Projection Regret: Reducing Background Bias for Novelty Detection via Diffusion Models

Summary

Recent methods have mainly utilized the reconstruction property of in-distribution samples to detect OOD by diffusion models. However, they often suffer from detecting OOD samples that share similar background information to the in-distribution data. Based on the observation that diffusion models can project any sample to an in-distribution sample with similar background information, the paper proposes Projection Regret (PR), an efficient novelty detection method that mitigates the bias of non-semantic information. To be specific, PR computes the perceptual distance between the test image and its recursive diffusion-based projection to detect the abnormality.

Strengths

1. The paper is written well and is easy to understand. 2. The studied problem is very important. 3. The results seem to outperform state-of-the-art.

Weaknesses

1. From the introduction, the projection cannot change the background information too much, which seems to be harmful on near-OOD detection, such as Cifar10 vs Cifar100. Since c100 and c10 has a similar background, then wouldn't this projection create a barrier further for solving the OOD detection problem? 2. The design and hyperparameter search (including the ensemble) seems to be very important on the results. How do the authors get the best configuration on a new ID/OOD pair? Is there any sensitivity analysis on more pairs? i.e., doe the current Figure 3 generalize to most ID/OOD pairs? 3. The computational budget is a little bit concerning in the algorithm. Is there any concrete comparison or explanation? 4. More results on large-scale benchmarks (i.e., ImageNet) and large (diffusion) models seem to be more meaningful for practical concerns and interpretability.

Questions

see above

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

3 good

Contribution

2 fair

Limitations

yes

Summary

This paper discusses a method in machine learning known as Novelty Detection, which is used to identify abnormal or out-of-distribution (OOD) samples. The authors suggest that diffusion models, a popular generative framework due to their strong generation performance, has recently become an attractive tool for novelty detection. However, the authors note a problem: while these models excel at generating high-quality results, they often struggle to detect OOD images with similar backgrounds. This issue, referred to as 'background bias', can lead to inaccurate novelty detection. To address this issue, the paper proposes 'Projection Regret' (PR), a method which reduces the impact of background bias and improves the accuracy of novelty detection. PR calculates the perceptual distance between a test image and its 'projection' (an in-distribution sample with similar background information created by the diffusion model). To further mitigate the effect of dominant background information, PR uses recursive projections to cancel out the background bias. The paper also introduces an ensemble of multiple projections for improved detection performance, calculated efficiently via a consistency model. A new perceptual distance metric using underlying features of the diffusion model is proposed and compared to other metrics, showing promising results. In conclusion, the paper's main contributions are: 1. Identifying the issue of background bias in novelty detection via diffusion models. 2. Proposing a solution, Projection Regret (PR), which mitigates this bias and enhances OOD detection. 3. Introducing an alternative perceptual distance metric computed from decoder features of the pre-trained diffusion model. Extensive experiments demonstrate that PR significantly outperforms previous diffusion-based novelty detection methods, showing potential for future applications.

Strengths

**Originality:** The paper presents a novel method called Projection Regret (PR) to address the problem of background bias in novelty detection using diffusion models. This is a creative combination of existing ideas, especially the use of perceptual distance and recursive projections to cancel out dominant background information. The proposed alternative perceptual distance metric using decoder features of the pre-trained diffusion model also adds to the originality. **Quality:** The research appears to be of high quality. The authors have carried out extensive experiments to demonstrate the effectiveness of their proposed method. They have also compared PR with other existing methods for novelty detection, showing that it outperforms them by a significant margin. **Clarity:** The paper is well written and structured. The authors clearly outline the problem of background bias in novelty detection, explain the concept behind PR, and provide detailed explanations of their experimental setup and results. They also do a good job of explaining complex concepts in an understandable way. **Significance:** The significance of this work lies in its potential to improve the accuracy of novelty detection in machine learning, which has broad applications in many fields, including medical diagnosis, autonomous driving, and forecasting. By addressing the issue of background bias, the proposed PR method could enhance the reliability and safety of deep learning applications. Additionally, the proposed perceptual distance metric could serve as a useful tool for researchers working on similar problems in the future.

Weaknesses

While the paper presents a significant contribution to the field, there are a few areas where it could be improved: **1. Evaluation metrics:** While the paper does an excellent job of comparing with existing methods and demonstrating the superiority of PR, it might benefit from more diverse evaluation metrics. At present, the paper primarily focuses on detection accuracy (AUROC). Incorporating additional measures such as precision-recall curves or F1 scores could provide a more comprehensive evaluation. **2. Real-world applications:** The paper demonstrates PR's effectiveness using standard datasets (CIFAR-10 vs CIFAR-100 etc.), but it would be valuable to see how PR performs in real-world scenarios. It is crucial to know how the method copes with the complexity and variability found in actual application scenarios including medical imaging or self-driving car data. **3. Deeper exploration of background bias:** Although the authors have proposed an innovative solution to tackle the issue of background bias, they could delve deeper into this problem. Understanding the nature of this bias, its origins, and why it is particularly problematic for diffusion models can further enhance the paper's impact. **4. Computational efficiency:** While the authors mention that PR can be calculated efficiently via a consistency model, specific data about computational requirements, time complexity, and scalability is missing. Providing these details can help readers assess whether PR is suitable for their specific use-cases, especially those that require real-time processing or deal with large-scale datasets. **5. Robustness analysis:** The paper could benefit from a detailed robustness analysis of PR against various types of noise and distortions. This would allow potential users to understand the limits of the approach and potential pitfalls in practical applications. By addressing these points, the paper could strengthen its contributions and appeal to a broader audience within the machine learning community.

