Multi-Label Test-Time Adaptation with Bound Entropy Minimization

We thanks for your valuable suggestions and we will try to address your concerns as follows. We are eager to engage in a more detailed discussion with you.

Weakness 1: Determine the setting of k. How to make sure the weak label set is reliable?

1.Determine the setting of k.

• K is determined based on the number of textual labels contained in the paired caption, it is not a hyperparameter. For each augmented view $\mathbf x_i$, we retrieve the most similar caption $\mathbf t_i$ for $\mathbf x_i$. For a specific pair $\langle \mathbf x_i, \mathbf t_i\rangle$, assuming $\mathbf t_i$ is A black Honda bicycle parked in front of a car, we follow the noun filtering in PVP [1] to extract the label set {bicycle, car} from $\mathbf t_i$ with the size of 2.
• This label set serves as both the strong label set for the caption and the weak label set for the view, hence the value of k is 2. If $\mathbf t_i$ is A group of girls enjoying a game of frisbee while sitting on chairs, the label set would be {girls, frisbee, chairs}, and the value of k would be 3.

2.How to make sure ... is reliable.

• Captions primarily describe salient visual information in the image and may not accurately reflect smaller object categories within the image. Therefore, for the caption “A black Honda bicycle parked in front of a car”, we refer to the label set {bicycle, car} as the strong label set for the caption as it is directly extracted from the caption. Likewise, it also serves as the pseudo labels for the view, $i.e.$, weak label set.
• Moreover, we aim to build a TTA framework for multi-label scenarios, demonstrating the feasibility of traditional entropy minimization methods in multi-label instances. In practical applications, we can consider employing a more robust similarity retrieval strategy, integrating label sets from multiple captions, or constructing a more comprehensive text description base to enhance the reliability of the weak label set.

Weakness 2: Explanation of "augmented view".

• We apologize for the confusion of the definition. Augmented view is a widely adopted method in the TTA domain, which involves generating a set of $N$ different views through data augmentations. Then, TTA selects the top 10% highest confidence views and minimizes the marginal entropy of these views, encouraging consistent and confident predictions.
• Our work follows the same method in the multi-label TTA scenario, performing $N$ data augmentations on multi-label instances and optimizing the view and caption prompts with the proposed Bound Entropy Minimization to enhance the consistency of model predictions.

Weakness 3: Add ML-TTA specific framework for comparison.

• Currently, TTA methods primarily focus on multi-class scenarios, adapting single-label instances through entropy minimization. However, for multi-label instances, considering only the top-1 class inevitably harms the prediction performance of other positive labels.
• To our knowledge, our work is the first to explore the feasibility of entropy minimization in multi-label scenarios. The proposed Bound Entropy Minimization (BEM) aims to simultaneously increase the confidence of multiple top-k labels. Therefore, we select the SOTA methods in the TTA field for multi-class scenarios as benchmarks, such as RCLF [2] and TDA [3]. Moreover, our work demonstrates the feasibility of entropy minimization in multi-label TTA and provides a basic framework for subsequent multi-label TTA tasks.

Weakness 4: Explanation of evaluation metric in Table 1. Marking the second-best result in the experiments.

• The evaluation metric in Table 1 is the widely used mean average precision (mAP) in multi-label classification tasks. mAP is the average of Average Precision (AP), where AP is the area under the Precision-Recall curve. Precision is the proportion of truly positive samples among all samples predicted as positive by the model, and Recall is the proportion of truly positive samples that are correctly predicted as positive by the model.

• In multi-label classification, for each category, we can draw a Precision-Recall curve and calculate the area under this curve, which is the AP for that category. Therefore, the calculation steps for mAP are: Calculate the AP for each category. Take the average of all category AP values to get mAP. The mAP is computed as follows: $mAP=\frac{1}{L}\sum_{i=1}^L AP_i$, where $L$ is the number of categories, and $AP_i$ is the average precision for the i-th category.

