CB-SLICE: Concept-Based Interpretable Error Slice Discovery

Researchers introduce CB-SLICE, a new method for identifying systematic errors in deep learning models by leveraging Concept Bottleneck Models to detect error patterns linked to human-understandable concepts. The approach outperforms existing techniques in uncovering model biases and provides more accurate, interpretable explanations of failure modes across multiple benchmarks.

CB-SLICE addresses a critical challenge in machine learning: deep learning models often fail systematically on specific population groups despite strong average performance. Traditional error slice discovery methods generate explanations disconnected from actual model inference processes, limiting their accuracy in root cause analysis. This research leverages Concept Bottleneck Models (CBMs), which make predictions based on human-interpretable semantic concepts, enabling direct traceability between concept failures and model errors.

The approach represents a meaningful evolution in model debugging and bias mitigation. By grouping samples with shared concept prediction failures, CB-SLICE identifies which specific concepts drive error patterns for different population groups. This interpretability advantage stems from CBMs' architecture, where failures directly map to concept mispredictions rather than opaque latent representations. The research demonstrates superior performance against existing state-of-the-art methods across multiple benchmarks.

For AI practitioners and organizations deploying models in production, this work has tangible implications. Interpretable error analysis enables faster debugging cycles and more targeted bias mitigation strategies. Rather than treating model failures as black boxes, teams can now pinpoint semantic-level failures and address them systematically. The concept-based explanations also enhance stakeholder trust by providing human-understandable failure narratives.

Looking forward, this research signals growing momentum toward interpretable-by-design machine learning architectures. As regulatory pressure increases for AI explainability and fairness, concept-based approaches may become standard practice. The methodology could extend beyond computer vision to other domains where semantic interpretability matters for compliance and deployment.

• →CB-SLICE groups error samples by shared concept prediction failures, enabling fine-grained bias identification tied directly to model inference
• →Concept Bottleneck Models provide inherent interpretability advantages over standard deep learning for error slice discovery
• →The method outperforms existing techniques while delivering richer, more faithful explanations of model failures across multiple benchmarks
• →Concept-based error analysis enables faster debugging and more targeted bias mitigation for production AI systems
• →This research reflects broader industry shift toward interpretable-by-design machine learning architectures for regulatory compliance