Questions

**Questions:** 1. Could you provide more details on the computational requirements of PR? For instance, what is the time complexity and how does it scale with the size of the dataset? 2. Would PR be equally effective in real-world applications where data might be more complex or noisy compared to standard datasets like CIFAR-10 or CIFAR-100? 3. Could you delve deeper into the issue of background bias? Understanding its origins and why it affects diffusion models specifically could strengthen the paper. 4. Have you investigated the robustness of PR against various types of noise and distortions? If not, do you anticipate that the method would be robust against such challenges? **Suggestions:** 1. Consider incorporating additional evaluation metrics such as precision-recall curves or F1 scores for a more comprehensive evaluation of PR's performance. 2. It would be beneficial to demonstrate how PR performs with real-world datasets, such as medical imaging data or self-driving car sensory data. This would help readers understand its applicability and effectiveness in practical scenarios. 3. Detailed information about computational efficiency, including time complexity and scalability aspects, would be useful for potential users assessing suitability for their specific use-cases. 4. Conducting a robustness analysis against various types of noise and distortions would further validate the efficacy of PR under different challenging conditions.

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The authors have addressed some limitations of their work, particularly the issue of background bias in diffusion models and how their proposed method, Projection Regret (PR), mitigates this. However, a few additional limitations could be further discussed: 1. **Computational Resources:** While the authors mention that PR can be calculated efficiently, the paper does not provide specific details on computational requirements or scalability. As advanced machine learning models often require significant computational resources, an evaluation of this aspect would be helpful for potential users. 2. **Robustness:** The paper does not explicitly discuss the robustness of PR to noise or distortions in the data. Given that real-world data can often be noisy or imperfect, discussing the limitations of PR in handling such situations would be beneficial. Regarding the broader societal impacts, the paper does not directly address this point. While novelty detection in machine learning can have multiple positive societal impacts, like enhancing medical diagnosis or improving safety in autonomous driving, it might also have potential negative impacts: 1. **Privacy concerns:** Improving the ability of machines to identify novel information could potentially lead to privacy concerns, as more accurate models might be misused for surveillance or unauthorized data collection. 2. **System Misuse:** In high-stakes applications, an over-reliance on automated novelty detection systems that might still make mistakes could lead to serious consequences. The authors could improve their paper by addressing these points, possibly in a new section dedicated to limitations and broader societal impacts. This would help readers appreciate the full context of the research, including its potential downsides.

Summary

This paper presents a novel approach to detecting and handling novelty in data using a generative diffusion model, with a specific focus on addressing biased backgrounds. The proposed method aims to transform noisy samples into perfect ones by leveraging the capabilities of the diffusion model. The central idea is that by reversing the effects of noise on an inlier sample (x+noise), the resulting output will closely resemble the original sample (X). Conversely, if X were to be treated as an out-of-distribution (OOD) sample, the diffusion model trained on similar OOD backgrounds could project the samples onto a background resembling the inlier data. Unlike traditional methods, such as auto-encoders that rely on reconstruction error, this approach emphasizes the significance of high and meaningful reconstruction errors for anomaly detection. Since the primary focus is on the foreground, which constitutes a small portion of the images, the reconstruction error becomes a more reliable indicator of anomalies. To validate the effectiveness of the proposed method, the authors conducted evaluations on various datasets, including CIFAR-10, CIFAR-100, SVHN, and ImageNet. Comparative analyses were performed against existing generative models commonly employed for novelty detection.

Strengths

Recently, there has been growing concern that existing methods for image classification, which are generally not specific to one-class classification or novelty detection tasks, face significant challenges. Additionally, studies in the field have highlighted the problem of bias in the background for novelty detection. In response, this paper addresses this ongoing challenge by proposing a solution. The paper leverages the reverse step of the diffusion model to restore noisy outliers, with the goal of aligning the background of these recovered samples more closely with the inlier samples. This approach is intriguing and tackles a specific aspect of the problem. The results demonstrate the effectiveness of the proposed idea.

