Prototype-Grounded Concept Models for Verifiable Concept Alignment

Abstract: Concept Bottleneck Models (CBMs) aim to improve interpretability in Deep Learning by structuring predictions through human-understandable concepts, but they provide no way to verify whether learned concepts align with the human's intended meaning, hurting interpretability. We introduce Prototype-Grounded Concept Models (PGCMs), which ground concepts in learned visual prototypes: image parts that serve as explicit evidence for the concepts. This grounding enables direct inspection of concept semantics and supports targeted human intervention at the prototype level to correct misalignments. Empirically, PGCMs match the predictive performance of state-of-the-art CBMs while substantially improving transparency, interpretability, and intervenability.

• The paper presents PGCMs that integrate symbolic labels with visual prototypes to enable verifiable concept alignment.
• It introduces a dual-decoding framework where each prototype maps to both visual and concept representations via a similarity-based selector.
• Experimental results demonstrate that prototype-level interventions can boost concept accuracy by over 4% in practical settings.

Prototype-Grounded Concept Models for Verifiable Concept Alignment

Introduction

Prototype-Grounded Concept Models (PGCMs) address a central limitation in Concept Bottleneck Models (CBMs)—the inability to verify alignment between learned concepts and their intended human semantics. While CBMs structure prediction through human-understandable intermediates, their latent concept spaces are not directly inspected, making interpretability contingent on the opaque assumption of correct alignment. PGCMs resolve this by introducing explicit visual grounding: concepts are linked to learned visual prototypes, enabling direct inspection, verification, and targeted intervention at the prototype level, without sacrificing predictive performance relative to state-of-the-art CBMs.

Figure 1: Comparison of a standard neural network, a CBM, and a PGCM; only PGCMs allow direct visual verification of concept semantics via prototypes.

Model Framework

PGCMs maintain the CBM paradigm—predicting the target via intermediate concept representations—but crucially alter the inference process. Each concept is not a scalar latent but is instantiated as a dual representation: (1) a human-level symbolic label, and (2) a set of visual prototypes. These prototypes are derived by extracting and embedding spatial image parts (via segmentation), which are then matched to a finite pool of learnable prototype embeddings through a similarity-based selector.

For each input image $x$, segmentation isolates $n$ parts $\{x_i\}$, and, for each part, a selector produces a distribution over $m$ prototypes. Each prototype is endowed with: (i) an image decoder mapping the prototype embedding to the pixel space, and (ii) a concept decoder mapping the embedding to a vector of Bernoulli-distributed concept probabilities. The concept prediction for an image part is thus the expected concept vector under the selected prototypes, making each concept not only interpretable but visually inspectable.

Figure 2: Illustration of the inference pipeline: segmentation, prototype selection, and dual decoding of each prototype for visual and conceptual semantics.

Importantly, an interpretability optimization ensures prototype embeddings are not arbitrarily abstract. Instead, embeddings must faithfully reconstruct their decoded images and, after an intermediate “prototype swapping” step, are replaced with nearest actual image parts from the training set. This grounds prototypes not only in the learned embedding space but also in realistic image regions.

Interpretability, Semantics, and Alignment

Standard CBMs operate under an implicit semantics for each concept: the mapping from data to concept is opaque, and human users cannot audit whether a prediction (e.g., “has stripes”) actually refers to the intended visual property. By contrast, PGCMs define semantics explicitly via a concept alignment table, where every prototype is linked to both an image representation and a concept activation vector. Thus, for any concept, the explicit semantics is the disjunction over the set of prototype images where that concept is active:

$$C_{\text{red}} \iff \exists j\ \left[\,\text{prototype } j\ \text{image looks like } x_j \ \land\ C_{\text{red}}\ \text{active for}\ j\,\right]$$

Direct inspection of all such prototypes enables verifiable alignment: the user can visually check, for each concept and prototype, whether the attributed semantics is consistent with their own interpretation.

