An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations

Official PyTorch Implementation of An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations (NeurIPS 2025)
Seonghwan Park, Jueun Mun, Donghyun Oh, and Namhoon Lee

---

Table of Contents

1. Abstract
2. Research Highlights
3. Setup
4. Data Preparation
   • CUB Dataset
   • AwA2 Dataset
5. Running the Experiments
   • Independent Model
   • Joint Model
   • Sequential Model
   • Test-Time Intervention
6. Contact
7. Citation
8. Acknowledgements

---

Abstract

Concept bottleneck models (CBMs) ensure interpretability by decomposing predictions into human interpretable concepts. Yet the annotations used for training CBMs that enable this transparency are often noisy, and the impact of such corruption is not well understood. In this study, we present the first systematic study of noise in CBMs and show that even moderate corruption simultaneously impairs prediction performance, interpretability, and the intervention effectiveness. Our analysis identifies a susceptible subset of concepts whose accuracy declines far more than the average gap between noisy and clean supervision and whose corruption accounts for most performance loss. To mitigate this vulnerability we propose a two-stage framework. During training, sharpness-aware minimization stabilizes the learning of noise-sensitive concepts. During inference, where clean labels are unavailable, we rank concepts by predictive entropy and correct only the most uncertain ones, using uncertainty as a proxy for susceptibility. Theoretical analysis and extensive ablations elucidate why sharpness-aware training confers robustness and why uncertainty reliably identifies susceptible concepts, providing a principled basis that preserves both interpretability and resilience in the presence of noise.

Research Highlights

Key Contributions

• Measuring Impact:
  We present the first comprehensive and systematic characterization of noise's impact on CBMs, demonstrating that even moderate level of noise significantly degrades predictive performance, interpretability, and intervention effectiveness.

• Understanding Mechanisms:
  To deepen our understanding of noise effects on \cbms, we identify a susceptible set, a small subset of concepts that suffer significant accuracy loss, and demonstrate that overall performance decline is largely driven by these localized failures.

• Mitigating Effects:
  To neutralize the risk of the susceptible set, we introduce two mitigation strategies: integrating sharpness minimization during training and implementing an uncertainty-based intervention at inference. Our suggestion is supported by theory and validated through extensive empirical evidence.

---

Setup

Clone the Repository

Clone the repository from GitHub and navigate into the directory.

git clone https://github.com/LOG-postech/CBM-Noise.git
cd CBM-Noise

Create and Activate Conda Environment

Create a new Conda environment named cbm-noise with Python version 3.9.12 and activate it.

conda create -y -n cbm-noise python=3.9.12
conda activate cbm-noise

Install PyTorch and TorchVision

Note: If you require a different CUDA version, please refer to the PyTorch official website.

Install specific versions of torch and torchvision using pip with the specified CUDA version.

pip install torch==1.13.0+cu117 torchvision==0.14.0+cu117 --extra-index-url https://download.pytorch.org/whl/cu117

Install Dependencies

Install all necessary dependencies listed in the requirements.txt file.

pip install -r requirements.txt

---

Data Preparation

CUB Dataset

Download the CUB dataset from http://www.vision.caltech.edu/visipedia/CUB-200-2011.html and place the downloaded data under the DATA/ directory.

AwA2 Dataset

Download the AwA2 dataset from https://cvml.ista.ac.at/AwA2/ and place it under the DATA/ directory.

---

Running the Experiments

Before running any experiment, set the FOLDER_DIR variable in each shell script to your working path.

Each shell script handles the full pipeline including training (X→C and C→Y where applicable) and inference. You can configure the following variables at the top of each script:

Variable	Options	Description
DATASET	cub, awa2	Target dataset
OPTIMIZER	SGD, SAM	Optimizer (SAM: Sharpness-Aware Minimization)
NOISE	0.0, 0.1, ...	Noise ratio (0.0 for clean)
NOISE_LOC	concept, class, both	Where to inject noise
BACKBONE	inception_v3, resnet18_cub, resnet101, vit_b16, vit_l16	Backbone architecture
RHO	0.1, ...	Perturbation radius for SAM

Independent Model

Trains the X→C model and C→Y model separately, then runs inference.

sh train_cbm_ind.sh

Joint Model

Trains the X→C→Y model end-to-end, then runs inference.

sh train_cbm_joi.sh

Sequential Model

Uses the X→C model from the Independent model, then trains C→Y on the predicted concept logits.

Note: The Sequential model requires a pre-trained Independent X→C model. Run train_cbm_ind.sh first.

sh train_cbm_seq.sh

Test-Time Intervention

Runs test-time intervention (TTI) on a trained Independent model. The intervention criterion can be configured via the CRITERION variable (e.g., ucp, rand, lcp, cctp, ectp, eudtp).

sh tti.sh

---

Contact

For any questions, discussions, or proposals, please contact:

📧 Email: shpark97@postech.ac.kr

---

Citation

If you use our code in your research, please consider citing:

@article{park2025analysis,
  title={An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations},
  author={Park, Seonghwan and Mun, Jueun and Oh, Donghyun and Lee, Namhoon},
  journal={arXiv preprint arXiv:2505.16705},
  year={2025}
}

---

Acknowledgements

Our experimental pipeline is built upon the following repository:

• CBM
• CBM-Intervention
• SAM

We express our gratitude to Koh et al., Shin et al., and Foret et al. for sharing their outstanding work and making their contributions available through open-source initiatives.