Towards Faithful Multimodal Concept Bottleneck Models

1. Problem Statement

Concept Bottleneck Models (CBMs) are interpretable AI architectures that route predictions through a layer of human-understandable concepts before making a final decision. While effective in vision and NLP, their application in multimodal settings (combining text and images) remains underexplored.

The core challenge addressed is faithfulness. For a CBM to be truly interpretable, its explanations must accurately reflect the model's reasoning. Current CBMs suffer from two main failures that compromise faithfulness:

1. Inaccurate Concept Detection: The model fails to correctly identify the presence of concepts in the input.
2. Leakage: The concept layer encodes unintended predictive information ("smuggling" signals) to bypass the bottleneck and improve accuracy. This occurs in two forms:
   • Task Leakage: Concepts encode task-relevant signals beyond their intended semantics.
   • Inter-Concept Leakage: Concepts encode information about other concepts beyond their natural correlations.

Existing methods typically treat concept detection and leakage mitigation as separate problems, often sacrificing predictive accuracy to reduce leakage or failing to detect concepts accurately.

2. Methodology: f-CBM

The authors propose f-CBM, a faithful multimodal CBM framework built on a CLIP-based vision-language backbone. It addresses the dual challenges of detection and leakage through two complementary strategies:

A. Differentiable Leakage Loss

To mitigate leakage, the authors introduce a differentiable leakage loss ($L_{leak}$) that can be incorporated directly into the training objective.

• Mechanism: Instead of using discrete binning (which destroys gradients), they use Kernel Density Estimation (KDE) to estimate Mutual Information (MI) between predicted concepts and task labels.
• Objective: The loss penalizes the excess mutual information between predicted concepts ($\hat{c}_i$) and the task label ($y$) compared to the ground truth concepts ($c_i$).
  $$L_{leak} = \left[ \frac{\hat{I}(\hat{c}_i; y) - \hat{I}(c_i; y)}{H(y)} \right]^2$$
• Effect: This forces the model to learn concept representations that contain only the information necessary to define the concept, removing "cheating" signals used for classification.

B. Kolmogorov-Arnold Network (KAN) Prediction Head

Standard CBMs use a linear layer to map concepts to classes. The authors replace this with a Kolmogorov-Arnold Network (KAN) layer.

• Rationale: A linear head may lack the expressiveness to model complex relationships, forcing the concept layer to encode extra information (leakage) to compensate.
• Implementation: KANs place learnable univariate functions on edges rather than fixed activation functions on nodes. This provides greater non-linear expressiveness while maintaining interpretability.
• Benefit: By allowing the final layer to handle complex mappings, the pressure on the concept bottleneck to encode extraneous task signals is reduced, thereby improving concept detection and reducing leakage.

C. Training Strategy

The model is trained jointly using a composite loss function:
$$L = L_{cls} + \tilde{\lambda}_C L_C + \tilde{\lambda}_{leak} \alpha L_{leak}$$

• $L_{cls}$: Cross-entropy classification loss.
• $L_C$: Concept prediction loss (MSE).
• $L_{leak}$: The differentiable leakage loss.
• Dynamic Scaling: Auxiliary losses are dynamically rescaled relative to the classification loss to ensure stable training.
• Curriculum Learning: The leakage loss weight ($\alpha$) follows a cosine annealing schedule from 0 to 1, allowing the model to first learn concepts before being penalized for leakage.

3. Key Contributions

1. First Faithful Multimodal CBM: Introduction of a CLIP-based architecture that processes both image and text inputs within a single concept bottleneck.
2. Preliminary Analysis: Empirical demonstration that concept detection accuracy, task leakage, and inter-concept leakage are positively correlated; improving detection and reducing task leakage simultaneously mitigates inter-concept leakage.
3. Differentiable Leakage Metric: A novel, fully differentiable measure of task leakage based on KDE, enabling end-to-end optimization to minimize information leakage.
4. KAN Integration: The first application of Kolmogorov-Arnold Networks in CBMs to enhance predictive flexibility without sacrificing interpretability, leading to better concept detection.

4. Experimental Results

The authors evaluated f-CBM on four datasets (N24News, CUB-200-2011, AGNews, DBpedia) using CLIP-base and CLIP-large backbones.

• Performance Trade-off: f-CBM achieves the Pareto-optimal trade-off between task accuracy, concept detection error (c-RMSE), and leakage reduction.
  • It matches the task accuracy of black-box models and other strong baselines (e.g., ~98% on N24News).
  • It significantly outperforms competitors (Indep.-CBM, Label-free, CT-CBM) in leakage reduction (lowest CTL and ICL scores).
  • It achieves superior concept detection (lowest c-RMSE) compared to leakage-mitigation baselines.
• Ablation Studies:
  • Replacing the linear head with KAN improved concept detection by 51% and reduced leakage by 30% with negligible accuracy loss.
  • Adding the leakage loss reduced task leakage by 73% with minimal impact on accuracy.
• Intervention Analysis (Crucial Validation):
  • In a "concept intervention" test (replacing predicted concepts with ground truth), f-CBM was the only method that consistently improved accuracy.
  • Competitors (like CT-CBM) saw accuracy degrade under intervention, proving they relied on leaked information rather than true concept semantics.

5. Significance

This work establishes a new standard for faithful multimodal AI.

• Practical Reliability: It proves that high accuracy and high interpretability are not mutually exclusive. By addressing leakage, f-CBM ensures that human interventions (correcting concept predictions) actually improve model behavior, a critical requirement for high-stakes domains like healthcare.
• Interpretability Depth: The use of KANs allows for the visualization of non-linear response curves, offering deeper insights into how specific concepts influence class predictions (e.g., saturation effects).
• Generalizability: The framework is versatile, working seamlessly across pure text, pure image, and multimodal datasets, suggesting a robust path forward for explainable multimodal AI.