Intervenability and Editability

PGCMs introduce new modes of intervention unattainable in standard CBMs:

• Prototype editing: Misaligned prototypes can be relabeled in the concept alignment table to correct their semantic mapping.
• Prototype addition/removal: Visual artifacts or spurious prototypes (e.g., those induced by label noise) can be removed, and new prototypes (manually curated or sourced from the data) can be introduced for coverage, extending concept representation.
• Prototype selection intervention: For a particular input part, the user may override prototype selection, enforcing that a part be matched (or not matched) to specific prototypes.
• Interventions propagate: Because concepts are not modeled independently—multiple concepts can be affected by prototype-level operations—a single edit can simultaneously correct several concepts.

These forms of editability not only correct task predictions but can fundamentally alter the operational semantics of the model post-training, an essential property for human-in-the-loop model governance.

Experimental Results

PGCMs are evaluated against strong concept-based baselines (CBMs, Concept Residual Models, Concept-based Memory Reasoner) and direct DNNs on CelebA, ColorMNIST+, and CLEVR-Hans. Task and concept accuracy are comparable to CBMs, and any slight capacity penalty is attributable to architectural constraints, not the semantic grounding (see Table 1 in the original paper for exact results).

Manual edit interventions on noisy-label data—removing or relabeling misaligned prototypes—produced a relative increase in concept accuracy exceeding +4% on validated test cases, a corrective method unavailable to black-box or even standard CBMs.

PGCMs support interventions at the concept and prototype level. On ColorMNIST+, prototype-level interventions lead to a sharper rise in concept accuracy with fewer direct interventions than in CBMs, demonstrating the value of shared and aligned prototype structure.

Figure 3: Concept accuracy improvement as a function of interventions; prototype-based interventions yield higher gains with fewer actions.

Ablation over the number of prototypes on CLEVR-Hans reveals the standard trade-off in prototype-based models: increasing prototype count improves concept and task accuracy but at a cost to interpretability due to increased cognitive complexity.

Figure 4: Effect of prototype count on concept and task accuracy—greater capacity yields higher accuracy but at interpretability cost.

Qualitatively, visualizations on CelebA and CLEVR-Hans confirm that selected prototypes correspond to semantically meaningful and human-recognizable image parts, in contrast to the less interpretable latent features of standard CBMs.

Figure 5: PGCM model outputs demonstrating visually grounded concept evidence, in contrast to concept-only predictions in CBMs.

Figure 6: Assignment of prototypes to image parts in CLEVR-Hans; real image regions after prototype swapping ensure grounding.

Figure 7: Pre-swapping learned prototypes in CLEVR-Hans—already capturing visual and conceptual structure before final grounding in data.

Practical and Theoretical Implications

PGCMs offer a path towards interpretable and auditable neural models, closing the gap between post-hoc explanation and architectural transparency. Their explicit dual semantic representation makes them an ideal platform for human-AI collaboration, model auditing, and data-centric debugging. In application domains where safety and regulatory compliance require not only interpretability but also correct concept alignment (e.g., medical imaging, autonomous driving), PGCMs provide actionable evidence for both model developers and end users.

Theoretically, PGCMs instantiate a formal semantics for neural concept prediction in the Tarskian sense, operationalizing the mapping from abstract symbol (concept) to grounded instance (prototype). This enables empirical studies on the calibratability and editability of learned conceptual spaces, and opens trajectory for future integration with foundation models, formal verification systems, and neurosymbolic reasoning frameworks.

Conclusion

PGCMs reconcile the interpretability aspirations of CBMs with the requirement for explicit, verifiable, and editable concept grounding. By coupling symbolic concepts to visual prototypes, the model’s semantics become inspectable and actionable. Comprehensive intervention mechanisms further empower users to correct and refine semantic mappings, making PGCMs suitable for real-world deployment in high-stakes environments demanding transparency, trust, and human oversight.