Fault-Tolerant Accelerators for Spiking Neural Networks

What is it about?

Spiking Neural Networks (SNNs) have shown great potential for achieving high accuracy in machine learning applications (e.g., classification tasks) with very low processing power/energy due to their sparse spike-based operations. To maximize the performance efficiency of SNN processing, specialized hardware accelerators have been employed. However, such SNN accelerators are vulnerable to soft errors due to energetic particle strikes during the operational time. In this paper, we propose a novel methodology, SoftSNN, that mitigates soft errors in the synapses and neurons of SNN accelerators without re-execution, hence maintaining high accuracy with low latency and energy overheads. To detect and mitigate soft errors at run time, our SoftSNN bounds the weight values and protects the neurons from faulty operations. Furthermore, our SoftSNN employs lightweight hardware enhancements to efficiently support the proposed technique.

Featured Image

Photo by Laura Ockel on Unsplash

Photo by Laura Ockel on Unsplash

Why is it important?

Hardware computing systems like SNN accelerators are subjected to soft errors due to energetic particle strikes, which may come from cosmic rays or packaging materials. These errors manifest as bit flips at the hardware layer and can propagate to the application layer, resulting in incorrect outputs (e.g., wrong classification) that can cause significant accuracy degradation. This condition poses serious reliability threats for SNN-based systems, especially when deployed in safety-critical applications (e.g., medical analysis, automotive, etc.). Towards this, our work is the first effort that studies the impact of soft errors on SNN accelerators and provides a cost-effective mitigation solution.

Perspectives