Fault Mitigation for SNN Classification of Neuromorphic Event Streams with Radiation-Induced Noise

The pursuit of autonomy for on-orbit processing drives the evolution of deep-learning (DL) systems for space applications. Hindered by limited compute capabilities on constrained embedded platforms, alternative technologies are being explored. Amidst the evolution of these DL systems, spiking neural networks (SNNs) coupled with neuromorphic sensing inputs offer advantages compared to traditional DL frameworks, as SNNs provide lower latency, memory, computation requirements, and subsequent lower power usage. These benefits of neuromorphic systems make them highly amenable to space applications. Given the limited space-flight heritage of neuromorphic systems, the characterization of radiation effects on such algorithms and sensors is valuable for reliable implementations in space. Specifically, characterization of the adverse effects of radiation on neuromorphic systems can lead to the development of effective fault-mitigation strategies. This research therefore aims to characterize the effects of simulated, radiation-induced noise on the classification capability of SNNs and explore contemporary fault-mitigation tactics. A fault-injection tool modeled from the results of a neutron radiation test is leveraged to simulate noise. Using this tool, noise is injected into event-based sensor data and subsequently passed to an SNN for classification. This simulated neuromorphic system provides insight into the radiation-resiliency of SNNs as network hyperparameters such as width and depth are varied, as well as identifies the limitations of current denoising methods. A novel event-rate denoising method is also introduced to increase the performance of SNN classification under higher noise rates, outperforming the standard spatio-temporal filters currently used with event data suitable for radiation environments. The proposed filtering methodology outperforms standard spatio-temporal filtering methodology with classification accuracy 1.7 × higher at high simulated noise rates (100 events per second) on the complex DVS Gesture dataset, with improved performance for a wider range of noise rates. The intrinsic reliability of SNNs, coupled with the enhanced event-rate filtering methodology offers a formidable DL alternative.