Multi-Prompt Alignment for Multi-Source Unsupervised Domain Adaptation

Most existing methods for unsupervised domain adaptation (UDA) rely on a shared network to extract domain-invariant features. However, when facing multiple source domains, optimizing such a network involves updating the parameters of the entire network, making it both computationally expensive and challenging, particularly when coupled with min-max objectives. Inspired by recent advances in prompt learning that adapts high-capacity models for downstream tasks in a computationally economic way, we introduce Multi-Prompt Alignment (MPA), a simple yet efficient framework for multi-source UDA. Given a source and target domain pair, MPA first trains an individual prompt to minimize the domain gap through a contrastive loss. Then, MPA denoises the learned prompts through an auto-encoding process and aligns them by maximizing the agreement of all the reconstructed prompts. Moreover, we show that the resulting subspace acquired from the auto-encoding process can easily generalize to a streamlined set of target domains, making our method more efficient for practical usage. Extensive experiments show that MPA achieves state-of-the-art results on three popular datasets with an impressive average accuracy of 54.1% on DomainNet.

Paper

Similar papers

Peer review

Reviewer 34CN6/10 · confidence 4/52023-07-03

Summary

This paper proposes an interesting pipeline for large pre-trained model-based UDA, and realizes a latent subspace tuning for continuous adaption.

Strengths

1. The motivation behind this article is very meaningful, and the proposed method supports the motivation well. 2. In terms of methodology, the design of this paper is clear and innovative for the combination of existing technologies. 3. This easy but effective baseline and latent subspace tuning ability may provide new paradigm for future UDA. 4. Good performances.

Weaknesses

Using CLIP as the backbone may limit the ability of MPA on only classification tasks, how to extend MPA to other tasks, like segmentation, referring, etc? And how to utilize other large pre-trained models like GPT-3?

Questions

Please see the weakness part.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors should provide the limitation of their work.

Authorsrebuttal2023-08-18

Have our rebuttal addressed your concerns?

Dear reviewer 34CN, we would be grateful if you could confirm whether our response has addressed your concerns. Please do not hesitate to let us know whether there is anything else you would like to see clarified or improved before the end of the rebuttal period.

Authorsrebuttal2023-08-20

Have our rebuttal addressed your concerns

Dear reviewer, as the end of the rebuttal period is approaching, we would like to know whether our rebuttal have addressed your concerns.

Reviewer UUa15/10 · confidence 4/52023-07-04

Summary

This paper introduces prompt learning to multi-source unsupervised domain adaptation (UDA). Firstly, individual prompts for each source and target pair are learned using a contrastive loss. Then, MPA aligns the learned prompt by an autoencoder-based step with an L_1 constraint to generate consistent results for the same target domain image. In addition, the LST strategy delivers the first target domain adaptation to the subsequent target domains efficiently. Experiments on ImageCLEF, Office-Home, and DomainNet validate the effectiveness.

Strengths

+ This paper's application of prompt learning to multi-source UDA problems is groundbreaking. + This paper is well-written and easy to understand.

Weaknesses

+ Regarding the prompt design part in sec3.2, can it only be realized by following Ge[10]? There is a lack of novelty and contribution due to the lack of original works in the whole prompt design. + Regarding the problem of reducing the dimensionality of the high-dimensional learned prompt, is this step introduced because the prompt design is not optimal? From the perspective of Tab. 6, the growth brought by AE is very limited, so are other methods of dimensionality reduction effective, or even the step of AE can be removed by considering redundant information in prompt design? 3 The LST in 3 sec 3.3 is a further extended use of MPA, which is somewhat insufficient as an independent innovation point. 4 This paper needs to weaken the sense of the existence of CLIP and the dependence on the previous prompt design method, otherwise very much like a simple application essay.

Questions

In addition to the questions in the above weaknesses that need to be answered, there are a few things that need to be clarified. + The generalization of CLIP is a huge advantage. Although the author discussed the impact of CLIP in 4.3, I hope that the author will discuss the design experiment of the prompt. In addition to using the original "a photo of [CLS]," the number of channels of the trainable prompt params can also be discussed. + Please state the problems and experimental results faced when the approach of Ge [10] is transferred to the tasks addressed in this paper. And for the problem and results, describe the differences of this paper. + For now, it is unknown why prompt, as a trained parameter, has class-specific and domain-specific attributes.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Not applicable

Authorsrebuttal2023-08-18

Have our rebuttal addressed your concerns?

Dear reviewer UUa1, we would be grateful if you could confirm whether our response has addressed your concerns. Please do not hesitate to let us know whether there is anything else you would like to see clarified or improved before the end of the rebuttal period.

Authorsrebuttal2023-08-20

Have our rebuttal addressed your concerns?

Dear reviewer, as the end of the rebuttal period is approaching, we would like to know whether our rebuttal have addressed your concerns.

Reviewer tF9m5/10 · confidence 4/52023-07-05

Summary

This paper deals with the multi-source domain adaptation problem. It proposes to tune the designed domain-invariant prompts and domain-specific prompts to enable the domain adaptation ability. Generally, the training consists of two objectives, i.e., the individual prompt learning objective and the de-noising objective via prompt auto-encoders. To enable a test-time domain adaptation ability and reduce the amount of learnable parameters, the authors propose LST. Experiments on various multi-source benchmarks verify the effectiveness of proposed method.

