Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Clarity: The Flexibility-Interpretability Trade-Off in Sparsity-aware Concept Bottleneck Models

Konstantinos P. Panousis, Diego Marcos

TL;DR

This work tackles the interpretability challenge of sparsity-aware concept bottleneck models (CBMs) in vision-language contexts by introducing Clarity, a metric that jointly accounts for sparsity, concept-prediction precision, and downstream accuracy. It systematically analyzes two CBM backbones (attribute-predictor-based and Vision-Language Model–based) under three per-example sparsity schemes: $\ell_1$, $\ell_0$ with Hard Concrete, and Bernoulli-based sparsity, using an amortized per-example selector $W_s$ to produce sparse concept activations. The framework enables objective interpretability evaluation on datasets with ground-truth attributes (CUB and SUN) and reveals a robust trade-off: methods with high flexibility can exhibit lower interpretability (clarity) unless carefully tuned, while higher sparsity often improves precision but may reduce task performance. The findings argue for evaluating interpretability with metrics beyond accuracy or sparsity alone and provide a principled path for designing CBMs that balance interpretability and performance in real-world applications. The work contributes a novel amortized sparsity mechanism, a formal clarity metric, and empirical evidence that sparsity-aware approaches require deliberate trade-off management to yield truly interpretable representations.

Abstract

The widespread adoption of Vision-Language Models (VLMs) across fields has amplified concerns about model interpretability. Distressingly, these models are often treated as black-boxes, with limited or non-existent investigation of their decision making process. Despite numerous post- and ante-hoc interepretability methods, systematic and objective evaluation of the learned representations remains limited, particularly for sparsity-aware methods that are increasingly considered to "induce interpretability". In this work, we focus on Concept Bottleneck Models and investigate how different modeling decisions affect the emerging representations. We introduce the notion of clarity, a measure, capturing the interplay between the downstream performance and the sparsity and precision of the concept representation, while proposing an interpretability assessment framework using datasets with ground truth concept annotations. We consider both VLM- and attribute predictor-based CBMs, and three different sparsity-inducing strategies: per example $\ell_1, \ell_0$ and Bernoulli-based formulations. Our experiments reveal a critical trade-off between flexibility and interpretability, under which a given method can exhibit markedly different behaviors even at comparable performance levels. The code will be made publicly available upon publication.

Clarity: The Flexibility-Interpretability Trade-Off in Sparsity-aware Concept Bottleneck Models

TL;DR

This work tackles the interpretability challenge of sparsity-aware concept bottleneck models (CBMs) in vision-language contexts by introducing Clarity, a metric that jointly accounts for sparsity, concept-prediction precision, and downstream accuracy. It systematically analyzes two CBM backbones (attribute-predictor-based and Vision-Language Model–based) under three per-example sparsity schemes: , with Hard Concrete, and Bernoulli-based sparsity, using an amortized per-example selector to produce sparse concept activations. The framework enables objective interpretability evaluation on datasets with ground-truth attributes (CUB and SUN) and reveals a robust trade-off: methods with high flexibility can exhibit lower interpretability (clarity) unless carefully tuned, while higher sparsity often improves precision but may reduce task performance. The findings argue for evaluating interpretability with metrics beyond accuracy or sparsity alone and provide a principled path for designing CBMs that balance interpretability and performance in real-world applications. The work contributes a novel amortized sparsity mechanism, a formal clarity metric, and empirical evidence that sparsity-aware approaches require deliberate trade-off management to yield truly interpretable representations.

