Harnessing Shallow Features in Pre-Trained Models for Out-of-Distribution Detection

[W4] Why chose to fine-tune a network pre-trained on ImageNet-21k?

[A4] The main reason is that our downstream tasks are ImageNet-1k and CIFAR-100. We aim to avoid overfitting of the model to the downstream datasets, as an overly fitted model would render the tasks overly simplistic and might not provide a comprehensive and accurate assessment of the method's performance and generalization ability. To ensure a fair comparison, we have chosen to follow the practices of existing works such as NeCo and ViM. Moreover, to further guarantee that the pre-training dataset and the downstream task datasets do not have excessive overlap, we have also conducted experiments using CLIP. This multi-faceted approach allows us to better evaluate the effectiveness and robustness of our proposed method across different datasets and pre-training setups, and we believe it contributes to a more reliable and meaningful research outcome.

Dear Reviewer cJWd, We are extremely grateful for the comprehensive and detailed review you have provided. Your insights and suggestions are invaluable in helping us improve our work.

[W1] The title appears overstated.

[A1] We are truly and sincerely grateful for your astute and perceptive advice regarding the title of our work. After meticulous and in-depth consideration, we have come to recognize the validity of your observation. Indeed, a more precise and accurate title is essential to aptly reflect the core scope and nature of our proposed method. We wholeheartedly concur with your suggestion and will accordingly modify the title to "Harnessing Shallow Features in Pre-Trained Transfomers for Out-of-Distribution Detection". This revised title, we believe, will not only enhance the clarity and specificity of our research's representation but also align more closely with the technical details and contributions of our work. Once again, we express our profound appreciation for your invaluable feedback, which has significantly contributed to the refinement and improvement of our manuscript.

[W2] An explanation for "measuring the degree of neural collapse of each layer".

[A2] We are extremely grateful for your incisive query regarding the mention of "measuring the degree of neural collapse of each layer" and its relation to determining the importance weights. Our fundamental motivation was grounded in the convergence behavior of the last layer. In machine learning, the typical aspiration is for the class center, classifier, and samples to converge. However, this poses a detriment to OOD detection as OOD samples may also converge to the same class-center, and this is equally true for OOD detection scores based on Mahalanobis distance. To address this, we aimed to introduce less convergent information into the sample features, thereby enabling them to escape the neural collapse condition, while still retaining a certain degree of convergent information. Through comprehensive statistical analysis and comparisons, we devised a method to dissolve the weights of different layers. We posit that the weights we currently employ can, to a certain extent, represent the convergence degree of each layer. For a more detailed understanding, please refer to the definition of alpha in Equation 5. To illustrate, if all samples within the same layer achieve optimal convergence, known as neural collapse, the variance of feature samples in that layer should be zero, meaning that the covariance matrix is zero except along the diagonal. Hence, moving towards neural collapse requires minimizing the off-diagonal elements while maximizing the diagonal elements, or the trace. Consequently, we assert that the definition of alpha provided by Eq5 effectively quantifies the extent of neurocollapse. We also acknowledge that the expression in lines 103 - 104 was suboptimal and lacked clarity. We will undertake a thorough revision of this description in the subsequent iteration to enhance the overall readability and comprehensibility of the paper.

[W3] To assess effectiveness.

[A3] We greatly appreciate your valuable feedback. We have seriously considered how to more suitably assess the proposed OOD score. We've carefully analyzed the dataset in OpenOOD v1.5. In terms of using CIFAR-10 as the in-distribution (ID) dataset, we found it to be relatively straightforward. As you might be aware, most methods we investigated achieve nearly perfect results, with the AUROC almost reaching 100 and the FPR95 being 0.00 when utilizing the ViT model. Currently, our CIFAR100 test set diverges only from Mnist. Therefore, for the sake of simplicity and also due to the time constraints during the rebuttal phase, we've opted to repeat the experiment with Imagenet-1k as the ID dataset in the OpenOOD v1.5 benchmark. Furthermore, we chose Imagenet-1K-LT as an ID as well since earlier experiments showed it is quite similar to Imagenet-1k but saves time. The outcomes of these experiments are shown in the following table. We sincerely hope that this strategy provides a more meaningful and challenging benchmark to evaluate our proposed OOD score while acknowledging practical constraints such as limited time.

Results for ImageNet-1k-LT in-distribution dataset on OpenOOD v1.5

We are highly appreciative of your valuable suggestion regarding the comparison with other training-data-dependent OOD scores. We have indeed made efforts in this regard. As you can see, we have presented the detailed results of Residual and ViM within our experimental results section. Additionally, to further address your concern and provide a more comprehensive evaluation, we have also includ the KL-matching results in the table below for direct comparison. This will enable a more in-depth analysis and a better understanding of the performance and characteristics of our proposed method in relation to these benchmarks. We hope that this additional information will enhance the overall quality and persuasiveness of our research. The results of ImageNet-21k pre-trained ViT are displayed in the top half of each table, while the results on CLIP are displayed in the bottom half.

Results for CIFAR-100 in-distribution dataset

Results for CIFAR-100-LT in-distribution dataset

[W6] May not be fair to compare with MCM.

