Spiktidbrus och Generalisering i Spikande Neurala Nätverk

Spiking Neural Networks (SNNs) are brain-inspired computational models that process information using discrete spikes over time. While promising for energy-efficient and temporally sensitive tasks, SNNs are often prone to overfitting and limited generalization when trained using biologically inspired learning rules such as Spike-Timing-Dependent Plasticity (STDP). Motivated by this, we investigate whether introducing spike-time jitter, which is originally observed in STDP contexts, can serve as a regularization mechanism, even under supervised learning with surrogate gradient descent. Two jitter strategies were implemented: stochastic Gaussian jitter and annealed jitter, alongside a benchmark model without any jitter. All models were trained using a shared deep convolutional SNN architecture on the CIFAR-10 dataset. Performance was assessed based on the average test accuracy over the final training epochs to evaluate generalization stability. The results partially support the hypothesis that moderate jitter improves generalization, as models with moderate noise performed better than those with excessive or insufficient jitter. However, none of the jittered variants outperformed the benchmark model. These findings suggest that while jitter influences training dynamics and can affect robustness, its benefits are sensitive to parameter tuning and not guaranteed. This study contributes to the understanding of noise-aware learning in spiking models and indicates that spike-time jitter might serve as a complementary regularization tool when carefully integrated with other training strategies.