Redundant Lagrange Interpolation for Fault-Tolerant Winograd Convolution

Most recent advances in artificial intelligence mainly focus on further acceleration, reliability and security in safety-critical applications including 6G StarLink, and automotive. We find that Winograd Convolutions possess both potentialities on acceleration and fault tolerance theoretically. Therefore in this paper, we firstly take the single-fanout property of Winograd products to develop an effective and efficient single-interpolation error model. Secondly, we improve the state-of-the-art interpolation-point selection algorithm by taking into the arithmetic weights of integer multipliers. And finally we propose some single-error detection and correction methods using redundant-interpolation techniques. From both derivations and evaluations, the proposed redundant-interpolation Winograd convolutions not only possess a beautiful and concrete theoretical basis, but also an error correction capability with only quite few area overhead.