[A6] The MCM method is naturally better suited for zero-shot OOD tasks compared to fine-tuning tasks. The prevalent fine-tuning approach, which mainly targets the visual encoder, tends to disrupt the initial alignment between the visual and text components after fine-tuning, resulting in less effective outcomes. Our goal in including the MCM method in our experiment was not to make a direct comparison but to empirically showcase that our proposed method enhances OOD detection performance. Conversely, methods like ViM and NECO are methodologically and conceptually more similar to our approach and, therefore, require a more thorough comparison. We will elaborate on this in the revised manuscript to provide readers with a clearer understanding of the comparative framework and the uniqueness of our contributions. Moreover, we present the results of MCM on the fine-tuned model (i.e. MCM-tuned) in the table below for comparison.

Results for CIFAR-100 in-distribution dataset

Results for CIFAR-100-LT in-distribution dataset

Results for ImageNet-1k-LT in-distribution dataset

[W5] Whether effectiveness relies on the pre-trained models and test with smaller pre-trained models.

[A5] Thank you so much for raising this important point regarding whether the effectiveness of our method relies on the pre-trained models and suggesting to test with smaller pre-trained models. Firstly, we wish to convey that our Out-of-Distribution (OOD) score's effectiveness is not contingent upon pre-trained models. We employ pre-trained models primarily due to their practical benefits. It's significantly simpler and requires fewer resources to fine-tune existing pre-trained models rather than developing a new model from scratch. Moreover, many contemporary state-of-the-art methods also rely on fine-tuning pre-trained models. This prevalent approach not only streamlines the experimental process but also facilitates more straightforward and fair comparisons across different approaches, which is essential for an objective assessment of our method's performance and contributions. Additionally, in light of your insightful suggestion, we have indeed verified the validity of our OOD score with smaller pre-trained models. As depicted in the table below, we have included models like vit_tiny_patch16_224 and vit_small_patch16_224, shown in the upper and lower sections of each table. The outcomes from these smaller models provide further confirmation that our OOD score remains robust and effective across various model scales, thereby enhancing the generalizability and reliability of our proposed approach.

Results for CIFAR-100 in-distribution dataset

Results for ImageNet-1k-LT in-distribution dataset

Results for CIFAR-100-LT in-distribution dataset

Dear Reviewer cJWd,

Thank you for your follow-up questions! We address your concerns below.

[Q1] Performance comparision with ViM based on pre-trained checkpoints as the backbone.

[A1] Thanks for the suggestion! We compare the performance of our method with ViM in the table below. The results are based on Imagenet-1K-LT as the ID dataset. Given that running experiments on the ImageNet-1k dataset is time-consuming, we'll update the results as promptly as possible once the testing phase finishes.

From the results, we can see that our approach still outperforms ViM by over 3% w.r.t. FPR95 averaged over 5 OOD datasets, also showing its effectiveness.

Table 1: Performance comparison between ViM and our approach using ImageNet-1k-LT in-distribution dataset. In this experiment, we utilize the pre-trained model as the backbone without fine-tuning.

[Q2] Clearification on the results in Table 7 of the paper

[A2] Sorry for the confusion! Table 7 in the paper reports results for both pre-trained models (denoted by zero-shot) and fine-tuned models (denoted by long-tailed or balanced). We will revise the description to improve the clearity.

As mentioned in the previous response, we present the comparison results for ViM and SFM (ours) on the ImageNet-1k dataset when utilizing the pre-trained checkpoints as the backbone. From the results, we draw a similar conclusion that our proposed SFM achieves superior performance than ViM, showing its robustness to different backbones.

Results for ViM and SFM on ImageNet-1k in-distribution dataset based on pre-trained model without fine-tuning.

Dear Reviewer 5FtG,

We are truly and deeply grateful for the substantial amount of time and painstaking effort you have committed to meticulously reviewing our paper and furnishing such highly detailed and exceptionally valuable feedback.

[W1] Emphasis on the novel multiple-feature fuse method might be more suitable.

[A1] We truly value your valuable feedback and suggestions regarding the focus of our paper. We will modify our statement in the contribution section to "proposed a new adaptive feature fusion strategy". We think better showcases the unique characteristics and advantages of our research and appreciate your help in improving and strengthening our manuscript.

[W2,W3] Innovation declaration and the comparison results with some simple fuse method.

[A2,A3] We appreciate your suggestions greatly. After carefully reviewing [ref1-4], we agree that these articles are pioneering in the use of multi-layer feature fusion. Consequently, contingent upon the acceptance of our paper, we shall ensure an accurate citation and comprehensive analysis of these works within the related work section. Regarding the novelty of our paper, we wish to clarify that we have proposed a novel adaptive feature fusion algorithm applicable to OOD detection. Unlike [ref1-3], which modify network architecture by utilizing residuals or attention mechanisms to combine multi-layer features, [ref4] uses different enhancements of the same sample. Our approach of adaptive feature fusion distinctly differs from these methodologies. To thoroughly investigate our algorithm's application in OOD detection, we have also examined other weight configurations, such as uniform or fixed ratio weights, and conducted ablation experiments detailed in the following table. Due to constraints on rebuttal time, we assessed CIFAR-100, CIFAR-100-LT, and ImageNet-1k-LT utilizing the pre-trained ImageNet21k model alongside CLIP. The findings demonstrate that the SFM approach consistently achieves either the top (bold) or the second-best (italic) position across all contexts. Remarkably, when ranked second, our method's performance closely rivals that of the first-place finisher. Specifically, our approach attains 96.17 and 17.82, whereas W_0.5 achieves 95.82 and 19.04, w.r.t. AUROC and FPR95 respectively, which shows the effectiveness of our SFM approach.

