Testing for Multiple Faults in Deep Neural Networks

Deep Neural Networks (DNNs) implemented on hardware accelerators are vulnerable to various faults. This necessitates the development of efficient testing methodologies to detect them in DNN accelerators. In this work, we propose a test pattern generation approach to detect fault patterns in DNNs' synaptic weight value representations at a bit level. The experimental results show that the generated test patterns provide 100% fault coverage for targeted fault patterns. Besides, a high compaction ratio was achieved over different datasets and model architectures (up to 50×), and high fault coverage (up to 99.9%) for unseen fault patterns during the test generation phase.