Spiking Neural Networks Generalization Bounds via Rademacher Complexity
A new theoretical paper on arXiv investigates generalization bounds of Spiking Neural Networks (SNNs) using Rademacher complexity. The study reveals that the empirical Rademacher complexity of SNNs is exponential to network depth and excitation probability, building on prior work that established excitation-dependent bounds. The research focuses on SNNs with stochastic firing and various integrate-and-fire schemes, aiming to clarify how well these bio-inspired models perform on unseen data. The findings highlight the challenge of depth in SNN generalization, with complexity growing exponentially. This work contributes to the theoretical understanding of SNNs, which are gaining traction in neuromorphic computing and sparse computation due to their efficiency. The paper is available on arXiv under ID 2605.02927.
Key facts
- Paper titled 'Generalization Bounds of Spiking Neural Networks via Rademacher Complexity'
- Published on arXiv with ID 2605.02927
- Investigates generalization bounds of SNNs using Rademacher complexity
- Empirical Rademacher complexity of SNNs is exponential to network depth and excitation probability
- Focuses on SNNs with stochastic firing and multiple integrate-and-fire schemes
- Builds on prior work showing excitation-dependent and architecture-related bounds
- SNNs are bio-inspired models used in neuromorphic computing and sparse computation
- Theoretical understanding of SNN generalization is still limited
Entities
Institutions
- arXiv