Table 1: Results on CIFAR-100 in-distribution dataset based on ImageNet-21k pre-trained ViT and CLIP, respectively.

Table 2: Results on CIFAR-100-LT in-distribution dataset based on ImageNet-21k pre-trained ViT and CLIP, respectively.

[W4] Open questions about the task motivation.

[A4] Regarding the initial point about whether detecting every out-of-distribution (OOD) sample during model training is always advantageous, the answer isn't straightforward. Many semi-supervised or label noise learning techniques recommend using the most accurate data possible during training. However, there are instances where including a small number of errors or OOD samples can be beneficial to the learning process. This inclusion might mimic the effects seen in self-supervised or contrastive learning approaches. Hence, we assert that identifying all OOD samples may not always be the best strategy.

For the second part of your inquiry, this paper primarily focuses on semantic OOD. It proposes that in-distribution (ID) and OOD classes are distinct, and training is conducted solely with ID data, yet OOD detection is necessary during testing. As pertains to your question, a covariate shift is expected in OOD detection. In such cases, cross-domain OOD becomes relevant during training, particularly when cross-domain test samples are involved. This is due to its potential to improve the model's ability to generalize and withstand varied data distributions.

We hope this provides a clearer and more satisfactory explanation. We appreciate your insightful feedback once more.

[W5] Minor: the bold font

[A4] We're sincerely thankful for your careful review and the suggestion about bold font. We'll use it appropriately in the table to enhance result readability. Appreciate your help in improving our work.

Table 3: Results on ImageNet-1k-LT in-distribution dataset based on ImageNet-21k pre-trained ViT and CLIP, respectively.

Dear Reviewer GG5e,

We sincerely appreciate the time and effort you have dedicated to reviewing our paper. We address your concerns below.

[W1] Experimental results for TRUSTED.

[A1] Thank you for your reminder. We previously overlooked including the comparison results with the Trusted method, but these have now been added to the revised manuscript in Table 1, 2, 3, 4. For your convenience, we have summarized the comparison between the Trusted method and our approach on CIFAR-100 dataset in tables below. As shown in the results, our method consistently outperforms Trusted.

Table 1: Results on CIFAR-100 in-distribution dataset for Trusted and SFM. The top two rows of the table present results based on the ImageNet-21k pre-trained ViT, while the bottom two rows show results based on CLIP.

Table 2: Results on CIFAR-100-LT in-distribution dataset for Trusted and SFM. The top two rows of the table present results based on the ImageNet-21k pre-trained ViT, while the bottom two rows show results based on CLIP.

[W2,Q1] Contribution of layerwise feature weighting in OOD detection.

[A2] Thank you very much for your valuable suggestion! We agree with the reviewer that the penultimate layer plays a crucial role in OOD detection. To further investigate this, we conducted comparative experiments by assigning different weights to the penultimate layer. For example, in the case of W_0.5, the penultimate layer was given a weight of 0.5, while the remaining layers equally shared the remaining 0.5 weight.

The results are summarized in the tables below. From these results, it is evident that our method achieves superior performance on most datasets compared to assigning fixed weights across all layers. Notably, the improvement is particularly significant on the ImageNet dataset.

Thank you again for your insightful comment, which helped us further validate the effectiveness of our approach.

Table 1: Results on CIFAR-100 in-distribution dataset based on ImageNet-21k pre-trained ViT and CLIP, respectively.

Table 2: Results on CIFAR-100-LT in-distribution dataset based on ImageNet-21k pre-trained ViT and CLIP, respectively.

Table 3: Results on ImageNet-1k-LT in-distribution dataset based on ImageNet-21k pre-trained ViT and CLIP, respectively.

Results for ImageNet-1k in-distribution dataset on Trusted and SFM

Results for ImageNet-1k-LT in-distribution dataset on Trusted and SFM

Dear Reviewer GG5e,

We sincerely thank you for taking the time to review our paper and providing us with your valuable feedback. Your insights are greatly appreciated and have been invaluable in improving our work.

May we kindly ask if our responses have sufficiently addressed your concerns? If there are any remaining questions or issues, we are fully committed to addressing them.

Once again, we deeply appreciate your time and effort in reviewing our work.

Best regards,

The Authors

We sincerely thank each reviewer for their valuable comments. We have addressed them in our responses and updated the manuscript accordingly. If the reviewers have any further questions, we are always ready to provide additional clarifications.

The Authors