Abstract

The widespread adoption of Vision-Language Models (VLMs) across fields has amplified concerns about model interpretability. Distressingly, these models are often treated as black-boxes, with limited or non-existent investigation of their decision making process. Despite numerous post- and ante-hoc interepretability methods, systematic and objective evaluation of the learned representations remains limited, particularly for sparsity-aware methods that are increasingly considered to "induce interpretability". In this work, we focus on Concept Bottleneck Models and investigate how different modeling decisions affect the emerging representations. We introduce the notion of clarity, a measure, capturing the interplay between the downstream performance and the sparsity and precision of the concept representation, while proposing an interpretability assessment framework using datasets with ground truth concept annotations. We consider both VLM- and attribute predictor-based CBMs, and three different sparsity-inducing strategies: per example and Bernoulli-based formulations. Our experiments reveal a critical trade-off between flexibility and interpretability, under which a given method can exhibit markedly different behaviors even at comparable performance levels. The code will be made publicly available upon publication.
Paper Structure (10 sections, 22 equations, 5 figures, 2 tables)

This paper contains 10 sections, 22 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background on sparse representations
  4. Proposed Framework
  5. Experimental Evaluation and Discussion
  6. Experimental Setup.
  7. Limitations & Conclusions
  8. Experimental Details.
  9. Concrete Relaxation
  10. Additional Visualizations

Figures (5)

  • Figure 1: Accuracy-Sparsity and Precision-Sparsity Curves for the CUB (top) and SUN (bottom) datasets. From left to right, we report results for predictor-based methods (first two plots) and VLM-based methods (last two plots). For both methods, all sparsity-inducing approaches maintain strong classification performance even at high sparsity levels. For the VLM-based setting, we observe a significant increase in classification accuracy when employing sparsity-aware methods. This suggests that, despite the limited zero-shot attribute prediction capabilities of CLIP in this scenario, appropriate concept selection can recover substantial discriminative power for the downstream task. At the same time, increased sparsity consistently leads to higher precision in the resulting representations.
  • Figure 2: Clarity-Accuracy Curves for both datasets. From left to right, we report the results for predictor-based methods (first two plots) and VLM-based methods (last two plots). While all methods can achieve similar classification accuracy, they do so at widely different clarity levels, highlighting that strong task performance does not guarantee interpretable representations.
  • Figure 3: Visualization of selected concepts for two examples from the CUB dataset by selecting the method with the best clarity (predictor-based Bernoulli) versus the one with the best performance (VLM-based $\ell_0$). Green highlighting denotes concepts that are present in the ground truth. Left:Brewer Blackbird example with 28 ground-truth active concepts. Right:Great Crested Flycatcher example with 30 ground-truth active concepts. Similar trends are observed in both cases, with the Bernoulli method selecting fewer, more precise concepts, and the $\ell_0$ method selecting a larger set with lower precision.
  • Figure 4: Visualization of selected concepts for two examples from the CUB dataset for the best performing models in terms of clarity (Predictor-based Bernoulli) and classification performance (VLM-based$\ell_0$) . Green highlighting denotes concepts that are active in the ground truth. Left:Cactus Wren example with 34 ground-truth active concepts; the Bernoulli-based method selects 12 concepts, 6 of which are correct, while the $\ell_0$ method selects 27 concepts, 2 of which are correct. Right:Belted Kingfisher example with 28 ground-truth active concepts; similar trends are observed, with the Bernoulli-based method selecting fewer, more precise concepts, and the $\ell_0$ method selecting a larger set with lower precision.
  • Figure 5: Visualization of selected concepts for two examples from the SUN dataset. Green highlighting denotes concepts that are active in the ground truth. Left:Chaparral example with 13 ground-truth active concepts; we choose a different set of models from the optimal in terms of clarity or accuracy, i.e., (Predictor-based$\ell_1$) and classification performance (VLM-based Bernoulli). The $\ell_1$-based method selects 20 concepts, 10 of which are correct, while the Bernoulli-based method selects 54 concepts, 11 of which are correct. Right:Natural History Museum example with 5 ground-truth active concepts; here, we choose the best performing models in terms of either clarity of accuracy as before, i.e., Predictor-based Bernoulli and VLM-based$\ell_0$. The Bernoulli-based method selects fewer, more precise concepts, and finds the 5 ground truth concepts, while the $\ell_0$ method selects a larger set with lower precision.