• We will introduce mAP in detail and mark the second-best result in the experiments in the future version.

[1]. TAI++:Text as Image for Multi-Label Image Classification by Co-Learning Transferable Prompt. IJCAI 2024

[2]. Test-Time Adaptation with CLIP Reward for Zero-Shot Generalization in Vision-Language Models. ICLR 2024

[3]. Efficient Test-Time Adaptation of Vision-Language Models. CVPR 2024

Dear Reviewer, we are looking forward to your professional suggestions and hope to receive your guidance to further discuss and refine the contents of the work. We are eager for your response. Thank you.

• We appreciate your careful review of our work. Throughout the research of ML-TTA, we conducted extensive investigation and discourse, including a large number of papers and surveys [1][2][3][4] on TTA. To our knowledge, our work is the first study to examine TTA within multi-label scenarios.
• To facilitate a relatively fair comparison, in our experiments, we selected and adjusted current SOTA methods for multi-class TTA to the multi-label scenarios. Following your suggestion, we further added and adjusted more latest multi-class TTA methods for comparison on ViT-B/16 architecture as below, highlighting the advantages of ML-TTA equipped with Bound Entropy Minimization (BEM) in multi-label scenarios. It is indicated that ML-TTA achieves the best performance across all benchmarks. Even though these methods employ innovative strategies such as class prototypes, optimal transport, and bias correction, their performance still does not significantly outperform CLIP in the multi-label scenarios.

• Our proposed Bound Entropy Minimization (BEM) explores the feasibility of the TTA paradigm in multi-label scenarios, and we hope that the innovation of ML-TTA will attract the attention of more researchers and inspire more excellent works in multi-label TTA. Once again, thank you for your valuable feedback, and we eagerly await your further guidance.

[1].A comprehensive survey on test-time adaptation under distribution shifts. IJCV 2024

[2].A comprehensive survey on source-free domain adaptation. TPAMI 2024

[3].In search of lost online test-time adaptation: A survey. IJCV 2024

[4].Beyond model adaptation at test time: A survey. arXiv 2024

[5].Dual Prototype Evolving for Test-Time Generalization of Vision-Language Models. NeurIPS 2024

[6].AWT: Transferring Vision-Language Models via Augmentation, Weighting, and Transportation. NeurIPS 2024

[7].Frustratingly Easy Test-Time Adaptation of Vision-Language Models. NeurIPS 2024

Dear reviewer LTFP, thanks for your previous suggestions for our work. We would like to further discuss the content with you and hope to receive your response to the manuscript.

Additionally, if you find that the overall quality of the manuscript has improved after re-evaluating these modifications, we kindly ask you to consider adjusting the rating score accordingly.

Looking forward to your feedback, thank you!

Thanks for your valuable suggestions, we will try to address your concerns and we are eager to engage in a more detailed discussion with you.

Weakness 1: More detailed motivation about ML-TTA.

• Test-time adaptation (TTA) refers to directly adapting test instances without the need to access original training data. However, existing TTA methods are based on entropy minimization and primarily focus on increasing the predicted confidence of the top-1 label. In multi-label scenarios, however, optimizing only the top-1 label may result in insufficient adaptation for other positive labels.

• To address this issue, we propose the Boundary Entropy Minimization (BEM) objective, which aims to simultaneously increase the confidence of multiple top-k labels, where k is determined by the retrieved paired caption. The core idea of BEM involves treating the weak label set of each augmented view and the corresponding strong label set of each caption as single-label, learning instance-level view and caption prompts to adapt to multi-label instances. By binding the top-k predicted labels, BEM mitigates the limitation of traditional entropy minimization and avoids over-optimizing the top-1 label.

Weakness 2: ML-TTA may introduce complexity in practice.

• ML-TTA consists of: view augmentation, caption retrieval, and label binding.

  • View augmentation is a widely adopted method in the TTA domain, which encourages the model to make consistent and confident predictions by minimizing the marginal entropy of predictions across multiple views.

