New techniques for assessing and enhancing the reliability of DNNs

Nowadays, machine learning (ML) algorithms are being exploited in a variety of applications, from health care to avionics, from computer vision to natural language processing. Furthermore, an increasing number of tasks exploiting deep learning are being performed on edge devices such as smartphones or drones through new programming paradigms, like dynamic neural networks or split computing. Semiconductors manufacturers produced AI accelerators, such as GPUs, able to provide outstanding power consumption and latency performances, leaving out reliability which is crucial factor in safety-critical applications, such as self driving vehicles. In fact, new generation systems on chip, produced using 7 nm technology (or later), are very likely to be effected by manufacturing defects during production or fault activation due to electromigration or aging during utilization. This work aims at evaluating the effects introduced by hardware permanent "stuck at" model) faults in GPU architectures while executing DNN workloads. Furthermore, it proposes new techniques for software-based detection and hardware hardening.