Towards Transient Fault Mitigation Techniques Optimized for Compressed Neural Networks

Abstract

In the last decades, Neural Networks have become widespread in many applications, leading to improved performance in the form of compressed neural networks. However, these compressed neural networks have not been analyzed in terms of resilience to transient faults, and faults per transistor have been constantly increasing along with scaling technology nodes. Therefore, we need novel analysis and mitigation techniques which are optimized for compressed neural networks. We propose a 4-step methodology to design and test optimized toolset, algorithms, fault effect models and mitigations, covering the first two steps and marking the path for the remaining two.