SurroCBM: Concept Bottleneck Surrogate Models for Generative Post-hoc Explanation

Automated concept discovery reduces cost barriers: Traditional explanation methods require expensive human annotation to define concepts. SurroCBM automatically discovers concepts without manual labeling, making explainability more practical and scalable for real-world applications.

Surrogate models can explain complex AI systems: The framework uses a simpler, explainable surrogate model to interpret black-box AI decisions post-hoc, offering a practical approach to understanding opaque systems without requiring access to their internal workings.

Shared concepts reveal model similarities: By identifying concepts that are shared across different black-box models, SurroCBM can help organizations understand where different AI systems make similar decisions and where they diverge—useful for auditing and comparing model behavior.

Self-generated training data improves explanations: The paper proposes a training strategy using self-generated data that continuously enhances explanation quality, suggesting that explanation systems can improve iteratively without requiring additional human input.

Concept-based explanations offer semantic understanding: Unlike traditional saliency methods that only highlight important image regions, SurroCBM provides semantic meaning through discovered concepts, helping stakeholders understand why models make decisions in more interpretable terms.