Parameter-Efficient Long-Tailed Recognition

We are grateful for the thoughtful reviews, and for commenting that the paper is well-written and the experiment results are strong. We will address the concerns below.

Concern #1 (Weakness #1): The technical novelty is a little incremental.

Response: Our work pioneers a systematical study on how to efficiently and effectively adapt pre-trained models for long-tailed tasks. Throughout our investigation, we have discovered many new insights, and we aim to disclose these fundamental findings and inspire future research.

1. Full Fine-tuning fails in long-tailed learning, and PEFT emerges as a superior alternative. This contradicts common sense on balanced datasets, where PEFT typically strives to match the performance of full fine-tuning [1-5].
2. We discover that most elaborately designed losses fail when applied to CLIP, except Logit-Adjusted loss (explained in Appendix B). This discovery motivates future research to design more suitable losses.
3. We point out the importance and advantages of classifier initialization, particularly for tail classes. This also inspires more advanced research in the future.
4. We first propose to employ TTE to remedy the inherent limitations of Vision Transformer. As a remedial measure, TTE is intuitive and effective (explained in Appendix C).

Moreover, the components we have introduced are both scalable and effective. We believe our findings and efforts can contribute to the long-tailed learning community.

Concern #2: (Weakness #2): Adopt ResNet as backbone.

Response: We have also considered this problem. However, there is no PEFT method specifically designed for ResNet, making it challenging to integrate our method with ResNet. Despite this constraint, we have explored some straightforward strategies, including 1) incorporating a scaling and shifting (SSF) [6] module after the backbone, and 2) fine-tuning solely the bias terms of ResNet. The results are presented below. In comparison to zero-shot CLIP and previous methods (detailed in Tables 1&2 in the main paper), our method achieves significantly superior performance with lower computational costs.

• ImageNet-LT

  	Backbone	Learnable Params.	#Epochs	Overall	Many	Medium	Few
  Zero-shot CLIP	ResNet-50	-	-	57.6	58.6	56.9	56.9
  PEL w/ SSF	ResNet-50	0.002M	10	66.9	72.0	66.5	54.1
  PEL w/ bias tuning	ResNet-50	0.034M	10	67.8	72.0	67.3	57.6
  PEL w/ bias tuning & SSF	ResNet-50	0.036M	10	68.3	72.5	67.8	58.2

• Places-LT

  	Backbone	Learnable Params.	#Epochs	Overall	Many	Medium	Few
  Zero-shot CLIP	ResNet-50	-	-	35.2	33.1	34.6	40.4
  PEL w/ SSF	ResNet-50	0.002M	10	46.7	47.5	48.7	40.7
  PEL w/ bias tuning	ResNet-50	0.034M	10	47.9	48.2	49.8	42.9
  PEL w/ bias tuning & SSF	ResNet-50	0.036M	10	48.1	48.1	50.0	43.9

[1] BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models.

[2] Visual Prompt Tuning.

[3] Parameter-Efficient Transfer Learning for NLP.

[4] LoRA: Low-rank adaptation of large language models

[5] AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition.

[6] Scaling & Shifting Your Features: A New Baseline for Efficient Model Tuning.

Concern #3 (Question #1): Why only use image cropping for test-time ensembling? Have you tried other data augmentation methods?

Response: We utilize image cropping tailored for Vision Transformer (ViT), which also uses image cropping to obtain $n\times n$ image patches. ViT carries the inherent risk of pattern separation. In response, we intuitively generate diverse cropped patches to remedy this issue. The detailed explanations are given in Appendix C. The results in Figure 12 demonstrate the effectiveness of our method: 1) When our cropped patches align with those of ViT, pattern separation persists, and the performance improvement is limited; 2) Conversely, when our cropped patches differ from ViT, the pattern separation can be alleviated, and the performance is enhanced.

Aside from TTE, we explore additional augmentation techniques such as TTE + Flipping and Random Augmentations [6] for multiple times. The results below demonstrate that TTE is more effective than other methods.

Concern #4 (Question #2): What kind of textual prompts are used for semantic-aware initialization?

Response: We employ the most widely used textual prompt, "a photo of a [CLASS]", where [CLASS] can be replaced by the class names. Additionally, we compare different prompting methods, including 1) the original class name ("[CLASS]") and 2) prompt ensembling [7] which applies different templates to class names. The experiments are conducted on ImageNet-LT. The results below show that these prompts have similar performance, and using "a photo of a [CLASS]" is adequate for generalization.

