Explaining Explanations in AI

Abstract

Recent work on interpretability in machine learning and AI has focused on the\nbuilding of simplified models that approximate the true criteria used to make\ndecisions. These models are a useful pedagogical device for teaching trained\nprofessionals how to predict what decisions will be made by the complex system,\nand most importantly how the system might break. However, when considering any\nsuch model it's important to remember Box's maxim that "All models are wrong\nbut some are useful." We focus on the distinction between these models and\nexplanations in philosophy and sociology. These models can be understood as a\n"do it yourself kit" for explanations, allowing a practitioner to directly\nanswer "what if questions" or generate contrastive explanations without\nexternal assistance. Although a valuable ability, giving these models as\nexplanations appears more difficult than necessary, and other forms of\nexplanation may not have the same trade-offs. We contrast the different schools\nof thought on what makes an explanation, and suggest that machine learning\nmight benefit from viewing the problem more broadly.\n

Keywords

Affiliated Institutions

• University of Oxford GB
• University of Surrey GB
• The Alan Turing Institute GB

Related Publications