  • For caption retrieval, ML-TTA pre-constructs an offline embedding base of text descriptions. Hence, in practice, only a single matrix multiplication is sufficient to retrieve the paired captions.

  • Label binding refers to making the logits of the top-k labels equal, as expressed by Eq.(6) in the manuscript: $ \tilde s_{ij}^{\mathbf x^{\text {test}}}=( ( m_i^{\mathbf x^{\text {test}}} - s_{ij}^{\mathbf x^{\text {test}}} ) + s_{ij}^{\mathbf x^{\text {test}}} ) \times \mathbb I( \mathrm {Rank}_{( s_{ij}^{\mathbf x^{\text {test}}}, \mathbf s_{i}^{\mathbf x^{\text {test}}})} \leq k^{\mathbf x_i^{\text {test}}} )+ s_{ij}^{\mathbf x^{\text {test}}} \times \mathbb I (\mathrm {Rank}_{(s_{ij}^{\mathbf{x}^{\text {test}}}, \mathbf s_i^{\mathbf x^{\text {test}}})} > k^{\mathbf x_i^{\text {test}}} )$. We can see that label binding involves some simple mathematical operations and stop-gradient operations $( ( m_i^{\mathbf x^{\text {test}}} - s_{ij}^{\mathbf x^{\text {test}}} ) + s_{ij}^{\mathbf x^{\text {test}}} )$, which only need a negligible time consumption during the model adaptation process.

• Furthermore, we conduct analysis of testing time per test instance on the MSCOCO-2014 dataset, comparing ML-TTA with others that also do not require retaining historical knowledge, as shown in the Table below:
  | Methods | TPT [1] | DiffTPT [2] | RLCF [3] | ML-TTA|
  | ------- | ------- | ------- | ------- | ------- |
  | Testing Time | $\mathbf {0.21}s$ | $0.41s$ | $0.45s$ | $\underline {0.24}s$ |
  | mAP | $48.52$ | $\underline {48.56}$ | $36.87$ | $\mathbf {51.58}$ |

• The result shows that compared to the benchmark TPT, ML-TTA exhibits an increase in testing time due to the simultaneous optimization of view and caption prompts. However, ML-TTA presents a significant advantage compared to DiffTPT, which involves generating multiple pseudo-images via a diffusion model, and RLCF, which requires distillation from a teacher model along with more gradient update steps.

Weakness 3: In real-world scenarios, captions may not always accurately represent the image content.

• Our work employs captions to determine the number of labels for views. Even if there is some deviation between captions and contents of views, the proposed BEM objective can effectively mitigate the limitation of traditional entropy minimization that only optimizes for the top-1 label.

• Indeed, in real-world application scenarios, the accuracy of retrieved paired captions may be affected by various factors. To address this, in the manuscript, we also adopt a confident-based filtering strategy, filtering out views and captions with high entropy ($i.e.$, low confidence) to reduce the impact of noise on the model's adaptation.

• Furthermore, we can explore more robust strategies to retrieve paired captions in future works, such as, constructing high-quality and content-rich text description databases, ensembling label sets from multiple captions, or improving the similarity retrieval strategy.

[1]. Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models. NeurIPS 2022

[2]. Diverse Data Augmentation with Diffusions for Effective Test-time Prompt Tuning. ICCV 2023

[3]. Test-Time Adaptation with CLIP Reward for Zero-Shot Generalization in Vision-Language Models. ICLR 2024

Dear Reviewer, we are looking forward to your professional suggestions and hope to receive your guidance to further discuss and refine the contents of the work. We are eager for your response. Thank you.

Thanks for your valuable suggestions, we will try to address your concerns and we are eager to engage in a more detailed discussion with you.

Weakness 1: Explanation of paired captions retrieval, label binding, and algorithmic details.

1.Paired captions retrieval.

