Introduction

A flowchart representing the structure of an article on selecting appropriate XAI techniques. On the left, questions prompt reflection on AI model interpretability, linking to respective article sections in the center. These sections are 'Building Inherently Interpretable Models from the Start,' 'Global Techniques for Understanding Model Behavior,' 'Analyzing Feature Effects,' 'Learning About An Individual Prediction,' and 'Leveraging Examples for Clarity.' To the right, the flowchart shows a sequence of potential impacts like 'Interpretability/Explainability,' 'Transparency,' 'Accountability,' 'Fairness,' and 'Trustworthiness.' This visual guide outlines the article's flow from questioning interpretability needs to exploring tools and strategies, and understanding their downstream impacts.
      Diagram created using Mermaid.js code written by Areal Tal, 2023.

With the increased relevance of AI products, understanding how they work has never been more important. Depending on your needs, traditional standards for performance might not be sufficient. To be deployment-ready, the model might need to not only be capable of unearthing intricate relationships in data but to be interpretable. To consider what your interpretability needs are, you can reflect on the following questions:

  • How can you ensure that the model is designed with transparency and interpretability at its core?
  • Can you concisely explain the model’s overall decision-making process to various stakeholders?
  • Do you have evidence, beyond the general gauge provided by cross-validation, confirming the model's reliable performance on new unseen data?
  • Are you able to diagnose & address issues effectively through an understanding of the relationships between features and targets?
  • Are you able to clearly explain to stakeholders the rationale behind how predictions are made?
  • When you encounter an unexpected prediction, can you easily trace the reasoning that led to it?
  • Can you assess the model’s strengths & limitations and identify specific areas for targeted data & feature adjustments?


Each question finds its detailed exploration and resolution in this article. We will unpack the available tools and strategies, their applications, and the insights they offer. When to harness these tools, and the situations where they might not be as relevant, will also be addressed. The narrative will extend slightly beyond the technical realm to incorporate ethical considerations in AI and underscore the necessity of clear communication to make complex AI systems accessible to a diverse audience.


Questions that are not thoroughly explored, but lightly touched upon in this article:

  • Does the model’s level of interpretability facilitate the identification of potentially varied outcomes for different demographic groups?
  • How can effective communication bridge the gap between technical and non-technical stakeholders?


Understanding the hierarchy of explanatory techniques ensures that you can communicate to every stakeholder, from the least to the most technical, such that they can grasp the model's behavior at a depth that aligns with their expertise and needs. It's about providing the right level of detail to the right people, ensuring that comprehension and trust in the model’s predictions are as widespread as possible.

Join us in this detailed guide, tailored to cater to both those who develop models and those keen on the transparent deployment of AI in real-world scenarios.

 

Covered Not Covered (or only touched upon)
Explanations for classification and regression models. Explanations for unsupervised models and dimensionality reduction
Explanations for models using tabular data (tables) Explanations for models trained on images or unstructured data (e.g. saliency maps, text, audio). Explanations for large language models or reinforcement learning models
Model-agnostic post-hoc explanations (local attributions, global techniques, surrogate models, analyzing feature effects, counterfactual explanations, example-based explanations) Model-specific post-hoc explanations (e.g. activation maximization)
What you can do with XAI methods (e.g. which XAI tools can be used to diagnose model issues) How to use XAI methods to accomplish your goals (e.g. how to diagnose model issues with XAI tools)
Inherently interpretable models
Clustering on local attributions
Technical challenges and limitations of current XAI techniques
Interactive or user-driven explanation tools
Ethical implications and regulations specific to AI interpretability
Visualizing explanations
Real-time or online interpretability
Model monitoring
How to incorporate XAI throughout the SDLC of AI development
Explaining anything that is not a model (e.g. a dataset)
A comprehensive list of XAI techniques
Any code