Weaknesses

One weakness of this paper is that the main idea of using projection and reconstruction error for novelty detection is not novel. There have been previous papers, such as "Adversarially One-Class Classifier for Novelty Detection" (CVPR 2018), that have presented similar ideas but with different approaches. Thus, the high-level novelty of the proposed method is questionable. However, the method is novel. Furthermore, the paper claims the method is less biased towards the background. However, the datasets used for evaluation, such as CIFAR and SVHN, are not known for having extreme biases toward the background. To address the bias problem effectively, evaluating the method on datasets that exhibit strong biases in the background would be more appropriate, such as "Hard ImageNet: Segmentations for Objects with Strong Spurious Cues." Other papers specifically tackle the issue of spurious correlations and out-of-distribution detection, such as "On The Impact Of Spurious Correlation For Out-of-distribution Detection" by Yifei Ming (2021). It would be valuable to compare the results of the proposed method with those papers that specifically address the bias problem rather than just comparing against other generative models. Moreover, the proposed method may fail in cases where the OOD samples have backgrounds that are very similar to the inliers but differ in the main concept or foreground. In such scenarios, the background would be ideally recovered using the proposed method, leading to false positives where OOD samples are detected as inliers. At the moment there are a lot of methods that try to handle the same problem to the reconstruction error can be a reliable score for novelty detection for example, see the Mem-autoencoder paper for novelty or anomaly detection. Addressing these weaknesses and providing a more thorough comparison with relevant papers and datasets would strengthen the overall contribution of the proposed method.

Questions

The idea and problem mentioned in this paper are interesting. However, the authors need to consider the following points: (1) They should review documents in the field of fairness for anomaly detection and those that attempt to mitigate bias. (2) There are benchmarks for evaluating methods biased against the background, such as Hard ImageNet. While these datasets may not be explicitly designed for out-of-distribution (OOD) detection, evaluating the proposed method on such data would be beneficial. Although CIFAR-10 and CIFAR-100 not have meaningful background concepts, additional evaluations on datasets with more pronounced background biases would provide valuable insights. (3) The comparison with existing methods needs improvement."

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

N/A

Summary

This paper proposes Projection Regret (PR) to mitigate the bias of background information for novelty detection. As an effective perceptual distance, it is able to detect abnormality by reducing the effect of dominant background using recursive projections. Experimental results show the effectiveness of the proposed novelty detection framework.

Strengths

1. Good motivation to explain the background bias problem. 2. well written and easy to follow this paper. 3. The effectiveness of Projection Regret.

Weaknesses

My main concerns lie in the following two aspects: (1) About the detection running time. Diffusion models are used in the proposed method, which leads to the increase of inference time when detecting the abnormality of test images. (2) It lacks experimental comparison on other practical industry product detection tasks, such as MVTecAD and BTAD benchmark, which seem to be more challenging to evaluate the effectiveness of the proposed method.

Questions

Refer to the Weaknesses.

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

upgrade my evaluation.

Thank you for your response. While many of my concerns have been addressed by the authors, there seems to be a lingering issue regarding (1) the novelty of the approach. I concur with the authors' assertion that their method diverges from a simple reconstruction-based model, instead employing a more refined approach with superior techniques. Additionally, (2) the computational costs associated with this method are notably higher than their predecessors. However, I still find the concern regarding novelty valid. Although the method is functional (and the perception of novelty can be pretty subjective), I have decided to revise and upgrade my evaluation.

A gentle reminder

Dear Reviewers and ACs, We again appreciate your valuable feedback on our submission. We have addressed your concerns and questions (e.g., scalability to the large-scale dataset, robustness, dataset with spurious correlation, and alternative metrics) throughout various experiments. To further discuss remaining concerns during this limited rebuttal period, we first (1) assert the motivation and virtue of our method and also (2) elaborate on the **computational cost** of our method through **new experimental results**. First, we propose an out-of-distribution detection (OOD) method via diffusion models (or consistency models) since they are widely used, and able to generate photorealistic images, hence may show potential in various downstream tasks. However, existing methods either show underwhelming OOD performance or require a high computational cost due to their requirement of multiple sampling. We show improvement on both sides: competitive performance in various domains (see Table 1, Table 2, Table 7 in Global Response PDF) and gain in computational cost. We further show that our method is robust while reducing computational costs for practical applications. For example, in Figure 4 (right), we show that even with 10x fewer computations, the AUROC of CIFAR-10 vs CIFAR-100 OOD detection performance decreases by less than 0.01. In such a hyperparameter, we take 0.081s/sample in wall-clock time and hence 8.7x faster than the computational cost of LMD. We report the OOD detection performance and the computational cost of the original version and the fast version compared to the LMD in the table below. ___ |Method |SVHN |CIFAR-100|LSUN |ImageNet|computational cost(sec/sample)| |--------|-----|---------|------|--------|------------------| |LMD [1] |0.979|0.620 |0.734 |0.686 |0.705 | |PR (ours, fast version)|0.992|0.767 |0.828 |0.806 |0.081 | |PR (ours) |0.993|0.775 |0.837 |0.814 |0.395 | Again, we truly appreciate your effort and time to review this work! Best, \ Authors **References**\ [1] Unsupervised out-of-distribution detection with diffusion inpainting, ICML 2023

After the rebuttal

Thanks for the authors' response. In the rebuttal. the authors cannot well address my concerns, w.r.t. running time and comparison problems. Although the proposed method seems to be novel, addressing my concerns is important to further evaluate the effectiveness of the proposed method. In the evaluated datasets, I think they are less biased toward the background. I will decrease my initial score to borderline accept.

Acknowledgement

Thanks for your response to address my questions. I have increased the score to 5.