CLAP4CLIP: Continual Learning with Probabilistic Finetuning for Vision-Language Models

Continual learning (CL) aims to help deep neural networks learn new knowledge while retaining what has been learned. Owing to their powerful generalizability, pre-trained vision-language models such as Contrastive Language-Image Pre-training (CLIP) have lately gained traction as practical CL candidates. However, the domain mismatch between the pre-training and the downstream CL tasks often calls for finetuning of the CLIP on the latter. Most existing finetuning methods exhibit deterministic nature. This makes them overlook the many possible interactions across the input modalities and deems them unsafe for high-risk tasks requiring reliable uncertainty estimation. To address these, our work proposes Continual LeArning with Probabilistic finetuning (CLAP) - a probabilistic modeling framework over visual-guided text features per task, thus providing more calibrated CL finetuning. Unlike recent data-hungry anti-forgetting CL techniques, CLAP alleviates forgetting by exploiting the rich pre-trained knowledge of CLIP for weight initialization and distribution regularization of task-specific parameters. Cooperating with the diverse range of existing prompting methods, CLAP can surpass the predominant deterministic finetuning approaches for CL with CLIP. We conclude with out-of-the-box applications of superior uncertainty estimation abilities of CLAP including novel data detection and exemplar selection within the existing CL setups. Our code is available at \url{https://github.com/srvCodes/clap4clip}.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ey1B8/10 · confidence 3/52024-07-09

Summary

The authors present a novel method for CL called CLAP (Continual LeArning with Probabilistic finetuning) applied to the CLIP model. The technique employs probabilistic modeling to refine task-specific modules aligned with visual-guided text features, improving model adaptation to new tasks while mitigating the forgetting of old ones. It also benefits from the knowledge of pre-trained CLIP for this purpose. This method also offers compatibility with prompt-based finetuning methods as well. The approach demonstrates superiority over traditional deterministic finetuning methods through enhanced performance and better uncertainty estimation in diverse experimental setups.

Strengths

1. **Originality and Significance:** The paper introduces a unique approach to CL by integrating probabilistic finetuning with the CLIP model. The proposed pipeline is innovative and significant, as it addresses the critical issue of catastrophic forgetting in CLIP when trained on streams of tasks using well-justified solutions. 2. **Quality and Clarity:** The paper is well-written, with clear explanations of the methodology and its advantages. 3. **Technical Soundness:** The experimental setup is comprehensive, covering several datasets and comparative baselines. The results convincingly demonstrate the effectiveness of CLAP4CLIP in improving in-domain performance and generalization to new tasks. Especially achieving positive back-ward transfer in VTAB (Table 12) was a very interesting finding.

Weaknesses

1. **Presentation Issues:** Figure 1 could be larger to improve readability and clarity. The paper would benefit from better visualization to help convey the complex mechanisms of the proposed method more effectively. 2. **Reference Order and Citations:** The order of the references seems incorrect, which could potentially confuse readers. Additionally, Reference [36] should be discussed in the related works section as well to better contextualize the contributions of the paper.

Questions

Please address the mentioned weaknesses.

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

The limitations of this work are discussed and presented in Appendix E.

Reviewer KrxH4/10 · confidence 4/52024-07-12

Summary

This paper introduces CLAP4CLIP, a method designed to enhance continual learning (CL) using CLIP, a pre-trained vision-language model. The method leverages probabilistic fine-tuning with task-specific adapters to mitigate the issue of catastrophic forgetting commonly faced in CL. By incorporating visual-guided attention (VGA) modules, the model aims to align text features with frozen visual features during incremental training. The proposed method is evaluated on several datasets, including CIFAR100, ImageNet100, ImageNet-R, CUB200, and VTAB, and compared against multiple baselines and state-of-the-art fine-tuning methods.

Strengths

1. The paper presents a novel approach by integrating probabilistic fine-tuning and visual-guided attention into CLIP for continual learning. The use of task-specific adapters and the Bayesian variational inference framework adds a unique angle to the existing methods. 2. The methodology is well-structured, and the experiments are thorough, comparing the proposed method against a wide range of baselines and state-of-the-art approaches. 3. Addressing catastrophic forgetting in CL is a significant challenge, and the proposed method offers a promising solution. The integration with CLIP, known for its zero-shot learning capabilities, highlights the potential impact on real-world applications.

Weaknesses

1. Given CLIP’s strong generalization abilities and its effectiveness in zero-shot learning, the reliance on replay strategies might seem redundant and potentially underutilizes CLIP’s full capabilities. 2. The paper does not explicitly detail whether the VGA module is updated during incremental tasks. This is a critical aspect, as the update strategy could significantly impact the model's performance and stability. 3. The method involves multiple components, such as probabilistic fine-tuning, VGA modules, and task-specific adapters, which may introduce significant computational overhead.

Questions

1. Could you provide more insights into why the replay strategy is necessary and how it complements the use of CLIP in your method? 2. How is the VGA module handled during incremental tasks? Are its parameters updated, and if so, what strategy is used to ensure consistency and stability across tasks? 3. What are the computational requirements of your method compared to the baselines, especially concerning the additional components introduced? How do you balance the trade-off between performance improvements and computational efficiency?

Rating

4

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

*

Reviewer bibK3/10 · confidence 4/52024-07-13

Summary

The paper emphasizes on the existing limitations of deterministic approaches in fine-tuning and highlights the need for probabilistic fine-tuning approach. Following this, it proposes a probabilistic parameter efficient fine-tuning method for continually learning vision language models like CLIP.

Strengths

1. The proposed approach seems novel.

Weaknesses

1. Justification of probabilistic modelling of the text feature and not the image feature space is not clear. 2. The approach is highly inefficient in terms of inference time. 3. Recent approaches like ConvPrompt[a], CODA-Prompt[b], HiDe-Prompt[c], and SLCA[d] not compared. [a]Roy, Anurag, et al. "Convolutional Prompting meets Language Models for Continual Learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024 [b]Smith, James Seale, et al. "Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [c]Wang, Liyuan, et al. "Hierarchical decomposition of prompt-based continual learning: Rethinking obscured sub-optimality." Advances in Neural Information Processing Systems 36 (2024). [d]Zhang, Gengwei, et al. "Slca: Slow learner with classifier alignment for continual learning on a pre-trained model." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.

Questions

1. Could you address the points I raised in the weakness section? 2. The paper demonstrates a ~2% performance improvement with the addition of the memory consolidation component. However, I'm interested in seeing how the proposed method performs when the replay memory size is set to zero.

Rating

3

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations of their work in the paper.

Reviewer giBN5/10 · confidence 4/52024-07-15

Summary

This paper proposes Continual Learning with Probabilistic Finetuning (CLAP) for class-incremental learning using CLIP. The key modules of the proposed idea are as follows. First, the authors introduce a CLIP-based probabilistic finetuning model using Bayesian Variational Inference to achieve better generalization during the continual learning process. Second, they propose a visual-guided attention (VGA) model to facilitate cross-modal feature alignment in the continual learning process. Lastly, to alleviate forgetting of pre-trained language-aware CLIP knowledge, they suggest past-task distribution generalization. Additionally, to enable parameter-efficient learning, a probabilistic adapter is used, along with a method for its initialization to enhance stability. Experimental results on various datasets demonstrate that the proposed algorithm achieves superior class-incremental learning performance compared to existing algorithms.

Strengths

The strengths of this paper are as follows: 1. The paper is well-written and easy to understand. 2. The proposed modules for successful class-incremental learning using CLIP are meticulously designed. Through various analyses and ablation studies, the roles of each module are thoroughly demonstrated. Although the algorithm appears somewhat complex compared to existing algorithms, parameter and time analyses show that the actual cost difference is negligible. 3. In class-incremental learning experiments using diverse datasets, the proposed algorithm consistently outperforms existing algorithms across various evaluations.

Weaknesses

1. I have no major concerns regarding the contribution of the proposed algorithm for class-incremental learning using CLIP. However, I have some questions based on the experimental results. 1-1) Unlike algorithms like L2P and DualPrompt, which utilize exemplar memory, the paper does not use exemplar memory. Are the results in L2P and DualPrompt sections of the paper based on using exemplar memory in a fair comparison? To ensure a fair comparison, results using exemplar memory should be shown. 1-2) The paper only considers prompt-based algorithms like L2P and DualPrompt using Vision Transformer (ViT). However, newer prompt-based algorithms (e.g., CODA-Prompt) that achieve better performance and representation-based algorithms (e.g., Ranpac and EASE) that generally improve performance have been proposed. The authors should consider these algorithms as additional baselines. For more details, please refer to this survey paper [1]. 1-3) Among existing baselines, PROOF is currently the state-of-the-art algorithm. While Table 1 shows PROOF's results, Tables 2 and 16 do not. To validate the superiority of the proposed algorithm, results for PROOF should be shown in these tables and other key experiments. 2. Recent research has highlighted discussions on hyperparameter tuning for class-incremental learning algorithms [2] and pointed out issues regarding computational costs[3]. From this perspective, I believe that further discussion and additional analysis on computational costs regarding the proposed algorithm would make the paper more convincing. 3. Lastly, I have personal concerns regarding the necessity of class-incremental learning research using CLIP. As shown in [1], algorithms using only ViT have already achieved excellent performance in class-incremental learning research. Despite using pretrained ViT, one of their major advantages is achieving superior performance without using exemplar memory. Considering this perspective, the proposed research not only uses a pretrained CLIP model but also requires text information and additional exemplar memory usage. Simply comparing numerical experimental results suggests that the proposed algorithm does not achieve overwhelmingly superior performance compared to algorithms using only pretrained ViT. In light of this, I would like to understand: 1) What justifies the necessity of class-incremental learning research using CLIP with exemplar memory? Additionally, 2) I am curious about the potential for the proposed algorithm's ideas to be applied to continual learning in other domains using CLIP or continual pre-training of CLIP itself. [1] "Continual learning with pre-trained models: A survey." arXiv preprint arXiv:2401.16386 (2024). [2] "Hyperparameters in Continual Learning: a Reality Check." arXiv preprint arXiv:2403.09066 (2024). [3] "Computationally budgeted continual learning: What does matter?." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.

Questions

See the Weakness section. I don't have significant concerns about the algorithm proposed in this paper, but I couldn't give it a higher score due to some experimental uncertainties and questions about the necessity of the setting. If the authors address these concerns in the rebuttal, I would gladly raise my score.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

There is no potential negative societal impact of this paper.

Reviewer giBN2024-08-09

I would like to thank the authors for conducting additional experiments and providing an author response to address the concerns raised in my review. I have thoroughly read both the author response and the updated PDF, and as a result, I am increasing my score to 5. I hope that the final version of the paper will fully incorporate the contents of the response.

Reviewer ey1B2024-08-11

Thank you to the authors for addressing my concerns. I will be keeping my original score.

Authorsrebuttal2024-08-11

Requesting feedback on our rebuttal

Dear Reviewer KrxH, We thank you again for taking the time to review this work. We put our best effort into preparing the rebuttal to your questions, including running experiments without memory replay. We would very much appreciate it if you could provide us with your feedback on our rebuttal. We would be glad to answer any further questions and clarify any concerns. Also, if you are satisfied with our answers, please consider revising your score. With best regards

Authorsrebuttal2024-08-11

Requesting feedback on our rebuttal

Dear Reviewer bibK, We thank you again for taking the time to review this work. We put our best effort into preparing the rebuttal to your questions, including reporting the experiments on zero replay memory size, comparing with CODA-Prompt, and justifying design choices/inference time overhead. We would very much appreciate if you could engage with us through your feedback on our rebuttal. We would be glad to answer any further questions and clarify any concerns. Also, if you are satisfied with our answers, please consider revising your score. With best regards

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC