Uncertainty-aware Language Guidance for Concept Bottleneck Models

This paper proposes an uncertainty-aware Concept Bottleneck Model that rigorously quantifies and incorporates the reliability of LLM-annotated concepts to mitigate hallucination risks and improve interpretability without requiring extensive expert labeling.

The Gist

Imagine you are trying to teach a robot how to identify different types of birds.

The Problem: The "Black Box" and the "Lazy Intern"

1. The Black Box:
Deep learning models (the robots) are incredibly smart at recognizing things, but they are like black boxes. You put a picture in, and they spit out an answer ("That's a goldfinch!"), but they can't tell you why. In high-stakes situations (like medical diagnosis or self-driving cars), we can't just trust a black box; we need to know the reasoning.

2. The Concept Bottleneck (The Good Solution):
To fix this, scientists created Concept Bottleneck Models (CBMs). Think of this as forcing the robot to act like a detective. Instead of guessing the bird directly, the robot must first list the clues it sees: "It has a yellow beak," "It has a red chest," "It has a small size." Only after listing these clues does it guess the bird. This makes the process transparent.

3. The Bottleneck (The Bad Problem):
Here's the catch: To train the robot to spot these clues, humans usually have to label thousands of pictures with these specific features. This is like hiring a team of expert ornithologists to sit and label every single photo. It's expensive, slow, and hard to scale.

4. The "Lazy Intern" (LLMs):
To speed things up, researchers started using Large Language Models (LLMs)—like the AI you are talking to right now—to act as the "intern" that labels the clues. The intern reads the picture and says, "I see a yellow beak!"

• The Risk: LLMs are great, but they sometimes hallucinate. They might confidently say, "I see a blue wing," when the bird is actually brown. If the robot learns from these fake clues, it becomes unreliable.
• The Second Risk: Even when the intern is unsure, current methods treat the answer as 100% fact. They don't ask, "How sure are you?" They just take the answer and move on, ignoring the fact that some clues are shaky.

---

The Solution: ULCBM (The "Strict Manager" with a "Safety Net")

The authors of this paper propose a new system called ULCBM (Uncertainty-aware Language Guidance for Concept Bottleneck Models). They solve the problems with two main tricks:

Trick 1: The "Strict Manager" (Uncertainty Quantification)

Instead of blindly trusting the intern (the LLM), this system acts like a strict manager who uses a special rulebook called Conformal Prediction.

• The Analogy: Imagine you are hiring a security guard to spot intruders. You don't just ask, "Did you see anyone?" You ask, "How sure are you?"
• The Process: The system tests the LLM on a small set of known images first. It sets a "confidence threshold."
  • If the LLM says, "I'm 99% sure this is a yellow beak," the manager accepts it.
  • If the LLM says, "I think it might be a blue wing, but I'm only 40% sure," the manager rejects it because it's too risky.
• The Magic: The paper proves mathematically that this method guarantees the system won't accept too many bad clues. It's like having a safety net that catches the hallucinations before they poison the robot's brain. They check three things:
  1. Discriminability: Is this clue actually unique to this bird? (Rejecting "has a beak" because all birds have beaks).
  2. Coverage: Did we miss any important parts? (Making sure we didn't forget the wings).
  3. Diversity: Are we repeating ourselves? (Rejecting "dark feathers" and "black plumage" as two separate clues when they mean the same thing).

Trick 2: The "Patchwork Quilt" (Targeted Data Augmentation)

Even with the strict manager, some clues are still rare. Maybe the LLM is very unsure about "yellow eyes" because it only saw them in 2 out of 1,000 pictures. If the robot tries to learn from such a tiny sample, it will fail.

• The Analogy: Imagine you are teaching a child to recognize a rare flower. You only have two photos of it. The child won't learn well. So, you take a piece of a real flower from a different photo and paste it onto a new background to create more examples.
• The Process: The system identifies these "rare and shaky" clues. It then takes a reliable patch of that feature from another image and inserts it into a new training image.
• The Safety: Crucially, the system is smart about where it pastes the patch. It looks at the "uncertainty map" and ensures it doesn't paste the new patch over an existing, reliable clue (like pasting a beak over an eye). It creates a "safety zone" for the new data.
• The Result: The robot gets plenty of practice on the rare, tricky clues, so it doesn't ignore them.

The Outcome

By using this "Strict Manager" to filter out bad guesses and the "Patchwork Quilt" to fix missing data, the new system:

1. Trusts less, verifies more: It filters out the LLM's hallucinations.
2. Learns better: It fills in the gaps where data is scarce.
3. Performs better: In tests, this method was more accurate and more reliable than previous methods, especially for the hardest-to-recognize categories.

In short: They built a system that uses AI to help train AI, but added a rigorous "quality control" layer to make sure the AI doesn't lie to itself, and a creative "data filling" layer to make sure it doesn't miss the important details.

