Prototype-Grounded Concept Models for Verifiable Concept Alignment

1. Problem Statement

Concept Bottleneck Models (CBMs) are a class of interpretable deep learning models that structure predictions through human-understandable intermediate concepts (e.g., "has stripes," "is metallic"). While CBMs allow for inspection and intervention at the concept level, they suffer from a fundamental limitation: lack of visual grounding.

• The Alignment Gap: In standard CBMs, concepts are learned as abstract scalar predictions. Even with human-supervised labels, there is no guarantee that the model's internal representation of a concept (e.g., "grey hair") aligns with the human's intended meaning. The model might rely on spurious correlations (e.g., background textures) rather than the actual visual feature.
• The Verification Deficit: Because the low-level visual evidence underlying a concept prediction is hidden, users cannot verify if the learned concept semantics match the intended semantics. This forces users to trust the model's "black-box" interpretation of the concept, undermining the core promise of interpretability.

2. Methodology: Prototype-Grounded Concept Models (PGCMs)

The authors propose PGCMs, a hybrid architecture that augments CBMs by grounding concepts in learned visual prototypes. Instead of representing a concept solely as an abstract prediction, PGCMs associate each concept with a set of localized visual patterns (prototypes) that serve as explicit evidence.

Core Architecture

The model operates as a three-stage mapping: $x \rightarrow p \rightarrow c \rightarrow y$ (Input $\rightarrow$ Prototypes $\rightarrow$ Concepts $\rightarrow$ Task).

1. Segmentation: The input image $x$ is decomposed into $n$ image parts $\{x_i\}$ using a segmentation network.
2. Prototype Selection (Discriminative): For each image part $x_i$, a prototype selector $q(S_i|x_i)$ computes a probability distribution over a set of $m$ learned prototypes. It selects the prototype most visually similar to the image part.
3. Dual Representation (Generative): Each prototype $j$ has two learned representations:
   • Image Representation: Generated by an image decoder $p(x_i|S_i=j)$, producing a concrete visual exemplar (modeled as a Gaussian).
   • Concept Representation: Generated by a concept decoder $p(c_i|S_i=j)$, producing a probability distribution over high-level concepts.
4. Prediction: Concept predictions are derived from the selected prototypes. The final task prediction $y$ is computed from the aggregated concept predictions.

Training Objective

The model is trained to maximize the joint log-likelihood of observed image parts, concepts, and task labels, utilizing an Evidence Lower Bound (ELBO) with four components:

• Task Loss: Standard classification loss for the final task.
• Concept Loss: Ensures the selected prototype correctly predicts the supervised concept labels.
• Reconstruction Loss: Encourages the prototype selector to pick prototypes visually similar to the input part and ensures prototypes are representative of the data.
• Regularization: Encourages high-entropy prototype selection to avoid local minima.

Key Optimization: To prevent prototypes from encoding "hidden" information, the authors enforce a constraint where the prototype embedding is derived only from its image representation (via a cycle: embedding $\to$ image $\to$ embedding). Additionally, a prototype swapping step replaces prototypes with the closest real image parts from the training set to ensure visual realism.

3. Key Contributions

A. Verifiable Concept Alignment

PGCMs introduce a Concept Alignment Table that explicitly maps prototype indices to their dual representations (visual image and concept vector).

• Mechanism: The semantics of a concept $C$ are defined as the disjunction of the image representations of all prototypes where $C$ is active.
• Benefit: Humans can directly inspect the "image representation" of a prototype to verify if it truly depicts the concept (e.g., does the prototype for "red" actually show red objects, or just red backgrounds?). This replaces the assumption of alignment with a verifiable mechanism.

B. Enhanced Intervenability

PGCMs unlock new forms of human intervention beyond standard concept editing:

1. Alignment Interventions (Model Editing): Users can edit the Concept Alignment Table directly. If a prototype is misaligned, users can:
   • Relabel the concepts associated with a prototype.
   • Remove spurious prototypes.
   • Add new prototypes (e.g., manually crafted image parts).
2. Prototype-Level Intervention: Users can intervene on the prototype selection process (e.g., forcing an image part to match a specific prototype) rather than just toggling a concept label.
3. Propagation: Because concepts are linked through shared prototypes, a single intervention on a prototype can simultaneously correct multiple concept predictions, unlike standard CBMs where concepts are often treated as conditionally independent.

C. Theoretical Grounding

The paper formalizes the semantics of PGCMs using a logic-based approach (Tarski semantics), where concepts are symbols and prototypes are the concrete entities they refer to, making the model's "meaning" auditable.

4. Experimental Results

The authors evaluated PGCMs on CelebA (face attributes), ColorMNIST+ (digit sums), and CLEVR-Hans (object properties).

• Predictive Performance: PGCMs achieve accuracy comparable to state-of-the-art CBMs (CBM, CRM, CMR) and Deep Neural Networks (DNNs).
  • Note: There is a slight drop in task accuracy on CelebA compared to DNNs, attributed to the capacity constraints of using a finite set of prototypes, but performance remains competitive with other concept-based models.
• Interpretability & Alignment:
  • Users could successfully verify concept alignment by inspecting the Concept Alignment Table.
  • Model Editing Experiments: On a noisy dataset (ColorMNIST+), removing or correcting misaligned prototypes significantly improved concept accuracy (from ~92.9% to ~96.9% and ~97.8% respectively), demonstrating the efficacy of prototype-level intervention.
• Intervenability:
  • PGCMs showed equal or superior responsiveness to concept interventions compared to standard CBMs.
  • The "PGCM*" variant (using prototype-level interventions) demonstrated that a single intervention could correct multiple concepts, outperforming independent concept models on ColorMNIST+.
• Capacity Trade-off: Accuracy scales with the number of prototypes. While more prototypes improve accuracy, they increase cognitive load, highlighting a trade-off between interpretability and model capacity.

5. Significance

This work addresses the "trust gap" in Explainable AI (XAI). While previous interpretable models (like CBMs) promised transparency, they relied on an unverified assumption that the model learned the correct semantics.

PGCMs fundamentally shift the paradigm by:

1. Making Semantics Explicit: Moving from abstract concept vectors to concrete visual evidence.
2. Enabling Verification: Allowing humans to audit what the model thinks a concept is, not just that it thinks a concept is present.
3. Facilitating Robust Correction: Providing a mechanism to fix model errors at the root cause (the visual prototype) rather than just the symptom (the concept label).

This framework bridges the gap between high-level symbolic reasoning and low-level visual perception, offering a more robust path toward trustworthy and controllable AI systems.