Explainable artificial intelligence

Explainable artificial intelligence (XAI) comprises methods and systems that make an AI model’s decisions or outputs understandable to people, through interpretable models or explanations of model behavior.

Interpretable machine learning: Machine-learning methods whose structure or outputs can be understood directly by people. Its models provide explanations through their own decision structure rather than a separate explanation tool.
Machine learning: The study of algorithms that learn patterns from data to make predictions or decisions. XAI methods explain systems built using learned patterns rather than manually specified rules.
Algorithmic decision-making: The use of computational procedures to make or support decisions about people or resources. Explanations can help expose how automated decisions are produced.
Transparent model: A model whose structure or decision process can be inspected directly without a separate explanation method. It offers built-in transparency, unlike explanations added to an opaque model.
Explanation fidelity: The degree to which an explanation accurately represents the behavior of the model it describes. A simple explanation is useful only if it preserves relevant aspects of the model’s behavior.
Feature attribution: Methods that assign portions of a model’s prediction to its input features. Attribution methods show which inputs contributed to a particular prediction.
Black-box model: A model whose internal operation is difficult to inspect or understand from its inputs and outputs. Many XAI techniques target models whose reasoning is not directly legible.
Fairness in machine learning: The study of whether machine-learning systems distribute errors, benefits, or decisions equitably. Explanations can help investigate suspected patterns of unfair treatment.
Post hoc explanation: An explanation generated after a model has made a prediction, without changing that model. This approach separates the explanation tool from the model being explained.
Explanation evaluation: The assessment of explanations for properties such as faithfulness, usefulness, and human comprehensibility. There is no single agreed measure for whether an XAI explanation succeeds.

Linked from 9 pages