Defining and Extracting generalizable interaction primitives from DNNs

DNNs can be simplified into understandable patterns: Deep neural networks make decisions through a small set of identifiable "interaction primitives"—basic patterns showing how input variables work together—rather than through incomprehensible black-box processes, making their reasoning more transparent.

Shared patterns reveal genuine knowledge: By comparing multiple neural networks trained on the same task, researchers can identify which interaction patterns are consistently used across different models. This approach helps distinguish between genuine patterns the networks have learned versus quirks specific to individual models, improving the reliability of AI explanations.

Generalization matters for trustworthiness: Previous methods could extract interaction patterns from individual networks, but these patterns often didn't generalize well or represent truly fundamental knowledge. This work addresses that limitation by focusing on interactions that multiple independent networks converge on, making explanations more faithful and trustworthy.

Practical implications for AI accountability: By extracting and understanding these shared interaction primitives, organizations can better audit what their AI systems have actually learned, potentially identifying biases or problematic patterns before deployment—an important step toward more responsible AI development.