ICSRS 2022

?? Abstract: The ability of CNNs to efficiently and accurately perform complex functions, such as object detection, has fostered their adoption in safety-related autonomous systems. These algorithms require high computational performance platforms that exploit high levels of parallelism. The detection, control, and mitigation of random errors in these underlying high computational platforms become a must according to functional safety standards. In this paper, we propose protecting, with a catalog of diagnostic techniques (??), the most computationally expensive operation of the CNNs, the matrix multiplication. However, this protection entails a performance penalty, and the complete CNN protection may be unaffordable for those systems operating with strict real-time constraints. This paper proposes a three-stage methodology to selectively protect CNN layers to achieve the required Diagnostic Coverage (DC) and performance trade-off:

1. Sensitivity analysis to misclassification per CNN layers using a statistical fault injection campaign.
2. Layer-by-layer performance impact and DC analysis.
3. Selective layer protection.

Furthermore, we propose a strategy to effectively compute the achievable DC of large matrices implemented on GPUs. Finally, we apply the proposed methodology and strategy in Tiny YOLO-v3, an object detector based on CNNs.

?? Link to previous work where we define the catalog of diagnostic techniques and how we implement it in CUTLASS (a high-performance GPU-based MMM implementation), whose title is "On the Safe Deployment of Matrix Multiplication in Massively Parallel Safety-Related Systems"

?? ?? Additionally, I refer the reader to the PowerPoint employed at the conference. Constructive feedback?is highly appreciated! Feel free to use, enjoy and spread my materials! Let me know if you are interested in a deep dive into this material and I will forward it to you ?? [email protected]

#Safety #Cnn #FaultDetection #GPUs #MatrixMultiplication