Cross-modality debiasing: using language to mitigate sub-population shifts in imaging

Weakness and Question 3

We agree with the reviewer that there are challenges in selecting appropriate text prompts when dealing with language (text) related deep models, and varying text prompts have a significant influence on the final performance, and this work does need a domain explicit validation dataset to guarantee reasonable worst-case performance, as mentioned in Section 3.2 and Section 5 (Experiment) and employed by related work in this line of research. (JTT and $\chi^{2}$-DRO generally use this validation dataset for another purpose, detailed in ‘stable worst-case accuracy across training epochs’ paragraph of section 5.1). However, our solution enables fine-tuning an extra adapter (in the last layers) with language guidance to mitigate subpopulations in the label space, which follows the trend of the broader and broader application scenarios of the (large) multi-modal models. We believe this type of condition closely satisfies real-world conditions instead of assuming it is a specific condition.

Question 1

Please refer to Appendix C, we have explored the failure mode on ViT-L/14. The root cause is the inconsistency of the effectiveness of text prompts over various models. Employing a prompt that is different from the default one achieves a more reasonable performance, it shows that the proposed method outperforms other baselines, as shown in Table 10.

Question 2

Table 6 empirically verifies the potential of L-DRO to help stabilize the existing methods since they are orthogonal, and it demonstrates that the proposed method embedded with other baselines generally improves the performance with CelebA but worsens it with the Waterbirds dataset. And, it is not surprising that some baseline methods actually worsen the performance compared with zero-shot performance such as ERM and JTT. Actually, it is well-established that inappropriate fine-tunings damage the distributional robustness of pre-trained models [1,2,3], which is also mentioned in the introduction.

It should be noted that parameter-efficient tuning, such as tuning the final layers only, demonstrates popularity under certain large model conditions. The existing work of DRO such as JTT typically tunes the whole model.

Question 4

It is true that training with fewer data points or data that are not enough to represent the whole distribution manifest performance decreasing. When data is not enough to represent the whole distribution, then the updating direction is hard to find the right one or even in the opposite direction. Table 3 shows the proposed method is data efficient and is possibly due to its parameter-efficient structure.

[1] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7959–7971, 2022.

[2] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022.

[3] Xin Zhang, Yusuke Iwasawa, Yutaka Matsuo, and Shixiang Shane Gu. Amortized prompt: Lightweight fine-tuning for clip in domain generalization. arXiv preprint arXiv:2111.12853, 2021.

We greatly appreciate your constructive comments. Please see our responses below;

Weakness 1

In our loss design, $\eta$ is expected to balance the two items, consistency loss and similarity loss. And we reported the performance under CelebA dataset and network structure CLIP (ViT-B/32) with varying $\eta$ here to show that the proposed method is not sensitive to the hyper-parameter.

η = 0.1, Avg. Acc. = 0.845±0.006, W.C. Acc. = 0.782±0.016.

η = 0.2 (default), Avg. Acc. = 0.836±0.003, W.C. Acc. = 0.792±0.013.

η = 0.3, Avg. Acc. = 0.845±0.003, W.C. Acc. = 0.771±0.028.

η = 0.4, Avg. Acc. = 0.844±0.004, W.C. Acc. = 0.777±0.020.

η = 0.6, Avg. Acc. = 0.844±0.008, W.C. Acc. = 0.763±0.035.

η = 0.8, Avg. Acc. = 0.842±0.006, W.C. Acc. = 0.771±0.037.

η = 1.0, Avg. Acc. = 0.842±0.007, W.C. Acc. = 0.764±0.040.

Weakness 2

We appreciate your careful observation of Table 1 that the improvements in worst-case accuracy seem to slightly improve the average accuracy under varying text prompts with RN50. We believe it may imply that our method improves the performance of the worst-case group meanwhile it will least damage the the performance of other groups in such conditions. So that the overall performance (average accuracy) gets improved. We conducted the following simple experiments. Under CelebA, CLIP (RN50), and text prompts "a picture of [] people", we report the average accuracy, worst-case accuracy, and additional accuracy for each subgroup over ten seeds.

Zero-shot - acc_avg= 0.837, acc_worst = 0.757, acc_each_group = [0.833, 0.757, 0.808, 0.907],

L-DRO - acc_avg = 0.859, acc_worst = 0.789, acc_each_group = [0.789, 0.814, 0.808, 0.908].

