Making Convolutions Resilient Via Algorithm-Based Error Detection Techniques

Convolutional Neural Networks (CNNs) are being increasingly used in safety-critical and high-performance computing systems. As such systems require high levels of resilience to errors, CNNs must execute correctly the presence hardware faults. Full duplication provides needed assurance but incurs a prohibitive 100 percent overhead. In this article, we focus on algorithmically verifying convolutions, most resource-demanding operations CNNs. We use checksums verify convolutions. identify feasibility performance related challenges that arise detecting errors convolutions optimized CNN inference deployment platforms (e.g., TensorFlow or TensorRT GPUs) fuse multiple network layers reduced-precision operations, demonstrate how overcome them. propose evaluate variations algorithm-based error detection (ABED) techniques offer implementation complexity, runtime overhead, coverage trade-offs. Results show ABED can detect all transient might otherwise corrupt output with low overheads (6-23 percent). Only about 1.4 total computations not protected by ABED, which be duplicated for full protection. compute-intensive duplicating rest at least 1.6× throughput compared duplication.