Explainable AI (XAI)
Also see index ethical AI .
The term XAI is coined by DARPA, cf. TFCA lecture 1: intro to XAI . See specifically D. Gunning, Explainable Artificial Intelligence (xAI), Technical Report, Defense Advanced Research Projects Agency (DARPA), 2017.
Definitions
- "explainability is defined as the ability a model has to make its functioning clearer to an audience." source, p.100
General
For an overview of notes on the more broader and philosophical question what "explanation" is: What is explanation? .
That index note also contains references to Bayesianism.
- Larsson and Heintz - Transparency in artificial intelligence
- Mittelstadt, Russell, Wachter - Explaining Explanations in AI
- Lipton - The mythos of model interpretability
- Trade-off between explainability and accuracy
- Why do we need explainability of AI?
- Explainability in support of algorithmic choice
- TODO Miller (2019) - Explanation in AI insights from social sciences
- TODO Rudin (2019) - Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and use Interpretable Models Instead
Technical Methods
Feature attribution
Example-based and counterfactual
- Mothilal (2019) - Explaining machine learning classifiers through diverse counterfactual explanations
- Paper behind DiCE library
- Van der Waa (2021) - Evaluating XAI Comparison of rule-based and example-based explanations
- Provides recommendations for user evaluations of XAI
- Results show that although this is often thought, two major contrastive explanation methods in XAI actually do not improve system understanding
- Wachter et al. 20XX
- SEDC algorithm for textual data: Martens & Provost 2014, "Explaining data-driven document classifications"
- Ramon, Martens, Provost, Evgeniou 2020, "A comparison of instance-level counterfactual explanation algorithms for behavioral and textual data: SEDC, LIME-C and SHAP-C"
- LIME-C and SHAP-C are proposed extensions to the "leave out" method of generating
Case-based reasoning
Case-based reasoning is also one of the hybrid strategies to explainable AI, e.g. identified in the deep learning taxonomy of BarredoArrieta. I'm collecting papers on that topic in a note on case-based reasoning.
Evaluation
- Van der Waa (2021) - Evaluating XAI Comparison of rule-based and example-based explanations
- Provides recommendations for user evaluations of XAI
- Results show that although this is often thought, two major contrastive explanation methods in XAI actually do not improve system understanding