Interpretable by Design: Learning Predictors by Composing Interpretable Queries.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI (XAI)

Background:

• High-performance machine learning models often lack transparency, hindering adoption in critical sectors like healthcare.
• Interpretability is essential for building trust and enabling effective use of AI in risk-sensitive applications.
• Current explanation methods are often post-hoc and may not align with domain-specific user needs.

Purpose of the Study:

• To develop machine learning algorithms that are interpretable by design.
• To ensure explanations are expressed in domain- and task-specific language understandable to end-users.
• To minimize the number of queries required for accurate predictions while maintaining interpretability.

Main Methods:

• Models are based on user-defined, task-specific binary functions of data for clear interpretation.
• Queries are selected sequentially based on information gain to minimize prediction effort.
• A Variational Autoencoder (VAE) and Unadjusted Langevin MCMC algorithm are used to select informative queries, handling conditional dependencies.

Main Results:

• The proposed method enables online determination of query chains of variable depth to resolve prediction ambiguities.
• Experiments on computer vision and natural language processing tasks show the approach's effectiveness.
• The method demonstrates superiority over existing post-hoc explanation techniques.

Conclusions:

• Interpretable machine learning by design, using domain-specific language, is crucial for sensitive applications.
• The novel query-selection strategy enhances prediction accuracy and transparency.
• This approach offers a more robust and user-centric alternative to post-hoc explanations in AI.