A redescription mining framework for post-hoc explaining and relating deep learning models

Deep learning models lack transparency despite their widespread use: The paper addresses a critical challenge in responsible AI—that while deep learning models achieve high performance across many domains, their enormous size makes them inherently difficult to explain, which can hinder trust and accountability in high-stakes applications.

A new tool can reveal what neural networks are actually doing: The proposed redescription mining framework identifies statistically significant patterns in how individual neurons activate, essentially translating the model's internal "decisions" into human-readable rules that connect neuron behavior to specific predictions or data attributes.

The framework works across different types of models and problems: Unlike some explainability methods tied to specific architectures, this approach is architecture-independent and can handle complex scenarios like multi-label predictions, making it potentially applicable to a broader range of AI systems used in practice.

Multiple ways to understand the same model: The framework supports both "pedagogical" explanations (understanding overall model behavior) and "decompositional" approaches (breaking down individual components), offering different perspectives on how a model works—similar to how different explanations can illuminate different aspects of a complex system.

Complements rather than replaces existing explainability methods: The authors note their approach provides "different information compared to existing explainable-AI approaches," suggesting it could be used alongside other explanation techniques to build more comprehensive understanding of model behavior for auditing and validation purposes.