Analyzing Failures in Artificial Intelligent Learning Systems (FAILS)

Abstract

We learn more from analyzing failures in engineering than by studying successes. There is significant value in documenting and tracking AI failures in sufficient detail to understand their root causes, and to put processes and practices in place toward preventing similar problems in the future. Similar efforts to track and record vulnerabilities in traditional software led to the establishment of National Vulnerability Database, which has contributed towards understanding vulnerability trends, their root causes, and how to prevent them [1], [3].