VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance

Dear Reviewer i5YZ,

Thank you for your feedback. Below are responses to your comments.

Q1: Therefore, also comparing the performance of VLG-CBM under conventional settings in previous work would make the comparison easier.

A1: In this paper we do not compare the results under a dense final layer setting mainly due to this setting is less informative: as shown in Fig. 3 of our draft, even random baseline achieve similar performance to SOTA CBM models, when NEC becomes large enough.

---

Q2: The idea of applying a detection or a grounding module to generate an image annotation is already widely explored (e.g., KOSMOS-2 [1]). The proposed data generation process does not significantly differ from previous approaches, limiting the originality of the proposed method.

A2: We would like to clarify our contribution in regards to grounding concepts for training CBMs. CBMs rely on LLMs for a concept set which usually provide generic concepts and might not be possible to visualize in images. We are the first to use open-vocabulary grounding models to improve the effectiveness of CBMs.

---

Q3: How is the performance of each method? Only the percentage of changed decisions are reported, therefore providing accuracies may help a better understanding of the effectiveness of each method

A3: We calculate the performance of each method:

We could see here most methods suffer significant performance drops. LF-CBM has a smaller gap, partially due to its sparse final layer.

---

Q4: Isn't it obvious that the percentage of changed decisions is lower in VLG-CBM since the sparsity constraint is not applied to other methods? Can you also provide the same analysis with sparsity constraint applied to other methods?

A4: Yes, what we want to illustrate here is to emphasize the importance of NEC control. A quite popular method to explain model decisions for CBMs is to show the users the top-5 contributing concepts. However, we found this approach has potential risk without NEC control: as shown in Tab. 3, if limited to only top-5 concepts, the decision of the model will change significantly. This suggests that showing top-5 concepts does not faithfully explain the model behavior.

To better illustrate this, we choose a baseline, LF-CBM and apply our NEC control on it. The below table shows that controlling NEC effectively reduces the decision change.

---

Q5: Providing qualitative examples of labeling results using Grounding-DINO would help in understanding the quality of the auto-labeled dataset, which seems to be absent in the current version of the paper.

A5: We have provided few examples of annotation obtained from GroundingDINO in global response supplementary pdf.

---

Summary

In summary, we have

• In Q1, we discuss why conventional benchmark is not that informative.

• In Q2, we clarify our difference with previous work.

• In Q3, we provide the performance change of each method in Table. 3.

• In Q4, we discuss how NEC constraint affect the decision changes and add LF-CBM experiments to support it.

• In Q5, we provide a few examples from GroundingDINO.

In response to the weakness part, we have

• addressed the reviewer’s concern in weakness #1 in Q1.

• addressed the reviewer’s concern in weakness #2 by clarifying our difference with previous work in Q2.

We believe that we have addressed all your concerns. Please let us know if you still have any reservations and we would be happy to address them!

Thank you for your response in a limited time.

Although some of my concerns have been addressed (e.g., questions regarding Table 3 and request for qualitative results), I still feel that the contribution is somewhat limited, which is also pointed out by reviewer QueL.

For me, the novel point of this paper seems to be

1. Introducing the idea of utilizing Grounding-DINO to remove noisy concepts generated by LLMs,
2. Proposing the new evaluation metric, NEC, quantifying the previously defined problem of information leaks.

Since the problem of non-visual concepts being generated by LLMs [1] and information leaks in CBM [2, 3] have been discovered before, it seems that the novelty in the problem definition or its solution is somewhat limited for me.

Therefore, could the authors provide additional points that I am missing that make the problem definition or its solution's novelty clearer?

[1] Kim et al., Concept bottleneck with visual concept filtering for explainable medical image classification, MICCAI Workshop 2023

[2] Roth et al., Waffling around for Performance: Visual Classification with Random Words and Broad Concepts, ICCV 2023

[3] Yan et al., Learning Concise and Descriptive Attributes for Visual Recognition, ICCV 2023

Thanks for your response. I appreciate the authors for providing a detailed comparison with previous works, which makes the position of the paper more clear.

