Kernel Surrogates are the Scales of Justice for AI
In this paper Dr. Gary Nan Tie and Dr. Bob Mark introduce kernel surrogates, which are mathematical functions that mimic AI behavior on a sample of its input-output for the purposes of understanding and explaining AI algorithm results. The known provable properties of a readily calculated surrogate enable assessment, communication, and fair application of AI algorithms when AI results are a black box, computationally expensive, or proprietary. They have focused on using kernel surrogates to model results from black box AI financial algorithms that are deployed in risk management. The scope, however, is much broader since kernel surrogates can be applied to any regression or classification model output. Kernel surrogates enable domain experts to understand and communicate black box model results to regulators, companies, and consumers.