Noise resilience in brain decoding driven by spiking neural networks for responsive robot control

Brain computer interface (BCI) is a promising technology, expanding on many fields ranging from medical to commercial applications. The diffusion of this technique is compromised by the complexity of the treated data and the unavoidable presence of noise in those systems. This study contributes to the solution of this problem by proposing a BCI pipeline that relies on the use of spiking neural networks (SNNs), quantifying their robustness to noise in the context of brain decoding through the classification of Electroencephalography (EEG). We benchmark a shallow convolutional network against two spiking counterparts based on leaky integrate-and-fire (LIF) and adaptive LIF neurons on the four-class BNCI 2014-001 motor imagery dataset. The conventional model has a slight advantage in clean conditions, but it degrades dramatically once the input is corrupted. In contrast, SNN demonstrates remarkable robustness under signal degradation. Under global Gaussian noise, CNN accuracy drops by 58.56%, whereas Adaptive LIF SNN accuracy drops by only 0.80%. Similarly, when trained on recordings with one electrode removed and tested on the complete montage, CNN loses 36.00% of its accuracy, compared with 0.34% for the SNN. With semi-simulated ocular, muscular and cardiac artifacts the gap is smaller, but the spiking models still degrade less than the CNN for every artifact type. This robustness is obtained at equal model size with 18.0 of the 24.6 MMAC required per trial by the baseline replaced by accumulate-only, spike-driven operations. Finally, the spiking model was integrated in a full decoding-to-actuation pipeline driving a Franka Emika Research 3 robot from a replayed EEG stream. These findings indicate that, compared with a dense counterpart, spiking computation substantially reduces the sensitivity of EEG decoding to signal corruption, which supports the use of SNNs in deployments where robustness matters more than peak accuracy.