Less is More: Discovering Concise Network Explanations

Simpler explanations can be just as effective: DCNE reduces the number of visual explanations needed to understand AI decisions by 97% (from CRP's output to 1/30) while maintaining nearly the same quality, demonstrating that "less is more" when explaining AI systems to humans.

AI explanations need to be human-friendly, not just accurate: The paper identifies a critical gap where existing explainability methods like CRP generate too many concepts for people to actually understand, highlighting that technical accuracy alone doesn't ensure responsible AI if humans can't comprehend the explanations.

Diverse explanations matter for trustworthiness: DCNE optimizes for three criteria simultaneously—few explanations, diverse perspectives, and human interpretability—suggesting that responsible AI requires balancing multiple factors rather than optimizing for just one metric.

Expert validation is essential for XAI methods: The researchers created a bird classification dataset with human expert-defined explanations to evaluate their method, emphasizing that AI explainability tools should be validated against human judgment to ensure they're actually meaningful to real users.

Interpretability is a practical tool for alignment: By making neural network decisions more accessible and transparent, DCNE supports the broader goal of "model alignment"—ensuring AI systems behave in ways humans can understand and trust.