Ablation Study of Hyperparameters in Spiking Neural Networks: A Case Study for Event-Based Gesture Recognition

Spiking Neural Networks (SNNs) are a promising paradigm for neuromorphic computing, particularly for processing event-based sensory data. Their performance, however, strongly depends on several design choices, including input encoding, preprocessing, neuron model, and training hyperparameters. We present an ablation study aimed at evaluating the impact of these factors on an event-based gesture recognition task using the IBM DVS128 Gesture dataset. Specifically, we analyze the effects of denoising and to-frame preprocessing parameters, together with the role of first-order and second-order Leaky Integrate-and-Fire (LIF) neuron configurations, including threshold, decay terms, and reset mechanism. Experimental results show that preprocessing parameters have a strong and non-linear influence on performance, with intermediate values providing the best trade-off between optimization and generalization. In addition, learned neuronal parameters, particularly thresholds and decay coefficients, improve out-of-sample accuracy in several configurations. Our findings indicate that SNN performance arises from a complex interaction among data representation, temporal aggregation, and neuron dynamics. Overall, our study highlights the importance of systematic hyperparameter tuning and provides practical insights for the design of effective SNNbased models for neuromorphic vision applications.