The XAI Discovery Platform provides a customizeable interface for exploring image classification data sets. Its goal is to help explore strengths and weaknesses of an image classifier, focusing on consistent errors, and patterns that help predict performance. It does not attempt to provide explanations on its own, but rather helps users understand the things that need to be explained, and to test and compare ideas.
The discovery platform allows a developer to better understand the global competency of a system through exploring patterns, contrasts, and edge cases. We have demonstration systems using the MNIST data set (10 classes of hand-written characters) as well as the FGVC aircraft data set (roughly 50 classes of aircraft).
The system is model-agnostic, but requires: (1) ground truth about images; and (2) machine classification with probabilistic or rank-order over the top-10 classification categories for each case. This enables searching by similarity/error patterns to find similar cases.
The system has been tested on data sets with 2000-5000 cases. Larger data sets may need to be sampled from to enable good performance.