Strengths

- The idea is simple and generally reasonable. - The experiment results are good.

Weaknesses

- In LST, I don't think it is reasonable to use randomly initialized representations as the input of the back-projector to perform prompt reconstruction, as the latent representations don't subject to the same distribution. And there is no empirical evidence to show such a way really works. - Missing comparisons to strong baselines. E.g., on DomainNet, [1] achieves a multi-source domain adaptation performance at 53.2 with a ResNet-101 backbone. The authors should compare to the strong baselines. [1] Contrastive adaptation network for single- and multi-source domain adaptation, TPAMI 2020.

Questions

See the weakness part.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

I don't see any serious potential negative societal impact of this work.

Authorsrebuttal2023-08-18

Have our rebuttal addressed your concerns?

Dear reviewer tF9m, we would be grateful if you could confirm whether our response has addressed your concerns. Please do not hesitate to let us know whether there is anything else you would like to see clarified or improved before the end of the rebuttal period.

Authorsrebuttal2023-08-20

Have our rebuttal addressed your concerns?

Dear reviewer, as the end of the rebuttal period is approaching, we would like to know whether our rebuttal have addressed your concerns.

Reviewer tSjx6/10 · confidence 4/52023-07-09

Summary

The paper proposes an extension of [10] (Domain Adaptation via Prompt Learning Ge et al., 2022) to the multi-source UDA set-up. (i) Distinct soft prompts are learnt via contrastive loss for each source-target pair; each source-target prompt is composed of class-wise source- and target-prompts. In the target domain, learning is achieved by utilizing pseudo-labels from the pretrained CLIP model, which possesses strong zero-shot ability. To encourage consistent target outputs across all learned prompts, a consistency L1 loss is applied to the soft outputs of all target samples. (ii) Prompt reconstruction: to eliminate redundant information that may impede performance, an autoencoder (AE) is learned to "denoise" the acquired prompts. (iii) Latent Subspace Tuning (LST): efficient adaptation to a new target domain, after learning from the first target domain, is accomplished by optimizing on the latent space of the learned AE using pseudo-labels of the new target. The proposed framework demonstrates superior results compared to previous State-of-the-Art (SOTA) methods in multi-source UDA.

Strengths

The proposed method leverages the powerful zero-shot capability of the pretrained CLIP model for the multi-source domain adaptation (DA) task. Empirical results demonstrate its effectiveness across various benchmarking setups. The presentation is good, enabling easy comprehension of the method.

Weaknesses

The main concern of this work is the lack of technical novelty and a more rigorous evaluation. The proposed framework is a straightforward extension of the work by Get et al. (2022) [10] to the multi-source setting. In terms of method evaluation, it is expected to compare against stronger CLIP-based baselines, both in the multi-source UDA experiments (Tables 1 and 2) and in the LST experiment (Table 3). Further details are provided in the following section.

Questions

Given that CLIP alone produces results close to the state-of-the-art (SOTA) as shown in Table 1, and even outperforms the SOTA as shown in Table 2, my main concern is whether the authors have made adequate efforts to create decent baselines using CLIP-pretrained models. I appreciate their effort in Section 4.3 when they ``swap MFSAN’s ResNet50 backbone pre-trained on ImageNet to CLIP’s image encoder''. However, I would love to know more details to know as if careful considerations were taken into account for MFSAN + CLIP or simply the authors swap the backbone's weight. For example, did the authors preserve the text classifier and the contrastive los of CLIP ? As the proposed framework is based on self-training with pseudo-labels, I wonder if how the CLIP + self-training and Single Prompt + self-training baselines perform. Furthermore, I am curious to know if CLIP with fixed prompts has been pushed to its maximum extent. In other words, how far can this naïve baseline reach with better prompt engineering instead of using the simple prompt "a photo of [CLS]"? One can imagine incorporating domain-specific information into the prompt or defining a set of templates rather than relying on just one template. In the LST experiment, how good is MPA (on the first domain) + self-training? For efficiency, one can try test time prompt tuning (TPT).

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

This work lacks significant technical novelty, and the experiment could be improved by using stronger baselines. Based on the current state of the submission, my recommendation leans towards the negative side. ---- After rebuttal ---- The rebuttal is convincing and has helped clarify most of my technical concerns. I believe that this work is indeed interesting and help advocate the usage of prompts in domain adaptation. While the novelty limitation still pertains, the original set of experiments, along with the new ones provided during the rebuttal, is sufficient. I believe the paper after revision would certainly pass the NeurIPS' bar.

Authorsrebuttal2023-08-18

Have our rebuttal addressed your concerns?

Dear reviewer tSjx, we would be grateful if you could confirm whether our response has addressed your concerns. Please do not hesitate to let us know whether there is anything else you would like to see clarified or improved before the end of the rebuttal period.

Authorsrebuttal2023-08-20

Have our rebuttal addressed your concerns?

Dear reviewer, as the end of the rebuttal period is approaching, we would like to know whether our rebuttal have addressed your concerns.

Reviewer UUa12023-08-20

Additional Comments

Thank you to the author for the careful responses that solved most of my questions. I'm keeping my score unchanged, mainly considering the lack of novelty of the prompt method proposed and the closeness of the implementation of LST to MPA. I appreciate the motivation for this article, which is the main reason for me to hold the current opinion. It would be better to do some prompt and structural innovation.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC