Exploring Image Selection for Self-Testing in Neural Network Accelerators

Hardware accelerators are essential to the accommodation of computation and memory-intensive neural network (NN) applications on resource-constrained edge devices. While hardware accelerators facilitate fast and energy-efficient convolution operations, their accuracy is threatened by various types of faults in their on-chip and off-chip memories, where millions of NN weights are held. To achieve fast and in-time fault detection, a self-test process that periodically runs a small set of test images in the accelerator can be adopted. This paper focuses on developing and comparing multiple numerical score based test image selection strategies. Various image selection criteria are studied, including output probability distribution, gradient sensitivity, and neuron coverage. Experimental studies show that images selected based on the output probability distribution offers high fault detection accuracy over a wide range of fault rates as well as low computation complexity. The small set of test images allows for real-time monitoring of the healthiness of DNN accelerators as well as the subsequent recovery and self-healing process.