The above results support our claims.

Question 1

We really appreciate your careful reading and thank you for pointing out the typo (y_p); it has been corrected (See the uploaded revised PDF). Your attention to detail is invaluable.

Thank you for taking the time to read and review our paper. We sincerely appreciate your keen interest in the technical aspects of our work and positive feedback. Please see our responses below;

Weakness

We admit the challenges in designing debiasing guidance from other modalities instead of language modality. And, our solution enables fine-tuning an extra adapter in the last layers with language guidance to mitigate subpopulations in the label space, it follows the trend of the broader and broader application scenarios of the (large) multi-modal models, relying on the guidance from natural language in learning (such as in-context learning), and parameter-efficient fine-tuning regime, which we believe closely matches real-world conditions.

Question 1 and 2

In current work, there are no such attempts to adapt L-DRO for modalities other than vision and language, however, we believe the principle developed in this work can be easily adapted to other types of data modalities. For Table 4, there is no clear pattern to describe divergences in worst-case accuracy with unaligned semantic relations currently. We do have similar observations that subgroups are (minorly) adversely affected, it serves as fundamentals of combining multiple debiasing sources, specifically, the last two rows.

We greatly appreciate your constructive comments. Please see our responses below;

Weakness 1

We appreciate the insights regarding our method. Projecting all the debiased feature representation onto the null-space of $[T(\hat{t_{1}}), T(\hat{t_{2}})]^{T}$ is not the only goal since our solution also encourages the similarity of feature representations before and after the adapter with a consistency loss. Without the consistency loss, we believe all the feature representations may collapse to, e.g., one fixed direction that is perpendicular with $T(\hat{t_{1}})$ and $T(\hat{t_{2}})$. However, with our consistency loss and $\eta$ (scalar to balance the above two terms), our results empirically negate the collapse where the average accuracy regarding the task basically does not change compared with zero-shot learning.

Weakness 2

The works cited ([1]) differ significantly from ours in scope. For work targeting mitigating sub-population shift, there are two categories: a) domain-oblivious in which the membership of each input sample is unknown during training, and b) domain-aware settings [1]. Our work targets the setting, domain-oblivious, as well as the selected baselines. However, the mentioned work [1] “re-trains the last linear layer on a small dataset where the backgrounds are not spuriously correlated with the foreground”, which belongs to the second setting, domain-aware. The domain-oblivious, as studied in this paper, is a more challenging setting than domain-aware ones such as [1], since the parameter updating does not have any guidance from the “good data”. As mentioned in Section 3.2, popular strategies along this line of work usually find oblivious evidence in the loss signal, and the high-level idea can be formulated as distributional robust optimization.

Weakness 3

Apologies for any confusion. A quick answer to the misalignment is due to this work only tunes a simple adapter network instead of full model tuning such as JTT and other baselines in the original papers. We mentioned this part in our objective Equation 4 and highlighted the architecture as I▷A▷T in Table 2. To further clarify, we added additional footnotes in Table 2 - only adapter A is tunable. (See the uploaded revised PDF). Note that existing works of DRO such as JTT typically tune the whole model while in this work and some large model conditions, parameter-efficient tuning, such as tuning the last two layers only, exhibits popularity.

Question 1

We want to clarify that ERM (Empirical Risk Minimization) in Table 2 is exactly the result you look for. As expected, ERM generally does not have satisfying performance over sub-population shifts.

Question 3

In Table 4, this work investigates the effect of unaligned debiasing and the combination of multiple debiasing sources. Firstly, the unaligned debiasing does not improve the performance under sub-population shifts. And, the last two rows in Table 4 demonstrate the performance of combining multiple debiasing sources.

Thank you for pointing out the minor issues, and we have corrected them.

[1] Zhai R, Dan C, Kolter Z, et al. Doro: Distributional and outlier robust optimization[C]//International Conference on Machine Learning. PMLR, 2021: 12345-12355.

Dear Reviewer FPNj,

We hope our response has answered your questions and clarified the concerns. If there are any further questions that require our attention, please feel free to inform us before the closure of the interactive rebuttal system.

Appreciate your time and happy Thanksgiving.