• Home
• Abstract submission
• Search abstract
• Topic keywords
• Streams
• F.A.Q.
• Program
• Help

Online presentation.

423. Exploiting Auto-Encoders for Explaining Black Box Classifiers

Invited abstract in session MD-32: Fair and explainable models 2, stream Multiple Criteria Decision Analysis.

Area: Decision support

Authors (first author is the speaker)

Abstract

Artificial Intelligence (AI) has nowadays a tremendous socio-economic impact and a pervasive adoption in every field of modern society. Many applications in different fields, such as credit score assessment, medical diagnosis, etc., are based on AI systems. Unfortunately, these systems often reach their impressive performance through obscure machine learning models that "hide" the logic of their internal decision processes to humans. For this reason, these models are called black boxes. The missing interpretability of black boxes is a limitation to AI adoption in socially sensitive contexts. As a consequence, the research in eXplainable AI (XAI) has recently caught much attention. A promising line of research in XAI exploits auto-encoders for explaining black box classifiers working on non-tabular data (images, time series, etc.). The ability of autoencoders to compress any data in a low-dimensional tabular representation, and then reconstruct it with negligible loss, provides the great opportunity to work in the latent space for the extraction of meaningful explanations, for example, through the generation of new synthetic samples that can be fed to a black-box to understand where its decision boundary lies. We discuss recent XAI solutions based on autoencoders that enable the extraction of meaningful explanations composed by factual and counterfactual rules, and by exemplars and counter-exemplars, offering a deep understanding of the local decision of the black box.

Keywords

• Artificial Intelligence
• Decision Support Systems
• Ethics

Status: accepted