Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Controllable Concept Bottleneck Models

Hongbin Lin, Chenyang Ren, Juangui Xu, Zhengyu Hu, Cheng-Long Wang, Yao Shu, Hui Xiong, Jingfeng Zhang, Di Wang, Lijie Hu

TL;DR

This work introduces Controllable Concept Bottleneck Models (CCBMs), a framework enabling post-deployment edits to CBMs at three granularities—concept-label-level, concept-level, and data-level—without full retraining. By deriving closed-form approximations from influence functions and accelerating Hessian computations with EK-FAC, CCBMs enable efficient, verifiable updates for correcting mislabeled or sensitive concepts, evolving concept sets, and incorporating new data. Empirical results across medical and vision benchmarks show dramatic runtime improvements (often >100x) with minimal utility loss, while interpretability is preserved through influence-based concept rankings and visualizations. The approach also demonstrates privacy benefits via data unlearning and supports continuous, sustainable AI in dynamic environments, highlighting practical significance for trustworthy, editable AI systems.

Abstract

Concept Bottleneck Models (CBMs) have garnered much attention for their ability to elucidate the prediction process through a human-understandable concept layer. However, most previous studies focused on static scenarios where the data and concepts are assumed to be fixed and clean. In real-world applications, deployed models require continuous maintenance: we often need to remove erroneous or sensitive data (unlearning), correct mislabeled concepts, or incorporate newly acquired samples (incremental learning) to adapt to evolving environments. Thus, deriving efficient editable CBMs without retraining from scratch remains a significant challenge, particularly in large-scale applications. To address these challenges, we propose Controllable Concept Bottleneck Models (CCBMs). Specifically, CCBMs support three granularities of model editing: concept-label-level, concept-level, and data-level, the latter of which encompasses both data removal and data addition. CCBMs enjoy mathematically rigorous closed-form approximations derived from influence functions that obviate the need for retraining. Experimental results demonstrate the efficiency and adaptability of our CCBMs, affirming their practical value in enabling dynamic and trustworthy CBMs.

Controllable Concept Bottleneck Models

TL;DR

This work introduces Controllable Concept Bottleneck Models (CCBMs), a framework enabling post-deployment edits to CBMs at three granularities—concept-label-level, concept-level, and data-level—without full retraining. By deriving closed-form approximations from influence functions and accelerating Hessian computations with EK-FAC, CCBMs enable efficient, verifiable updates for correcting mislabeled or sensitive concepts, evolving concept sets, and incorporating new data. Empirical results across medical and vision benchmarks show dramatic runtime improvements (often >100x) with minimal utility loss, while interpretability is preserved through influence-based concept rankings and visualizations. The approach also demonstrates privacy benefits via data unlearning and supports continuous, sustainable AI in dynamic environments, highlighting practical significance for trustworthy, editable AI systems.

Abstract

Concept Bottleneck Models (CBMs) have garnered much attention for their ability to elucidate the prediction process through a human-understandable concept layer. However, most previous studies focused on static scenarios where the data and concepts are assumed to be fixed and clean. In real-world applications, deployed models require continuous maintenance: we often need to remove erroneous or sensitive data (unlearning), correct mislabeled concepts, or incorporate newly acquired samples (incremental learning) to adapt to evolving environments. Thus, deriving efficient editable CBMs without retraining from scratch remains a significant challenge, particularly in large-scale applications. To address these challenges, we propose Controllable Concept Bottleneck Models (CCBMs). Specifically, CCBMs support three granularities of model editing: concept-label-level, concept-level, and data-level, the latter of which encompasses both data removal and data addition. CCBMs enjoy mathematically rigorous closed-form approximations derived from influence functions that obviate the need for retraining. Experimental results demonstrate the efficiency and adaptability of our CCBMs, affirming their practical value in enabling dynamic and trustworthy CBMs.
Paper Structure (45 sections, 27 theorems, 275 equations, 26 figures, 1 table)

This paper contains 45 sections, 27 theorems, 275 equations, 26 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Controllable Concept Bottleneck Models
  5. Concept Label-level Controllable CBM
  6. Concept-level Controllable CBM
  7. Data-level Controllable CBM
  8. Data Unlearning (Removal)
  9. Data Addition (Incremental Adaptation)
  10. Acceleration via EK-FAC.
  11. Experiments
  12. Experimental Settings
  13. Evaluation of Utility and Editing Efficiency
  14. Results on Interpretability
  15. Improvement via Harmful Data Removal
  16. ...and 30 more sections

Key Result

Theorem 4.2

The retrained concept predictor $\hat{g}_{e}$ defined by (concept-label:g) can be approximated by $\bar{g}_{e}$, defined by: where $H_{\hat{g}} = \nabla_{\hat{g}} \sum_{i,j} G^j_C(x_i,{c}_i;\hat{g})$ is the Hessian matrix of the loss function with respect to $\hat{g}$.

Figures (26)

  • Figure 1: An illustration of Controllable Concept Bottleneck Models with four settings.
  • Figure 2: The workflows of Controllable Concept Bottleneck Models with four settings.
  • Figure 3: Impact of edition ratio on three settings on CUB dataset.
  • Figure 4: F1 score difference after removing most and least influential concepts given by CCBM.
  • Figure 5: RMIA scores of data before and after removal.
  • ...and 21 more figures

Theorems & Definitions (49)

  • Definition 4.1
  • Theorem 4.2
  • Theorem 4.3
  • Theorem 4.4
  • Lemma 4.5
  • Theorem 4.6
  • Theorem 4.7
  • Theorem 4.8
  • Theorem 4.9
  • Theorem 4.10
  • ...and 39 more