Moreover, Reviewer BCXb has proposed an intriguing situation, i.e., when the class names are OOD or unseen for CLIP. In response to this, we explore an alternative approach by incorporating class descriptions (which can be crafted manually or generated using large language models). The results show that using class descriptions can also enhance the performance compared to random initialization.

[6] Deep residual learning for image recognition.

[7] Learning Transferable Visual Models From Natural Language Supervision.

Concern #4 (Weakness #4): Compare the computational cost with PEL methods.

Response: Thank you for the suggestion. We compute the time cost per epoch for different parameter-efficient fine-tuning methods. The experiments are conducted using a single NVIDIA A100 GPU. The results indicate that the time cost for each PEFT method is very close.

Concern #5 (Weakness #5): Comparing various long-tailed learning strategies.

Response: We have compared different classifiers and losses in Tables 9 and 10 in Appendix A. Moreover, we give an analysis of Logit Adjusted loss in Appendix B. Please refer to the appendix for detailed explanations.

To enhance clarity, we will add a concise outline in the paper to summarize the experiments conducted in the appendix.

Concern #6 (Question #1): How is PEFT helping?

Response: As illustrated in Figure 4 in the main paper, the utilization of PEFT enhances the capability to extract more separable representations. This is because the PEFT module is integrated into the backbone, and enhances the feature extraction process of the pre-trained model. In this way, it is more effective than solely fine-tuning the classifier.

Moreover, PEFT freezes the pre-trained model while introducing a small number of learnable parameters for adaptation. This approach ensures the retention of the generalization ability of the pre-trained model. In contrast to full fine-tuning, PEFT can mitigate overfitting, as analyzed in Table 8 in Appendix A.

Concern #7 (Question #2): Why are the training epochs chosen as 10/20? How would the results differ if the training period were extended?

Response: We set the training epochs to 10 for ImageNet-LT, Places-LT, and CIFAR-100-LT. This is because PEL can converge rapidly by learning a small number of parameters, along with a classifier initialization.

For iNaturalist 2018, we train 20 epochs considering it has much more data. We have presented an analysis of different training epochs in Table 7 in Appendix A. The results show that extending the training epochs can result in improved performance. However, this improvement comes at the expense of increased computational overhead.

We also conduct experiments across different epochs on ImageNet-LT and Places-LT, and the results are presented below. Increasing the training epochs does not yield significant improvements. In practice, the selection of the optimal training epoch depends on the data scale and the task complexity. Generally, 10-20 epochs are appropriate for most cases.

• Results of PEL on ImageNet-LT by training different epochs.

  #Epochs	Overall	Many	Medium	Few
  5	77.5	80.9	77.2	69.0
  10	78.3	81.3	77.4	73.4
  20	78.3	81.8	77.1	72.8
  30	78.0	82.2	76.6	71.1

• Results of PEL on Places-LT by training different epochs.

  #Epochs	Overall	Many	Medium	Few
  5	51.3	51.7	53.0	46.8
  10	52.2	51.7	53.1	50.9
  20	51.8	51.3	52.3	51.6
  30	51.3	51.4	51.5	50.5

Concern #8 (Question #3): PEL incurs larger(4x) computation cost than LPT.

Response: We have discussed this in Paragraph Results on iNaturalist 2018 in Section 4.2: Although LPT uses fewer learnable parameters, we can reduce the parameters of PEL to reach a comparable quantity (i.e., reduce the bottleneck dimension r to 64, more details are given in Figure 6). In this case, PEL achieves an accuracy of 77.7% (without TTE) / 79.0% (with TTE), which still outperforms LPT. In fact, due to the large number of classes of iNaturalist 2018, the classifier already contains 6.25M parameters. Therefore, the parameter quantity of PEL does not lead to too much cost.

We appreciate the reviewer for the thorough comments, and for mentioning that the topic is important and the paper is well structured. Following we will respond to the concerns.

Concern #1 (Weakness #1): Include references and detailed explanations of existing studies concerning parameter-efficient fine-tuning approaches.