Since now I better understand the novelty of the paper, I am raising the score by one.

Still, I highly recommend the authors include a detailed comparison with previous works in the manuscript. Since the contributions of the paper are built upon previous observations, clearly stating the novel points and importance of the Grounding-DINO approach and NEC metric may help readers better grasp the value of the paper.

Dear Reviewer MvHs,

Thank you for the positive feedback. Please see below our responses to address your comments.

Q1: It is not clear if NEC controls information leakage. As defined in [2], information leakage happens because of a `soft’ concept layer (that predicts probability of a concept instead of presence and absence). Though a small NEC reduces the possibility of information leakage and is a useful metric to check the goodness of the concept set.

A1: Information leakage is a problem that haunts CBM models. As discussed in Sec 4 of [2], not only soft CBL has this problem, using hard concepts still shows information leakage, suggesting it’s hard to fully eliminate it in CBM. Though the reason for this has not been fully understood, we present Thm. 4.1 to provide a theoretical understanding of this problem. Additionally, in order to understand the real semantic information learned by CBL, we compare model performance with a random baseline as a “control group”, where no semantic information is learned. Our results in Figure 3 show that this random baseline drops significantly when NEC is low. This suggests that reducing NEC is a way to control information leakage.

---

Q2: A study of time and compute complexity is needed, as every image has to be passed through a large object-detector. The effort to generate concepts is much higher (compared to [1], who generate concepts class wise).

A2: Following your suggestion, we provide a comparative analysis of GPU-hours required to obtain annotations for three datasets: CUB, ImageNet, and Places365 in the table below. We would also like to clarify that we use concepts relevant to the class of an image when using Grounding-DINO. This helps reduce the false-positive rate of the open-vocabulary detection model and significantly decrease computation time. Please refer to Appendix D for more details.

---

Q3: A study of the impact of proposed visually-grounded concept filtering would be interesting. For ex. What fraction of quantitative gains are because of the removed concept set? How much does augmentation contribute to overall performance gains?

A3: We would like to clarify that concept filtering is a natural step in our training pipeline. An ablation here may not change the performance as it would be equivalent to training in a multi-label setting with certain labels never appearing in the training dataset.

Further, following your request, we reported our provided an ablation study on data augmentation in the Global Response #2.1.

---

Q4: The theoretical analysis, though appreciated, seems out of place. The result presented in Thm 4.1 is not specific to concept models, but can be seen as a general statement on fidelity of performing dimensionality reduction with a linear random projection. It is not clear to me how NEC follows from here, as NEC is a measure of the sparsity of the final layer, which can be low even for large concept sets.

A4: Thank you for the comments, please allow us to clarify. Thm. 4.1 shows that the approximation error goes down linearly with the number of concepts, when CBL weights are randomly selected. This supports previous observations that a random CBL layer could also achieve good performance, suggesting the existence of information leakage. By applying NEC constraint, we actually constraint the number of concepts in Thm 4.1 (as other concepts do not contribute to the decision and could be ignored). Thus, the information leakage could be controlled as Thm. 4.1 suggested.

---

Reference

[1] Oikarinen et al. “Label-Free Concept Bottleneck Models.” ICLR 2023

[2] Mahinpei et al. “Promises and Pitfalls of Black-Box Concept Learning Models.” 2021

---

Summary

In summary,

• In Q1, we discuss why controlling NEC could help control information leakage

• In Q2, we provide a study on computational time and complexity.

• In Q3, we provide an ablation study on the contribution of our data augmentation.

• In Q4, we discuss how our theoretical result is connected to the NEC.

In response to the weakness part, we have addressed the reviewer’s concern

• in weakness #1 in Q1.

• in weakness #2 by providing an analysis on computational time and complexity in Q2

We believe that we have addressed all your concerns. Please let us know if you still have any reservations and we would be happy to address them!

Dear Reviewer QueL,

Thank you for the feedback, we believe there are some misunderstandings based on the comments. Please allow us to clarify below to address your comments.

Q1: This paper is an incremental work to previous LLM-guided CBMs [1, 2] by adding a step to filter concepts with an object detection model.

