This note is a summary of the survey paper by Arun et al.1
Definitions and preliminaries
Explainable artificial intelligence (XAI) is a field of artificial intelligence (AI) that promotes a set of tools, techniques, and algorithms that can generate high-quality interpretable, intuitive, human-understandable explanations of AI decisions.
Importance of XAI
- Improves transparency: XAI improves transparency and fairness by creating a human-understandable justification to the decisions.
- Improves trust: A scientific explanation or logical reasoning for a sub-optimal decision is often better than a highly confident decision that lacks explanations.
- Improves model bias understanding and fairness: XAI promotes fairness and helps mitigate biases introduced to the decision either from input datasets or poor architecture.
Keyword definitions
- Interpretability
- Desirable quality or feature of an algorithm which provides enough expressive data to understand how the algorithm works
- Interpretation
- Simplified representation of a complex domain to meaningful concepts which are human-understandable and reasonable
- Explanation
- Additional meta information to describe the feature importance or relevance of an input instance towards a particular output classification
- White-box
- A model is considered a white-box if the model parameters and the model architecture information are known
- Black-box
- A model is considered a black-box if the model parameters and the model architecture information are hidden
- Transparency
- A model is considered transparent if it is expressive enough to be human-understandable
- Trustability
- A measure of confidence, as human end-users, in the intended working of a given model in dynamic real world environments
- Bias
- Indicates the disproportionate weight, prejudice, favour, or inclination of the learnt model totowards subsets of data
- Fairness
- The quality of a learnt model in providing impartial decisions without favouring any populations in the input data distribution
Categories
Scope of explanation
Scope of explanations can be either local or global, or even both.
Local explanation
Locally explainable methods focus on a single input data instance to generate explanations by utilising the different data features.
- Activation maximisation
- Image-specific class saliency
- Layer-wise relevance propagation
- Local interpretable model-agnostic explanations
- Shapley additive explanations
References
Footnotes
-
Das, A., & Rad, P. (2020). Opportunities and challenges in explainable artificial intelligence (xai): A survey. arXiv preprint arXiv:2006.11371. ↩