Response: This is a good suggestion. We will add more explanations as follows:

Parameter-efficient fine-tuning (PEFT) aims to adapt the pre-trained models to downstream tasks by learning a small number of parameters. Over the past few years, various PEFT approaches have been proposed for both computer vision (CV) and natural language processing (NLP) tasks [1]. These approaches can be broadly categorized into three groups [1]: 1) Prompt Tuning, such as VPT-shallow [3] and VPT-deep [3], is inspired by prompting strategies in NLP. It prepends learnable patches to images or sentences to serve as the contexts for downstream tasks; 2) Adapter Tuning, such as Adapter [4] and AdaptFormer [6], incorporates lightweight modules sequentially or parallelly to the pre-trained models; 3) Parameter Tuning, like BitFit [2] and LoRA [5], selectively modifies specific parts of the pre-trained models, such as the weight part or the bias part. Recent work [7] offers a unified view of different PEFT methods and proposes that they are intrinsically similar. Nevertheless, different PEFT modules may excel in distinct tasks due to variations in their design approaches and integration positions.

Thanks for the thoughtful comments. We will update the related works in the next version.

Concern #2 (Weakness #2): The technical contribution of this paper is not clear.

Response: We summarize the contribution of this work as follows.

PEL is a general framework, with all its components scalable and efficient. Its design allows updates in alignment with the development of the related field. What is more, we aim to reveal fundamental insights and inspire future research regarding long-tailed learning with pre-trained models.

1. Full Fine-tuning fails in long-tailed learning, and PEFT emerges as a superior alternative. This contradicts common sense on balanced datasets, where PEFT typically strives to match the performance of full fine-tuning [1-5].
2. We discover that most elaborately designed losses fail when applied to CLIP, except Logit-Adjusted loss (explained in Appendix B). This discovery motivates future research to design more suitable losses.
3. We point out the importance and advantages of classifier initialization, particularly for tail classes. This also inspires more advanced research in the future.
4. We first propose to employ TTE to remedy the inherent limitations of Vision Transformer. As a remedial measure, TTE is intuitive and effective (explained in Appendix C).

To the best of our knowledge, this is the first work that systematically studies the parameter-efficient adaptation of pre-trained models on long-tailed tasks. We believe our findings and efforts can contribute to the long-tailed learning community.

Concern #3 (Weakness #3): In Table 4, the authors seem only conduct experiments with existing PEL methods, the performance of the proposed method is missing.

Response: There may be some misunderstandings. PEL is a framework that can adopt any existing PEFT methods. Table 4 is an ablation study on different PEFT methods. The results show that PEL with AdaptFormer achieves superior performance currently. If some new PEFT methods are proposed in the future, PEL can seamlessly incorporate them to enhance its performance.

Moreover, we apologize for the confusion in Table 4. It should be clarified that zero-shot CLIP, full fine-tuning, and classifier fine-tuning are not PEFT methods. The remaining methods are PEFT methods that can be integrated into PEL. We will update Table 4 in the next version.

[1] Visual Tuning.

[2] BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models.

[3] Visual Prompt Tuning.

[4] Parameter-Efficient Transfer Learning for NLP.

[5] LoRA: Low-rank adaptation of large language models

[6] AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition.

[7] Towards a Unified View of Parameter-Efficient Transfer Learning.

Dear Reviewer,

We appreciate you for your further feedback. We would like to address your concerns below.

Concern: A major concern is the technical contribution, as also noted by Reviewer BCXb and Reviewer 4hL6. This concern isn't adequately addressed by merely outlining empirical findings and their broader implications.

Response: First, as we further responded to Reviewer BCXb, the main contributions of this paper lie in the proposed framework and the state-of-the-art performance it achieves. Our method achieves significant performance improvements, which can help take an important stride forward in the field of long-tailed recognition.

Second, the technical contribution does not necessarily mean proposing a completely new technology. Applying classic methods to specific fields, as well as proposing uncovering empirical findings that have not been previously explored, can also help advance the field. Numerous works are built upon this approach, and the following are just some typical examples:

• Kang et al. [1] explore the widely used two-stage learning framework to help long-tailed recognition.
• Yang et al. [2] study the effects of self-supervised learning and semi-supervised learning on long-tailed recognition.
• Zhong et al. [3] employ the popular mixup method and label smoothing to help long-tailed representation learning.
• Menon et al. [4] revisit the classic idea of logit adjustment based on label frequencies and use it to help long-tailed learning.
• Park et al. [5] chose the CutMix technology to improve the data augmentation in long-tailed recognition.
• Dong et al. [6] utilize visual prompt tuning, and previously proposed A-GCL loss to improve long-tailed recognition.

All of these works have not proposed new technologies. Instead, they proposed a framework that utilizes classical methods to enhance long-tailed recognition tasks. However, they all provide novel insights into the field and have inspired several works.