A1: We would like to clarify our contributions below and demonstrate that our contribution is much more than just filtering the concepts with object detection models. As described in the end of Sec 1, we have 3 main contributions spanning from a new methodology in sec 3, rigorous theoretical analysis in sec 4, to extensive evaluation in sec 5:

1. We are the first to propose an end-to-end pipeline for training CBM with annotations from grounding DINO and provide a method to utilize bounding boxes information obtained from the Grounding DINO, as shown in Sec 3.
2. As demonstrated and discussed in Sec 4, our work is the first to provide rigorous theoretical analysis for information leakage in CBMs and demonstrate that CBM can achieve optimal accuracy even on random concepts when the number of concepts is large enough
3. As discussed in Sec 4, we propose a new metric called the Number of Effective Concepts (NEC), which facilitates fair comparison between different CBMs by controling information leakage and helps improve interpretability of the CBMs.

---

Q2: The findings that "CBM with random concepts can achieve good performance" have also been shown in previous works [3, 4]...

A2: CBM with random concepts can achieve good performance as has been observed before, as you point out. However, we are the first to provide a theoretical analysis for it. As we prove in Thm. 4.1, The random concept CBL can approximate any linear function with approximation error that goes down linearly with the number of concepts, and reaches zero error when the number of concepts gets to the dimension of embedding.

---

Q3: However, the detected bounding boxes are not used in the model training and inference. Therefore, the technical contribution is also limited…

A3: We would like to clarify that the detected bounding boxes are indeed used during model training. As mentioned in Ln 134-136, we augment the training dataset by cropping images to a randomly selected bounding box and modifying the target one-hot vector to predict the concept corresponding to the bounding box.

---

Q4: The proposed method relies on an open-vocab detection model. However, open-vocab detection is still a very challenging task, and from my experience, the off-the-shelf models do not work well in many cases. For example, …

A4: We would like to highlight that the open-vocabulary object detectors primarily suffer from false-positive detections, and we address this issue in the following way:

• Rather than using the entire concept set for detecting concepts in an image, we limit the concepts relevant to the ground truth class of the image. This ensures that the concepts have a high likelihood of being present in the image.

• We used confidence threshold to filter bounding boxes with low confidence. In the Appendix D, we have evaluated concept annotations at different thresholds. We use CUB dataset for comparison which contains ground-truth for fine-grained concepts present in each image and report precision and recall metric to measure the quality of annotations from Grounding-DINO. As demonstrated in the appendix along with Table 2, Fig 4, Fig G.1 and G.2, the effect of false-positives is minimal and VLG-CBM is able to faithfully represent concepts in the CBL.

---

Q5: This concern is also reflected by the small set of datasets in their experiments, only 5 datasets, compared to what LaBo evaluated (11 datasets).

A5: Results on Flower102 and Food101: The Food-101 dataset achieved an ACC@NEC=5 of 81.68% and an Avg. Acc of 80.31%, while the Flower-101 dataset achieved higher scores with an ACC@NEC=5 of 90.58% and an Avg. Acc of 92.94% with CLIP-Resnet50 backbone.

---

Q6: The linear probing baseline is missing, which is necessary to justify the effectiveness of CBMs.

A6: The linear probing results for Table 2 models are: 88.80% for CIFAR10, 70.10% for CIFAR100, 76.70% for CUB, 48.56% for Places and 76.13% for ImageNet. The gaps between our models (avg. acc) to the linear probing is 0.17% for CIFAR10, 3.62% for CIFAR100, 0.88% for CUB, 5.99% for Places and 2.15% for ImageNets.

---

Q7: For CUB, Place365, and ImageNet, the numbers of LM4CV and LaBo are missing.

A7: Following your request, we reported our results in the Global Response #1 Additional models and datasets. It can be seen that our VLG-CBM still outperforms the baselines under this setting.

---

Q8: In general, the numbers shown in Table 2 are quite low, far behind linear probing based on my experience. It seems the proposed method sacrifices a lot of performance for interpretability, which is not validated properly (see weakness 4)...