Technical Summary

1. Problem Statement

Concept Bottleneck Models (CBMs) offer inherent interpretability by mapping inputs to human-understandable concepts before making a final prediction. However, their practical adoption is hindered by the high cost of manual concept annotation by human experts. While recent works utilize Large Language Models (LLMs) to automate concept generation, they face two critical limitations:

1. Unquantified Uncertainty & Hallucinations: Existing methods treat LLM-generated concepts as deterministic ground truth, ignoring the inherent uncertainty and potential for hallucinations (irrelevant or imprecise concepts). This lack of uncertainty quantification leads to unreliable concept bottlenecks.
2. Ineffective Training with Sparse Signals: Current approaches fail to incorporate the degree of uncertainty into the training process. When LLMs generate concepts with varying reliability, rigorous filtering often results in sparse training samples for specific concepts. Standard CBM training on such data tends to ignore these sparse but informative signals, degrading model performance.

2. Methodology: ULCBM

The authors propose ULCBM, a framework that rigorously quantifies LLM-annotated concept uncertainty with distribution-free guarantees and integrates this uncertainty into the training pipeline via targeted data augmentation.

A. Uncertainty-Aware Concept Generation (Language Guidance)

Instead of accepting LLM outputs directly, ULCBM employs a calibration process based on Conformal Prediction (CP) to construct concept sets with formal guarantees.

• Candidate Generation: An LLM generates candidate concepts for each class, and a grounded object detector (Grounding-DINO) identifies bounding boxes for these concepts in images.
• Three Complementary Criteria: The quality of a concept set is evaluated using three loss functions:
  1. Discriminability ($\ell_{dis}$): Ensures selected concepts are highly specific to the true class compared to competing classes.
  2. Coverage ($\ell_{cov}$): Ensures the selected set comprehensively covers the semantic scope of the class (no missing key features).
  3. Diversity ($\ell_{div}$): Penalizes semantic redundancy (e.g., selecting both "dark plumage" and "black feathers" if they mean the same thing).
• Conformal Calibration: To avoid distributional assumptions (i.i.d.), the method uses a calibration set to find a global threshold $\hat{\lambda}$. This threshold is selected such that the expected loss for each criterion remains below a user-specified risk level ($\alpha$) with distribution-free guarantees.
  • Theorem 1 proves that under exchangeability, the constructed uncertainty set satisfies the risk constraints for any new sample.

B. Targeted Data Augmentation

To address the sparsity of reliable concepts caused by strict filtering:

• Identification: The system identifies "sparse" concepts (those with few reliable occurrences).
• Synthesis: It synthesizes new training samples by inserting visual patches of the sparse concept from source images into target images.
• Uncertainty-Guided Placement: Crucially, the insertion location is constrained to avoid overlapping with existing high-reliability concepts (determined by the calibrated threshold $\hat{\lambda}$). This ensures the augmented data respects the spatial layout of the image while injecting the missing supervisory signal.

C. Training Procedure

The model is trained on the augmented dataset ($\hat{D}_{tr}^{aug}$) using a joint loss function:

• Concept Loss ($L_C$): Binary Cross Entropy between predicted and generated concept labels.
• Task Loss ($L_Y$): Cross Entropy for the final classification.
• Regularization: Elastic-net regularization is applied to the concept-to-class predictor to encourage sparsity and interpretability.

3. Key Contributions

1. Rigorous Uncertainty Quantification: The first method to provide distribution-free, theoretical guarantees for LLM-annotated concepts in CBMs using Conformal Prediction, effectively mitigating hallucination risks.
2. Multi-Dimensional Quality Control: Introduces a novel framework evaluating concepts across discriminability, coverage, and diversity, ensuring the concept set is relevant, comprehensive, and non-redundant.
3. Uncertainty-Aware Data Augmentation: Proposes a targeted augmentation pipeline that synthesizes training data for sparse, high-reliability concepts based on derived uncertainty metrics, preventing the model from ignoring valuable but rare signals.
4. Theoretical Analysis: Provides formal proofs (Theorem 1) ensuring that the calibrated thresholds satisfy prescribed risk levels for the defined loss functions.

4. Experimental Results

Experiments were conducted on CIFAR-10, CIFAR-100, and CUB (Caltech-UCSD Birds) datasets, comparing against baselines like LaBo and VLG-CBM.

• Validity (Risk Control): ULCBM successfully kept empirical losses for discriminability, coverage, and diversity below the specified risk thresholds (e.g., $\alpha_{dis}=0.7$). In contrast, baselines (LaBo, VLG-CBM) frequently exceeded these thresholds, indicating a failure to control uncertainty.
• Concept Compliance Accuracy (CCA): A new metric measuring the percentage of samples that are both correctly classified and have concept sets satisfying all quality constraints. ULCBM consistently achieved the highest CCA across all datasets and concept counts.
• Test Accuracy:
  • Overall Accuracy: ULCBM with data augmentation outperformed baselines (e.g., 75.5% on CUB vs. 74.4% for VLG-CBM).
  • Worst-Class Accuracy: The improvement was most significant here (25.0% for ULCBM vs. 16.7% for LaBo), demonstrating that the targeted augmentation effectively mitigates data scarcity for difficult/rare concepts.

5. Significance

This work bridges the gap between the interpretability of CBMs and the scalability of LLMs. By rigorously quantifying uncertainty and actively managing data scarcity through augmentation, ULCBM enables the deployment of reliable, interpretable AI models in high-stakes domains without relying on expensive human annotation. It establishes a new standard for trustworthy concept learning, ensuring that the "black box" of LLM generation is tamed by formal statistical guarantees before being fed into decision-making models.