SoftSNN: Low-Cost Fault Tolerance for Spiking Neural Network Accelerators under Soft Errors
TimeTuesday, July 12th1:52pm - 2:15pm PDT
Location3005, Level 3
Event Type
Research Manuscript
AI/ML Design: System and Platform
DescriptionSpiking Neural Network (SNN) accelerators are vulnerable to soft errors that can corrupt the weights and neuron operations, thereby degrading accuracy. State-of-the-art have not thoroughly explored specialized soft-error mitigation techniques for SNNs. This paper proposes SoftSNN, a lightweight methodology to mitigate soft errors without re-execution. SoftSNN analyzes the SNN characteristics under soft errors and leverages this analysis to bound the weights and protect the neuron operations. We also propose necessary hardware enhancements to efficiently support the proposed techniques. SoftSNN maintains accuracy degradation below 3%, while reducing latency and energy by up to 3x and 2.3x, respectively compared to the re-execution.