A8: First, we want to clarify that our accuracy is not far behind linear probing baseline, only 0-6% gap (refer to the numbers in A6) while other baselines have much larger gaps . In Table 2, we control NEC to be small (NEC=5) mainly for fairly compare different models: as shown in Figure 3, under dense setting (linear probing) even random CBL could achieve comparable performance to SOTA CBMs, which largely weaken its usefulness in comparing different CBMs. In practice, users have the flexibility to choose the NEC to trade-off between performance and interpretability.

---

Q9: …without concrete evidence like large-scale human evaluation, it is hard to justify VLG-CBM as more interpretable by just showing a few qualitative examples.

A9: Following your suggestion, we have conducted additional large-scale human evaluations and demonstrated that our method outperformed the baselines. Please see the results in Global Response #3 Human study.

Dear Reviewer UnrN,

Thank you for the positive feedback! Please see our responses below to address your comments.

---

Q1: How well the grounding detector can do in discovering the concept candidate. And if the detector fails to detect a key concept, will the final recognition result be affected?

A1: In Appendix D, we provided a study evaluating concept annotations obtained from Grounding-DINO. We use CUB dataset for comparison which contains ground-truth for fine-grained concepts present in each image and report precision and recall metric to measure the quality of annotations from Grounding-DINO.

The confidence threshold controls when a concept will be used in training the model. It is possible that a concept is present with a confidence score below threshold in an image and will not be used during training. Please see the ablation study for this quantity in Appendix C.1.

---

Q2: The reliance on pre-trained VLMs and object detectors may limit the approach's applicability to domains where such models are not readily available or effective.

A2: We would like to clarify that this is a general drawback with the use of foundational models and most existing CBMs [1-3] use foundation models to remove the need for human annotation in various phases of the CBM pipeline. For example, LF-CBM[1] uses CLIP model for assigning score to concepts, and existing methods like LaBO[2], LM4CV[3], and LF-CBM use LLMs for generating concept sets.

---

Q3: How would the VLG-CBM perform for OOD image recognition like NICO or Water-Bird, or how could it handle more complex tasks like VQA?

A3: Following your suggestion, we have conducted additional experiments to evaluate how VLG-CBM performs in OOD datasets. We trained our VLG-CBM on the CUB dataset and tested it on the Water-Bird dataset to evaluate the OOD performance. The Water-bird dataset is constructed by cropping out birds from the CUB dataset and transferring them onto backgrounds from the Places dataset. We compare the performance of VLG-CBM with the standard black-box model trained on the CUB dataset. It can be seen that our VLG-CBM generalizes well as the standard model does, which shows that our VLG-CBM is competitive and has very small accuracy trade-off with the interpretability compared with the standard black-box model.

For the question on more complex tasks like VQA, it is a very different task than the image classification task that current CBMs [1-3] are designed for, as a typical model for VQA task would involve the image embedding and question embeddings. Nevertheless, we think that our VLG-CBM can be potentially useful to convert the non-interpretable image embedding to an interpretable embeddings through a concept bottleneck layer. We will include this interesting question to future work and add a discussion in the revised draft.

---

Reference

[1] Oikarinen etal. Label-free concept bottleneck models. ICLR 23

[2] Yang etal. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. CVPR 23.

[3] Yan etal. Learning concise and descriptive attributes for visual recognition. ICCV 23.

---

Summary

In summary,

• In Q1, we added discussion on the performance of the grounding detector.

• In Q2, we respond to the concern on reliance on pretrained VLMs.

• In Q3, we conducted an extensive study on the waterbird dataset to show generalizability of our model to OOD datasets.

In response to the weakness part, we have addressed the reviewer’s concern

• in weakness #1 in Q1.

• in weakness #2 by comparing with other baselines in Q2.

• in weakness #3 by adding an extensive study on the waterbird dataset in Q3.

We believe that we have addressed all your concerns. Please let us know if you still have any reservations and we would be happy to address them!

Dear Reviewer hsJW,

Thank you for the positive feedback! Please see our responses below to address your comments.

