FireNN: Neural Networks Reliability Evaluation on Hybrid Platforms

Abstract

Modern neural network complexity has grown dramatically in recent years, leading to the adoption of hardware-accelerated solutions to cope with the computational power required by the new network architectures. The possibility to adapt the network size and performance to different platforms enhanced the interests of safety-critical applications such as automotive and avionic. Today, the reliability evaluation of neural networks is still premature and requires platforms to measure the safety standards required by mission-critical applications. For this reason, the interest in studying the reliability of neural networks is growing. In this work, we propose a new approach for evaluating the resiliency of neural networks by using programmable hardware of hybrid platforms. The approach relies on the reconfigurable hardware for emulating the target hardware platform and performing the fault injection process. The main advantage of the proposed approach is to involve the on-hardware execution of the neural network in the reliability analysis without modifying the hardware implementation of the network under analysis, and addressing specific fault models. The implementation of FireNN, the platform based on the proposed approach is detailly described in the paper. Experimental analyses are performed using fault injection on the AlexNet Convolutional Neural Network. The analyses are carried out by means of the FireNN platform and the obtained results are compared with the outcome of traditional software-level evaluations. Results are commented taking into account the insight into the hardware level achieved by using the FireNN platform.