• Given a test image $\mathbf x$, $\mathbf x$ is first augmented $N$ times to obtain a set of different views {$\mathbf x_i|i=1,2,3,...,N$}. The goal of paired caption retrieval is to retrieve the most similar caption for each view. Initially, we collect massive text descriptions following PVP [1]. Then, CLIP is used to extract text embeddings and construct an offline database of size $B\times d$, where $B$ denotes the number of test descriptions and $d$ denotes the embedding dimension.
• For a given augmented view $\mathbf x_i$, defined as a $d$-dimensional vector, we directly compute the similarity between $\mathbf x_i$ and all text embeddings in the database, resulting in a $B$-dimensional similarity vector. The text description corresponding to the highest similarity is considered as the retrieved paired caption for $\mathbf x_i$.

2.Label binding.

• Bound Entropy Minimization (BEM) aims to simultaneously increase the prediction confidence for the top-k labels, whereas the traditional entropy minimization can only enhance the confidence of the top-1 label. Proposition 2 in the manuscript states that the key point in BEM is to equalize the logits of the top-k labels, $i.e.$, label binding process. The Eq.(6) in the manuscript is as follows:

  $ \tilde s_{ij}^{\mathbf x^{\text {test}}}=( ( m_i^{\mathbf x^{\text {test}}} - s_{ij}^{\mathbf x^{\text {test}}} ) + s_{ij}^{\mathbf x^{\text {test}}} ) \times \mathbb I( \mathrm {Rank}_{( s_{ij}^{\mathbf x^{\text {test}}}, \mathbf s_{i}^{\mathbf x^{\text {test}}})} \leq k^{\mathbf x_i^{\text {test}}} )+ s_{ij}^{\mathbf x^{\text {test}}} \times \mathbb I (\mathrm {Rank}_{(s_{ij}^{\mathbf{x}^{\text {test}}}, \mathbf s_i^{\mathbf x^{\text {test}}})} > k^{\mathbf x_i^{\text {test}}} )$.

• We take a 3-class classification task with class labels of (1,2,3) as an example, assuming $k^{\mathbf x_i^{\text {test}}}$ is 2, and the label binding process is $\mathbf s = [\mathbf {0.9}, \mathbf {0.7}, 0.3] \rightarrow \mathbf s^{'} = [\mathbf {0.9}, \mathbf {0.9}, 0.3]$. $\tilde s_{ij}^{\mathbf x^{\text test}}$ represents the logit of the j-th class in the i-th augmented view after label binding, $e.g.$, $\tilde s_{i2}^{\mathbf x^{\text test}}$ changes from $\mathbf {0.7}\rightarrow\mathbf{0.9}$. $m_i^{\mathbf x^{\text {test}}}$ denotes the maximum value of $\mathbf s$, which is $\mathbf {0.9}$. $\mathbb I(\cdot)$ is the indicator function. $\mathrm {Rank}_{(a, \mathbf b)}$ indicates the descending rank of $a$ within b, $e.g.$, $\mathrm {Rank}_{(0.7, \mathbf s)} = 2$. The process for computing the bound logit for each class is as follows:

  $ \tilde s_{i1}^{\mathbf x^{\text {test}}}=( ( 0.9 - 0.9 ) + 0.9 ) \times \mathbb I( \mathrm {Rank}_{( 0.9, \mathbf s)} \leq 2)+ 0.9 \times \mathbb I (\mathrm {Rank}_{(0.9, \mathbf s)} > 2) = 0.9 \times \mathbb I( 1 \leq 2)+ 0.9 \times \mathbb I (1 > 2) = 0.9$

  $ \tilde s_{i2}^{\mathbf x^{\text {test}}}=( ( 0.9 - 0.7 ) + 0.7 ) \times \mathbb I( \mathrm {Rank}_{( 0.7, \mathbf s)} \leq 2)+ 0.7 \times \mathbb I (\mathrm {Rank}_{(0.7, \mathbf s)} > 2) = 0.9 \times \mathbb I( 2 \leq 2)+ 0.7 \times \mathbb I (2 > 2) = 0.9$

  $ \tilde s_{i3}^{\mathbf x^{\text {test}}}=( ( 0.9 - 0.3 ) + 0.3 ) \times \mathbb I( \mathrm {Rank}_{( 0.3, \mathbf s)} \leq 2)+ 0.3 \times \mathbb I (\mathrm {Rank}_{(0.3, \mathbf s)} > 2) = 0.9 \times \mathbb I( 3 \leq 2)+ 0.3 \times \mathbb I (3 > 2) = 0.3$

3.Algorithmic details.

