If Concept Bottlenecks are the Question, are Foundation Models the Answer?

TL;DR

This work scrutinizes Concept Bottleneck Models (CBMs) and the trend of using Vision-Language Model (VLM) supervision to replace expert concept annotations. It introduces four concept-quality metrics—$\\mathsf{AUC}(C)$, LEAK, DIS, and OIS—to quantify accuracy, leakage, disentanglement, and unwanted correlations, and applies them to three datasets (Shapes3D, CelebA, CUB) across several architectures. The empirical findings show that VLM supervision can produce correct labels even when concept quality is poor, with label accuracy not reliably tracking concept quality, and that leakage and impurity can be substantial depending on dataset and VLM used. The paper concludes that while VLM-CBMs broaden applicability, they do not yet match expert annotations in preserving interpretable concept representations, and it outlines strategies for making VLM-CBMs more robust and trustworthy in practice.

Abstract

Concept Bottleneck Models (CBMs) are neural networks designed to conjoin high performance with ante-hoc interpretability. CBMs work by first mapping inputs (e.g., images) to high-level concepts (e.g., visible objects and their properties) and then use these to solve a downstream task (e.g., tagging or scoring an image) in an interpretable manner. Their performance and interpretability, however, hinge on the quality of the concepts they learn. The go-to strategy for ensuring good quality concepts is to leverage expert annotations, which are expensive to collect and seldom available in applications. Researchers have recently addressed this issue by introducing "VLM-CBM" architectures that replace manual annotations with weak supervision from foundation models. It is however unclear what is the impact of doing so on the quality of the learned concepts. To answer this question, we put state-of-the-art VLM-CBMs to the test, analyzing their learned concepts empirically using a selection of significant metrics. Our results show that, depending on the task, VLM supervision can sensibly differ from expert annotations, and that concept accuracy and quality are not strongly correlated. Our code is available at https://github.com/debryu/CQA.

Figures (8)

• Figure 1: A prototypical VLM-CBM. Left: given textual descriptions $\mathcal{T}$ of visual concepts, a VLM is used to annotate a training set $\mathcal{D}$ with per-concept labels or activation scores and (optionally) masks indicating where each concept activates. Right: the supervision is typically used to fine-tune a backbone $f$ that extracts concepts from new inputs and a linear layer $g$ for inferring predictions from the concepts. See \ref{['sec:prelims']} for architecture-specific differences.
• Figure 2: Issues with CBM concepts. An illustration of the four issues with concept quality and usage affecting CBMs on the CelebA task (cf. \ref{['sec:experiments']}): (1) Attaining high label accuracy does not prevent learning inaccurate concepts. (2) CBMs can maximize label performance by learning leaky concepts that include irrelevant contextual cues. (3) Entangled concepts encode unwanted information about one another, affecting out-of-distribution behavor. (4) Impure concepts encode unwanted correlations that do not exist among ground-truth concepts.
• Figure 3: Gaps of the concept leakage test in CelebA and CUB. The higher the gap, the more the subset of learned concepts leaks information for label prediction. (Left) Gaps on CelebA upon varying the subset of concepts from $1$ to $39$. All model concepts always display a non-negative gain in label $F_1$ compared to ground-truth concepts. (Right) Gaps on CUB upon varying the subset of concepts from $1$ to $112$. Only CBM is affected by leakage when considering smaller subsets of concepts.
• Figure 4: Gaps in concept leakage test on Shapes3d dataset. Negative values indicate that the classifier using predicted concepts has lower $F_1$ score than the same which has access to ground-truth concept annotations.
• Figure 5: Annotations obtained using G-DINO on the 3 different datasets. On CelebA (left) and Shapes3d (center) the VLM makes several mistakes highlighted by red crosses. On the other hand, the model does a good job annotating CUB (right).