The ability of Convolutional Neural Networks (CNNs) to accurately process\nreal-time telemetry has boosted their use in safety-critical and\nhigh-performance computing systems. As such systems require high levels of\nresilience to errors, CNNs must execute correctly in the presence of hardware\nfaults. Full duplication provides the needed assurance but incurs a prohibitive\n100% overhead. Algorithmic techniques are known to offer low-cost solutions,\nbut the practical feasibility and performance of such techniques have never\nbeen studied for CNN deployment platforms (e.g., TensorFlow or TensorRT on\nGPUs). In this paper, we focus on algorithmically verifying Convolutions, which\nare the most resource-demanding operations in CNNs. We use checksums to verify\nconvolutions, adding a small amount of redundancy, far less than\nfull-duplication. We first identify the challenges that arise in employing\nAlgorithm-Based Error Detection (ABED) for Convolutions in optimized inference\nplatforms that fuse multiple network layers and use reduced-precision\noperations, and demonstrate how to overcome them. We propose and evaluate\nvariations of ABED techniques that offer implementation complexity, runtime\noverhead, and coverage trade-offs. Results show that ABED can detect all\ntransient hardware errors that might otherwise corrupt output and does so while\nincurring low runtime overheads (6-23%), offering at least 1.6X throughput to\nworkloads compared to full duplication.\n