Any time you simplify data using a summary statistic, you lose information. Model accuracy is no different. When simplifying your model's fit to a summary statistic, you lose the ability to determine where your performance is lowest/highest and why. In this post we discuss the code behind IBM's FreaAI, an efficient method for identifying data slices with low accuracy. In prior posts, we covered the method at a high level and did a deep dive into leveraging HPD to find areas of model weakness. Here, we will walk through a an MVP implementation of the paper for a binary classifier.

But here is the problem. Only a third of developers seem to know how to test these systems. And many companies have capabilities to only test them partially, risking the reliability of the system as a whole. We've decided to develop a way to know when AI algorithms work and when they don't. While it may not be critical if a movie recommendation is not that accurate, the results can be devastating if an algorithm performs poorly in an autonomous car or a medical app.