CLIP-QDA: An Explainable Concept Bottleneck Model

Explainability is built into the model's design: Unlike "black box" AI systems that require separate explanation tools, CLIP-QDA integrates interpretability directly into its architecture. Each neuron in the model is linked to a specific word, making it possible to understand why the system makes classification decisions without additional analysis.

The model balances accuracy with transparency: CLIP-QDA achieves comparable accuracy to state-of-the-art concept bottleneck models while providing explanations that are faster to compute than existing explainability methods, suggesting that interpretability doesn't necessarily require sacrificing performance.

Two levels of explanations support different needs: The system provides both "local explanations" (why a specific image was classified a certain way) and "global explanations" (how the model works overall), offering flexibility for different stakeholders—from end-users to auditors—to understand model behavior.

Statistical simplicity enables trustworthiness: By using a Mixture of Gaussians framework, the model relies on simple, well-understood statistical distributions rather than complex neural network operations, making the underlying logic more transparent and potentially easier to verify for bias or errors.

A middle ground between performance and interpretability: CLIP-QDA represents a "greybox" approach that combines the strong performance of large foundation models with the transparency of simpler models, potentially offering a practical path for deploying AI systems where both accuracy and explainability matter.