[1] Decoupling Representation and Classifier for Long-Tailed Recognition. ICLR 2020

[2] Rethinking the Value of Labels for Improving Class-Imbalanced Learning. NeurIPS 2020.

[3] Improving Calibration for Long-Tailed Recognition. CVPR 2021.

[4] Long-Tail Learning via Logit Adjustment. ICLR 2021.

[5] The Majority Can Help The Minority: Context-rich Minority Oversampling for Long-tailed Classification. CVPR 2022.

[6] LPT: Long-tailed Prompt Tuning for Image Classification. ICLR 2023.

Concern: The current version lacks an in-depth analysis and a clear understanding of how different PEFT methods benefit long-tailed learning. Furthermore, the main body of the experimental part (Table 1-3) seems inadequate as it appears to focus more on confirming SOTA performance, rather than providing a comprehensive evaluation of various PEFT methods.

Response: In fact, we have already conducted plenty of work to analyze these PEFT methods in our paper.

1. We have analyzed the impact of PEFT modules on the representation in Figure 4 in Section 4.3.
2. We have analyzed the effect of the PEFT module on different layers in Figure 8 in Appendix A.
3. We have analyzed the influence of the PEFT module on model convergence in Figures 9 and 10 in Appendix A.
4. We have provided model parameter quantity analysis in Appendix E, and demonstrate that these PEFT methods are at the same complexity level and are all lightweight.
5. We have compared different PEFT methods in Table 4 in Section 4.3, and find that most of them yield enhanced performance.

Moreover, we have updated our manuscript and added Figures 12 and 13 to give more analyses on the representation separability of different PEFT methods. The results show that all of these PEFT methods contribute to more distinctive features. Notably, Adapter and AdaptFormer are more effective in enhancing separability.

We would like to argue that, our paper is not a simple "empirical evaluation". We have presented comprehensive analyses of the effects of the employed PEFT modules in multiple aspects.

Concern: It seems PEL is sensitive to the training epochs. However, it remains unclear how to best determine the length of training epochs by data scale and the task complexity.

Response: It is necessary to mention that, previous works also need to elaborately choose a proper training epoch. For instance, their training epochs range from 90 to 400 on ImageNet-LT, and 30 to 100 on Places-LT (please refer to Tables 1 and 2 for more details). However, none of the previous works have explained why they chose a specific number of training epochs.

Moreover, the number of the training epochs is the only hyperparameter that needs to be manually set in our method. As we explained in the previous comments, training for 10-20 epochs is appropriate for PEL in most cases.

We are grateful for the valuable reviews, and for the positive comments that the experimental results are promising and the paper is well-written. We will address the concerns below.

Concern #1: Initialization for vision-only pre-trained models.

Response: When employing vision-only pre-trained models, utilizing class mean features is a viable choice. As delineated in Table 5 in the main paper, this approach yields significant performance improvements. It surpasses most existing methods by comparing the results in Tables 1&2.

Furthermore, we have tried linear probing with re-weighted (RW) or logit-adjusted (LA) loss and report the results below. The results are better than random initialization but are inferior to class mean features.

It remains an intriguing challenge how to initialize the classifier for visual-only models with limited data. Anyway, employing class mean features proves to be an intuitive and effective approach at present.

• ImageNet-LT

  	Overall	Many	Medium	Few
  Random initialization	76.1	80.8	75.9	63.2
  Linear probing w/ RW	77.1	81.8	76.4	66.6
  Linear probing w/ LA	77.2	81.3	76.4	68.4
  Class mean features	77.5	81.3	76.8	69.4

• Places-LT

  	Overall	Many	Medium	Few
  Random initialization	49.6	51.2	52.1	41.0
  Linear probing w/ RW	49.9	51.2	51.0	45.0
  Linear probing w/ LA	50.2	51.3	51.1	46.2
  Class mean features	51.3	51.6	52.1	48.8

Concern #3 (Question #1): What if the class is OOD/unseen for CLIP? Will such class prompt features still be a good initialization?

Response: We posit that CLIP has seen sufficient language corpus, considering its pre-training on web-scale datasets. However, it is noteworthy that CLIP may fail to recognize specific class names. This probably stems from its limited vocabulary size (CLIP contains approximately 49K of vocabulary) or encountering uncommon or novel concepts. In this case, the initialization may regress to random initialization.

One solution is to design some descriptions for this class, which can be crafted manually or generated using large language models. We follow [6] to generate the descriptions. Following are some examples:

We conduct experiments on ImageNet-LT by assuming that the class names are unseen. We solely use the descriptions for initialization. The results shown below demonstrate that the descriptions can enhance generalizations, particularly for tail classes.

