ExplainabilityTransparencyTrustworthinesscs.LG
A general approach to compute the relevance of middle-level input features
Andrea Apicella, Salvatore Giugliano, Francesco Isgrò, Roberto Prevete
October 16, 2020
ICPR Workshops
7 citations
Key Takeaways
-
Middle-level explanations bridge a gap between raw data and human understanding: The paper addresses a key problem in AI explainability—that low-level explanations (like individual pixels in images) require significant effort from users to understand what the model actually "sees." Middle-level explanations using more intuitive features (like shapes or objects) make AI decisions more accessible to non-experts.
-
A new framework fills a methodological gap: The authors propose the first general approach to properly evaluate which middle-level features are actually relevant to a model's decisions. Previously, there was no standardized way to measure this, which limited the reliability of middle-level explanations across different applications.
-
This work supports more trustworthy AI systems: By enabling clearer explanations of how ML models make decisions, this framework helps organizations and users better understand and audit AI systems—a critical requirement for responsible AI deployment in high-stakes domains like healthcare or criminal justice.
-
The approach is broadly applicable: The framework is designed as a "general approach" rather than being limited to specific tasks (like image classification), suggesting it could improve explainability across diverse AI applications and industries.
Abstract
This work proposes a novel general framework, in the context of eXplainable Artificial Intelligence (XAI), to construct explanations for the behaviour of Machine Learning (ML) models in terms of middle-level features. One can isolate two different ways to provide explanations in the context of XAI: low and middle-level explanations. Middle-level explanations have been introduced for alleviating some deficiencies of low-level explanations such as, in the context of image classification, the fact that human users are left with a significant interpretive burden: starting from low-level explanations, one has to identify properties of the overall input that are perceptually salient for the human visual system. However, a general approach to correctly evaluate the elements of middle-level explanations with respect ML model responses has never been proposed in the literature.