• Algorithm 1 in the manuscript describes the process of label binding. Likewise, taking $\mathbf s = [\mathbf {0.9}, \mathbf {0.7}, 0.3] \rightarrow \mathbf s^{'} = [\mathbf {0.9}, \mathbf {0.9}, 0.3]$ with k being 2 as an example, since $\mathbf {0.9}$ and $\mathbf {0.7}$ are all within top-2, the logits of $\mathbf {0.9}$ and $\mathbf {0.7}$ are bound together $\rightarrow \mathbf {0.9}$ and $\mathbf {0.9}$. However, 0.3 is not in the top-2, so 0.3 will not be bound $\rightarrow 0.3$.

[1]. TAI++:Text as Image for Multi-Label Image Classification by Co-Learning Transferable Prompt. IJCAI 2024

Weakness 2: Discussion about the selection of baselines and underperformance reasons.

• Current mainstream Test-Time Adaptation (TTA) methods primarily adapt to multi-class instances by entropy minimization, with the core idea of increasing the prediction confidence of top-1 label. However, for multi-label instances, focusing solely on the top-1 label inevitably impairs the adaptation for other positive labels.

• To our knowledge, our work is the first to investigate the feasibility of traditional entropy minimization in the multi-label setting. Therefore, we select the SOTA methods in the TTA area for multi-class scenarios as our baselines, including methods that do not require retaining historical knowledge (TPT [1], DiffTPT [2], RCLF [3]) and those that do (DMN [4], TDA [5]).

  • For instance, DMN [4] introduces a dual-memory network that preserves historical knowledge from single-label instances, which intensifies the optimization bias towards the top-1 label when adapting to multi-label instances.

  • TDA [5] proposes a dynamic key-value cache that retains only a small number of high-quality labels as key-value pairs at each step. Similar to DMN [4], it faces challenges in adapting to multi-label instances due to the erroneous accumulation of historical knowledge.

  • DiffTPT [2] tends to neglect small object categories when generating multi-label pseudo-images, causing the model to focus more on optimizing prominent object categories.

  • RLCF [3] employs teacher model logit distillation and more adaptation steps, which also results in excessive optimization for the top-1 label, thereby damaging the adaptation performance for other positive labels.

[1]. Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models. NeurIPS 2022

[2]. Diverse Data Augmentation with Diffusions for Effective Test-time Prompt Tuning. ICCV 2023

[3]. Test-Time Adaptation with CLIP Reward for Zero-Shot Generalization in Vision-Language Models. ICLR 2024

[4]. Dual memory networks: A versatile adaptation approach for vision-language models. CVPR 2024

[5]. Efficient Test-Time Adaptation of Vision-Language Models. CVPR 2024

Dear Reviewer, we are looking forward to your professional suggestions and hope to receive your guidance to further discuss and refine the contents of the work. We are eager for your response. Thank you.

Thanks for your valuable suggestions, we will try to address your concerns and we are eager to engage in a more detailed discussion with you.

Weakness 1: Explanation of Eq.(6) and $\tilde s_{ij}^{\mathbf x^{\text {test}}}$. Recognize weak and strong label sets.

1.Explanation of label binding and $\tilde s_{ij}^{\mathbf x^{\text {test}}}$ in Eq.(6) in manuscript.

• Label binding refers to making the top-k predicted logits equal, as expressed below:
  $ \tilde s_{ij}^{\mathbf x^{\text {test}}}=( ( m_i^{\mathbf x^{\text {test}}} - s_{ij}^{\mathbf x^{\text {test}}} ) + s_{ij}^{\mathbf x^{\text {test}}} ) \times \mathbb I( \mathrm {Rank}_{( s_{ij}^{\mathbf x^{\text {test}}}, \mathbf s_{i}^{\mathbf x^{\text {test}}})} \leq k^{\mathbf x_i^{\text {test}}} )+ s_{ij}^{\mathbf x^{\text {test}}} \times \mathbb I (\mathrm {Rank}_{(s_{ij}^{\mathbf{x}^{\text {test}}}, \mathbf s_i^{\mathbf x^{\text {test}}})} > k^{\mathbf x_i^{\text {test}}} )$.

• Since label binding (making ... equal) is non-differentiable, we employ the stop-gradient operation in VQ-VAE [1] for backpropagation, $i.e.$ $( ( m_i^{\mathbf x^{\text {test}}} - s_{ij}^{\mathbf x^{\text {test}}} ) + s_{ij}^{\mathbf x^{\text {test}}} )$ to perform label binding. Taking a 3-class classification task as an example with class labels of (1,2,3), assuming $k^{\mathbf x_i^{\text {test}}}$ is 2, and the label binding process is $\mathbf s = [\mathbf {0.9}, \mathbf {0.7}, 0.3] \rightarrow \mathbf s^{'} = [\mathbf {0.9}, \mathbf {0.9}, 0.3]$. $\tilde s_{ij}^{\mathbf x^{\text test}}$ represents the logit of the j-th class in the i-th augmented view after label binding, $e.g.$, $\tilde s_{i2}^{\mathbf x^{\text test}}$ changes from $\mathbf {0.7}\rightarrow\mathbf{0.9}$. $m_i^{\mathbf x^{\text {test}}}$ denotes the maximum value of $\mathbf s$, which is $\mathbf {0.9}$. $\mathbb I(\cdot)$ is the indicator function. $\mathrm {Rank}_{(a, \mathbf b)}$ indicates the descending rank of $a$ within b, $e.g.$, $\mathrm {Rank}_{(0.7, \mathbf s)} = 2$. The process for computing the bound logit for each class is as follows:

  $ \tilde s_{i1}^{\mathbf x^{\text {test}}}=( ( 0.9 - 0.9 ) + 0.9 ) \times \mathbb I( \mathrm {Rank}_{( 0.9, \mathbf s)} \leq 2)+ 0.9 \times \mathbb I (\mathrm {Rank}_{(0.9, \mathbf s)} > 2) = 0.9 \times \mathbb I( 1 \leq 2)+ 0.9 \times \mathbb I (1 > 2) = 0.9$

  $ \tilde s_{i2}^{\mathbf x^{\text {test}}}=( ( 0.9 - 0.7 ) + 0.7 ) \times \mathbb I( \mathrm {Rank}_{( 0.7, \mathbf s)} \leq 2)+ 0.7 \times \mathbb I (\mathrm {Rank}_{(0.7, \mathbf s)} > 2) = 0.9 \times \mathbb I( 2 \leq 2)+ 0.7 \times \mathbb I (2 > 2) = 0.9$

  $ \tilde s_{i3}^{\mathbf x^{\text {test}}}=( ( 0.9 - 0.3 ) + 0.3 ) \times \mathbb I( \mathrm {Rank}_{( 0.3, \mathbf s)} \leq 2)+ 0.3 \times \mathbb I (\mathrm {Rank}_{(0.3, \mathbf s)} > 2) = 0.9 \times \mathbb I( 3 \leq 2)+ 0.3 \times \mathbb I (3 > 2) = 0.3$

  The logits after binding are $[\mathbf {0.9}, \mathbf {0.9}, 0.3]$, and we will introduce the process in detail in future version.

2.Recognize weak and strong label sets.