Concern #4 (Question #2): Different loss functions are all about the tradeoff between tail and head class performance.

Response: The phenomenon you highlighted does indeed exist. In the context of long-tailed learning, there exists a competition for decision boundaries between head and tail classes. By optimizing re-balancing losses, the model tends to predict samples as tail classes, leading to a performance deterioration of head classes, which is called the seesaw dilemma.

Most recently, some works have proposed to address this issue through multi-objective optimization [7] [8]. Although the solution may not be fully mature, the topic remains intriguing and warrants exploration in the future.

[6] Visual classification via description from large language models.

[7] Long-Tailed Learning as Multi-Objective Optimization. In https://arxiv.org/abs/2310.20490.

[8] Pareto Deep Long-Tailed Recognition: A Conflict-Averse Solution. In https://openreview.net/forum?id=b66P1u0k15.

We appreciate the reviewer for the thoughtful reviews and for the encouraging comments such as nice benchmarking and strong motivation. Following are our responses to the concerns.

Concern #1 (Weakness #1): The novelty of PEL.

Response: Since it is the first work that systematically studies the parameter-efficient adaptation of pre-trained models on long-tailed tasks, we aim to reveal fundamental insights and inspire future research.

1. Full Fine-tuning fails in long-tailed learning, and PEFT emerges as a superior alternative. This contradicts common sense on balanced datasets, where PEFT typically strives to match the performance of full fine-tuning [1-5].
2. We discover that most elaborately designed losses fail when applied to CLIP, except Logit-Adjusted loss (explained in Appendix B). This discovery motivates future research to design more suitable losses.
3. We point out the importance and advantages of classifier initialization, particularly for tail classes. This also inspires more advanced research in the future.
4. We first propose to employ TTE to remedy the inherent limitations of Vision Transformer. As a remedial measure, TTE is intuitive and effective (explained in Appendix C).

Furthermore, PEL is a general framework, with all of the components efficient and scalable. We believe our findings and efforts can contribute to the long-tailed learning community.

[1] BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models.

[2] Visual Prompt Tuning.

[3] Parameter-Efficient Transfer Learning for NLP.

[4] LoRA: Low-rank adaptation of large language models

[5] AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition.

Concern #2 (Weakness #2): A systematic ablation of components.

Response: Thanks for your valuable comments. We ablate the components in PEL separately, including 1) the PEFT module, 2) the Logit-Adjusted (LA) loss, 3) Semantic-Aware Initialization (abbreviated as SAI), and 4) Test-Time Ensembling (TTE). The results are reported below.

The results demonstrate the effectiveness of each component. Specifically, 1) PEFT enhances performance on both head and tail classes; 2) without the LA loss, predictions tend to be biased to head classes; 3) Semantic-Aware Initialization (SAI) consistently improves the generalization, particularly on tail classes; 4) Test-Time Ensembling further boost the performance across both head and tail classes.

• Ablation study on ImageNet-LT. The baseline involves learning a classifier using CE loss.

  PEFT	LA	SAI	TTE	Overall	Many	Medium	Few
  				60.9	82.6	56.7	13.8
  $\checkmark$				68.9	84.7	66.3	33.7
  $\checkmark$	$\checkmark$			74.9	79.7	74.7	61.7
  $\checkmark$	$\checkmark$	$\checkmark$		77.0	80.2	76.1	71.5
  $\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	78.3	81.3	77.4	73.4

• Ablation study on Places-LT. The baseline involves learning a classifier using CE loss.

  PEFT	LA	SAI	TTE	Overall	Many	Medium	Few
  				34.3	53.6	30.5	7.5
  $\checkmark$				38.9	55.6	35.3	16.4
  $\checkmark$	$\checkmark$			48.7	50.6	50.8	40.5
  $\checkmark$	$\checkmark$	$\checkmark$		51.5	51.3	52.2	50.5
  $\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	52.2	51.7	53.1	50.9

Dear Reviewer,

We sincerely appreciate your follow-up comments. We would like to provide further clarification on the concerns below.

Concern: At first sight, this paper seems a "performance" one by stacking up different available techniques to obtain SOTA results, which raises novelty concerns. Now the authors clarify that the paper is more about empirical evaluations of possible techniques for the same goal of "parameter-efficient finetuning for long-tailed tasks''.

