Explainable AI via learning to optimize.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning models often function as "black boxes", hindering trust and adoption in critical applications.
• There is a growing need for explainable artificial intelligence (XAI) that provides transparency and interpretability.
• Encoding prior knowledge and flagging untrustworthy inferences are key challenges in current XAI.

Purpose of the Study:

• To develop concrete tools for XAI that allow encoding of prior knowledge.
• To propose methods for flagging untrustworthy inferences from machine learning models.
• To enhance the transparency and interpretability of data-driven algorithms.

Main Methods:

• Utilized the "learn to optimize" (L2O) methodology, framing inferences as data-driven optimization problems.
• Implemented L2O models that directly incorporate prior knowledge and offer theoretical guarantees.
• Introduced interpretable certificates for verifying the trustworthiness of model inferences.

Main Results:

• L2O models demonstrated straightforward implementation and direct encoding of prior knowledge.
• Theoretical guarantees, such as constraint satisfaction, were achieved.
• Numerical examples in signal recovery, CT imaging, and cryptoasset trading validated the approach.

Conclusions:

• The L2O methodology provides a practical framework for building explainable AI systems.
• Interpretable certificates enhance the trustworthiness and verifiability of AI inferences.
• This approach offers a promising direction for XAI in diverse scientific and financial domains.