• Given a test image $\mathbf x$, $\mathbf x$ is first augmented $N$ times to obtain different views {$\mathbf x_i|i=1,2,3,...,N$}. Then, for each $\mathbf x_i$, a most similar caption is retrieved to form $N$ view-caption pairs, defined as {$\langle \mathbf x_i, \mathbf t_i\rangle|i=1,2,3,...,N$}.
• For example, given a pair $\langle \mathbf x_i, \mathbf t_i\rangle$, where $\mathbf t_i$ is "A black bicycle parked in front of a car". We follow the nouns filter strategy in PVP [2] and extract label set {bicycle, car} from $\mathbf t_i$. This label set serves as the strong label set for $\mathbf t_i$ and also as the weak label set for $\mathbf x_i$. The term "weak" is called because $\mathbf t_i$ may not include all the labels presented in $\mathbf x_i$, for example, the truth label set of $\mathbf x_i$ could be {bicycle, car, dog}.

[1]. Neural Discrete Representation Learning. NeurIPS 2017

[2]. TAI++:Text as Image for Multi-Label Image Classification by Co-Learning Transferable Prompt. IJCAI 2024

Weakness 2: The learnable parameters.

• The learnable parameters in ML-TTA are the view prompt and caption prompt shown in Figure 2 in the manuscript. The image and text encoders of CLIP are frozen.

Weakness 3: Motivation of the view prompt and caption prompt. Why are they useful?

• The goal of ML-TTA is to enable adaptation to the multi-label test instance with varying distributions during the testing stage. Prompt tuning adapts to new data by adjusting the input context of the CLIP, thus not distorting the original knowledge of the pretrained CLIP model. Therefore, we also adopt prompt tuning strategy, treating prompt tuning at test-time as a way to furnish customized context for individual test instances.

• Benefit from the aligned visual-language space of CLIP, the feature representations of images and texts share similar semantic information, therefore, the paired caption can be considered as "pseudo image" with accurate textual labels. This mitigates the potential limitation of weak label set, which may not fully capture the content of augmented views. Additionally, within the aligned space of CLIP, the model is to learn visual-related knowledge from text captions. Therefore, we adopt both view prompts and caption prompts to learn complementary information from views and captions jointly.

Dear Reviewer, we are looking forward to your professional suggestions and hope to receive your guidance to further discuss and refine the contents of the work. We are eager for your response. Thank you.

• We appreciate your feedback. Your understanding of the weak and strong label sets is right. They are the same in both quantity and content in ML-TTA. We differentiate them by their respective action scopes, hence the terms "weak" and "strong".

  • Weak Label Set: Represents the pseudo-true labels for each augmented view. Since the true labels for each view are inaccessible and cannot be directly obtained. Therefore, we retrieve the most similar caption for each view and extract the textual labels to form the weak label set for that view. These textual labels, acting as an approximation of the view's true labels, provide as accurate label information as possible to the view.

  • Strong Label Set: Represents the known true labels corresponding to each paired caption. Owing to the aligned visual-language space of CLIP, captions can be regarded as pseudo-images with known true labels. Therefore, the textual labels extracted from the caption are utilized directly as the true labels for the caption, which we refer to as the strong label set. These textual labels help the model capture visual-related knowledge from the caption and the aligned CLIP space.

• Although the weak and strong label sets are the same in quantity and content, they differ in their action scope. The weak label set is defined by approximating the true labels of each view, whereas the strong label set is derived directly from the corresponding paired caption. In addition, we employ a confidence filtering strategy to filter out views and captions with high entropy (low confidence), ensuring that the label sets more accurately reflect the true label information of the views and captions.

Dear reviewer jmBy, thanks for your previous suggestions for our work. We would like to further discuss the content with you and hope to receive your response to the manuscript.

Additionally, if you find that the overall quality of the manuscript has improved after re-evaluating these modifications, we kindly ask you to consider adjusting the rating score accordingly.

Looking forward to your feedback, thank you!