Response: There may be some misunderstandings. The primary objective of this paper is to reveal the limitations of conventional fine-tuning methods and to propose a new framework for long-tailed recognition. Therefore, the main contributions of this paper lie in the proposed framework and the state-of-the-art performance it achieves.

Moreover, it is essential to clarify that our work is not about empirical evaluations. The response to Concern#1 is not our motivations; instead, it summarizes the novel discoveries emerging from our works. These discoveries can be viewed as broader impacts, and we hope these discoveries can inspire future research.

Concern: The exploration of 4 independent techniques without good motivation of why they are selected (but not others) and/or enough analysis about the synergy between them.

Response: We understand your concern. However, we have to clarify that these 4 modules are all with sufficient motivation.

• Motivation for PEFT: We have discussed the limitation of 1) zero-shot CLIP, 2) full fine-tuning, 3) classifier fine-tuning in Section 3.2, as well as the limitation of 4) partial fine-tuning in Paragraph Partial fine-tuning vs. parameter-efficient fine-tuning in Appendix A. Therefore, the question arises: How to adapt the pre-trained model to downstream long-tailed tasks? Adopting Parameter-efficient Fine-tuning (PEFT) emerges as an intuitive, economical, and effective approach.

  Among various PEFT methods, we employ AdaptFormer due to its state-of-the-art performance. Additionally, we explain why it helps in Paragraph Effects of the AdaptFormer module on different layers in Appendix A.

• Motivation for LA loss: It is widely recognized that employing re-balancing loss is a common method to alleviate the long-tailed problem. Additionally, we have theoretically explained that LA loss is more generalizable across different backbone models. Furthermore, we conduct experiments in Paragraph Comparison of different losses in Appendix A to demonstrate its superiority over other elaborately designed losses.

• Motivation for Semantic-aware initialization: As we explained in Section 3.3, semantic-aware initialization is a simple approach to leverage the semantic relationships, which can effectively make up for the data scarcity issues on the tail classes. The results in Table 5 demonstrate its ability to enhance tail-class performance. We also demonstrate its effectiveness on fast convergence in Figure 5.

• Motivation for TTE: As we explained in Section 3.3 and Appendix C, TTE is specially designed to compensate for the limitation of Vision Transformer, and can enhance the generalization on both head and tail classes.

All these modules are specially designed for long-tailed learning with pre-trained models and they have been proven effective in enhancing generalization, particularly on tail classes.

Concern: The results seem to suggest that using class descriptions, despite being better than random initialization, is still inferior to using simple class names (lagging behind on the "few" group by a large margin).

Response: Thanks for your concern. Your understanding is correct since the class descriptions do not contain the class names at all. When employing these descriptions for zero-shot predictions, the performance is notably inferior compared to using class names, as illustrated in the table below.

• Zero-shot CLIP performance with class descriptions or class names

  	Overall	Many	Medium	Few
  class descriptions	25.8	28.1	24.7	23.3
  class names	68.3	69.2	67.6	67.7

Nevertheless, leveraging class descriptions remains a viable alternative approach when the class names are unachievable or OOD for CLIP. We propose this in response to your concerns "What if the class is OOD/unseen for CLIP?" The results demonstrate that using class descriptions can effectively mitigate such challenges.

Thanks for your detailed feedback and new ablations. Below are my follow-up comments.

Since it is the first work that systematically studies the parameter-efficient adaptation of pre-trained models on long-tailed tasks, we aim to reveal fundamental insights and inspire future research.

At first sight, this paper seems a "performance" one by stacking up different available techniques to obtain SOTA results, which raises novelty concerns. Now the authors clarify that the paper is more about empirical evaluations of possible techniques for the same goal of ``parameter-efficient finetuning for long-tailed tasks''. Then it seems to me that the paper needs a major rewriting. More focus should be put on the comparisons and findings/insights from different threads of investigations, rather than on SOTA performance or any technical contributions. At least the authors need to tone down on the performance and technical contribution sides. The paper in its current form (e.g., title, abstract, the way it describes PEL, and the whole message) is misleading, and it seems to need a lot of time to be publication ready, which is concerning. Another lingering concern is the exploration of 4 independent techniques without good motivation of why they are selected (but not others) and/or enough analysis about the synergy between them.

The results shown below demonstrate that the descriptions can enhance generalizations, particularly for tail classes.

The results seem to suggest that using class descriptions, despite being better than random initialization, is still inferior to using simple class names (lagging behind on the "few" group by a large margin). Am I interpreting the results right?