to merge outputs from two residual branches, reducing spike redundancy and information loss.
- XOR meta-residuals select pre-learning residuals using Exclusive-OR operation, cutting redundant computation in the backbone.
- Outperforms state-of-the-art gradient-optimized deep SNNs on four datasets: Fashion-MNIST, CIFAR-10, CIFAR-100, miniImageNet.
Why It Matters
Paves way for deeper, more efficient spiking neural networks in neuromorphic computing and edge AI systems.