A machine-learning-guided framework for fault-tolerant DNNs

Deep Neural Networks (DNNs) show promising performance in several application domains. Nevertheless, DNN results may be incorrect, not only because of the network intrinsic inaccuracy, but also due to faults affecting the hardware. Ensuring the fault tolerance of DNN is crucial, but common fault tolerance approaches are not cost-effective, due to the prohibitive overheads for large DNNs. This work proposes a comprehensive framework to assess the fault tolerance of DNN parameters and cost-effectively protect them. As a first step, the proposed framework performs a statistical fault injection. The results are used in the second step with classification-based machine learning methods to obtain a bit-accurate prediction of the criticality of all network parameters. Last, Error Correction Codes (ECCs) are selectively inserted to protect only the critical parameters, hence entailing low cost. Thanks to the proposed framework, we explored and protected two Convolutional Neural Networks (CNNs), each with four different data encoding. The results show that it is possible to protect the critical network parameters with selective ECCs while saving up to 79% memory w.r.t. conventional ECC approaches.