BendVLM: Test-Time Debiasing of Vision-Language Embeddings

Thank you for your review and thoughtful questions. We address them below:

Why Does Performance Not Degrade After Debiasing? Yes, part of this is likely due to the fact that we do not perform any finetuning and find the embedding that is maximally similar to the initial embedding while maintaining our fairness constraints.

Bend-VLM Runtime: We’ve now conducted a runtime analysis, and determined that the time added by our current implementation of Bend-VLM over the baseline falls in the range of 1.53 to 1.76 seconds 95% of the time. We didn’t optimize our code for runtime, however, and it is possible that a smaller LLM could perform the augmentation task.

Is Step 2 Of Bend-VLM Necessary?

We’ve added a new ablation experiment to confirm that yes, both steps contribute to the success of Bend-VLM.

Table 1 in the PDF attached to the general rebuttal shows that while most of the Worst-Group Accuracy performance comes from Step 1, utilizing only step 1 results in a much more biased retrieval metric by having a much higher KL divergence from a fair distribution. Utilizing step 2 alone results in a fair retrieval roughly equivalent to the full Bend-VLM approach, but does not have as good of a Worst Group Accuracy. We achieve the best results by combining Step 1 and Step 2 to make the full Bend-VLM approach.

Why are CCF and CDF Not Reported For All Methods? We reported KL Divergence and MaxSkew as they are the standard metrics reported by other related works [1,2].

[1] Berg, Hugo, et al. "A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning." arXiv preprint arXiv:2203.11933 (2022).

[2] Chuang, Ching-Yao, et al. "Debiasing vision-language models via biased prompts." arXiv preprint arXiv:2302.00070 (2023).

The Example Shown in Figure 1 Could Be Improved: Thank you for this suggestion. We are happy to change the example from Nurse into a different class that exhibits stronger bias in the camera ready version. We just used Nurse for this figure since that class was the running example referenced in the rest of our paper.

Thank you for your review and constructive feedback. We address your questions below:

The “Open Set” Ability of Bend-VLM: We don’t need to know the class of the instance prior to training (and there is not a distinct training phase at all, since we do test time adaptation). We do need to know the set of attributes that we wish to debias, e.g. we know that we want to debias for race and gender. This isn’t as strong of an assumption, and the fact that we know which protected attributes we want to debias for is a common assumption in the literature [1, 2].

[1] Berg, Hugo, et al. "A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning." arXiv preprint arXiv:2203.11933 (2022).

[2] Chuang, Ching-Yao, et al. "Debiasing vision-language models via biased prompts." arXiv preprint arXiv:2302.00070 (2023).

How does Bend-VLM Overcome Catastrophic Forgetting?: We argue that Bend-VLM is less susceptible to catastrophic forgetting than fine tuning methods for the following reasons:

• We update only the vector embedding, but do not alter any weights of the model itself.
• Bend-VLM can be selectively applied to input, such as only to input queries where the subject is something that can have a gender/race/other protected attribute.
• Step 2 of Bend-VLM finds the vector embedding minimally distant from the original vector embedding that satisfies the fairness constraint — something not guaranteed by fine tuning approaches. We thus minimize the risk of breaking or modifying other associations that the VLM has learned.

Definitions and Theorems Should Be Reorganized: We are happy to take this suggestion, and reformat the definitions/theorems for readability and formality in the camera-ready version.

Thank you for your review and thoughtful questions. We address them below:

W1: Whether Bend-VLM works when there is a distribution shift between the reference and target domains is an excellent question. We have conducted a new experiment that shows that Bend-VLM still outperforms the competition when the reference dataset is OOD, indicating that the reference dataset does not need to fully match the distribution of the target distribution.

In this new experiment, we use FairFace as the reference dataset while CelebA is the target dataset. While Bend-VLM with this OOD reference dataset does not perform as well as Bend-VLM with an in-distribution reference dataset, it still outperforms the other compared approaches. See Table 2 in the PDF attached to the general rebuttal for details.

W2:

We look at intersectional bias in a new experiment where we debias FairFace with respect to Gender for HairColor queries, but evaluate on Race. We do not expect to see improvements with respect to racial bias after gender debiasing for any method. We observe that racial bias goes up for all debiasing methods after gender debiasing (see the Table below). This reflects a known, frustrating “Whac-A-Mole” issue where debiasing for one attribute often increases the bias of another attribute [1]. Interestingly, we do not see racial bias increase when performing only Step 2 of the Bend-VLM debiasing, indicating that this shortcut issue is most strongly affected by the orthogonalization operation performed in Step 1. The other debiasing methods also perform a similar orthogonalization step and likewise experience this shortcut problem.

[1] Li, Zhiheng, et al. "A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.

Thank you for your thoughtful feedback. We address your concerns below:

Missing Ablation Study: We’ve added an ablation study based on your suggestion. The ablation study results are in Table1 in the PDF attached to the general rebuttal. Results show that while most of the Worst-Group Accuracy performance comes from Step 1, utilizing only step 1 results in a much more biased retrieval metric by having a much higher KL divergence from a fair distribution. Utilizing step 2 alone results in a fair retrieval roughly equivalent to the full Bend-VLM approach, but does not have as good of a Worst Group Accuracy. We achieve the best results by combining Step 1 and Step 2 to make the full Bend-VLM approach.

Efficiency: We’ve added a runtime analysis, and determined that the time added by our current implementation of Bend-VLM over the baseline falls in the range of 1.53 to 1.76 seconds 95% of the time. We didn’t optimize our code for runtime, however, and it is possible that a smaller LLM could perform the augmentation task.

Can Bend-VLM Work With Non-CLIP Models? Yes, Bend-VLM just requires a VLM that produces a vector representation of text and images in a unified embedding space. We added a new experiment where FLAVA [1] is used as the VLM instead of CLIP. Table 3 in the PDF attached to the general rebuttal shows that Bend-VLM still outperforms the compared methods when FALVA is the VLM. Results shown for the CelebA dataset. Note that there are no “ground truth” labels for the stereotype queries, so it isn’t possible to compute AUC for them.