Q1. Missing comparison with recent works in CMB, for example, [1] and [2]

A1: Thank you for providing the reference [1, 2]. Below we discuss the differences between our method and [1] and [2]

• [1]: Similar to us, [1] also uses an open-vocabulary object detection model to provide an explainable decision. However, their model is directly adapted from an OWL-ViT model, while our VLG-CBM uses an open-vocabulary object detection model to train a CBL over any base model, providing more flexibility. Additionally, their model requires pretraining to get best performance, while our VLG-CBM could be applied post-hoc to any pretrained model.
• [2]: This paper also aims at eliminating the information leakage. They evaluate it by measuring the performance drop speed after removing top-contributing concepts, which can be controlled by our proposed NEC metric. This is because the performance should reach minimum after removing all contributing concepts.

We will cite the papers [1,2] and discuss the differences in the revised manuscript.

[1]. Pham, Thang M., et al. "PEEB: Part-based Image Classifiers with an Explainable and Editable Language Bottleneck." arXiv preprint arXiv:2403.05297 (2024).

[2]. Sun, Ao, et al. "Eliminating information leakage in hard concept bottleneck models with supervised, hierarchical concept learning." arXiv preprint arXiv:2402.05945 (2024)

---

Q2: How good is LLM at generating candidate concepts, and how the different LLMs' quality would impact the generated candidate quality as well as final performance?

A2: We would like to clarify that the focus of this work is on proposing a novel CBM training pipeline by grounding concepts with open-domain object detectors and not on obtaining the concept set. As mentioned in Ln 110-112, we use the concept set from LF-CBM for training VLG-CBM and this could be replaced with concept sets obtained from other methods.

---

Q3: How good is the grounding model in terms of object class matching and box location prediction?

A3: In the Appendix D, we have provided a study evaluating concept annotations obtained from Grounding-DINO. We use CUB dataset for comparison which contains ground-truth for fine-grained concepts present in each image and report precision and recall metric to measure the quality of annotations from Grounding-DINO. As demonstrated in the appendix along with Table 2, Fig 4, Fig G.1 and G.2, the effect of false-positives is minimal and VLG-CBM is able to faithfully represent concepts in the CBL.

---

Q4: The ablation of different grounding models and even a random baseline should be added.

A4: Following your suggestion, we have conducted additional experiments and provided an ablation study with two different versions of GroundingDINO: Swin-T and Swin-B, where Swin-B is a larger model and has better open-vocabulary detection performance. We observe an increase in accuracy when using Swin-B for both NEC=5 and avg accuracy metric. This shows the potential of our method to scale as the performance of open-domain object detectors continues to improve.

For the random baseline, we have provided results in Table 2 in the original draft by starting with a randomly initialized CBL layer and training a linear layer with our sparsity constraints. This would be equivalent to randomly assigning concepts from the concept set to each image and training the final layer.

---

Q5: Gains with training augmentation based on bounding box

A5: Following your request, we reported our results in the Global Response #2.1.

---

Q6: If forcing the NEC to certain number, would the model automatically learn to assign different numbers of concepts to different classes? An analysis is needed.

A6: In Appendix E, we have shown the distribution of non-zero weight on CUB and Places365 models. As the Fig E.1 shows, our method automatically learns to use a different number of concepts for classifying different classes.

---

Summary

In summary,

• In Q1, we discussed and compared with recent works [1] and [2]
• In Q2, we answered the question on concept set generation.
• In Q3, we discussed the effect of grounding models on concept annotations
• In Q4, we provided an ablation study on the grounding model used in VLG-CBM.
• In Q5, we added an ablation study to study the impact of our augmentation probability.
• In Q6, we answered the question on the distribution of non-zero weights between different classes.

In response to the weakness part, we have addressed the reviewer’s concerns

• in weakness #1 by providing a comparison in Q1.
• in weakness #2 in Q2
• in weakness #3 by adding ablation study on object detection models and augmentation based on bounding box in Q3 and Q4
• in weakness #4 in Q6

We believe that we have addressed all your concerns. Please let us know if you still have any reservations and we would be happy to address them!