Out-of-distribution Detection Learning with Unreliable Out-of-distribution Sources

Summary

This paper contributes a new data-generation method to train a OOD detector. The paper proposes to first create an auxiliary task for a generator to generate in-distribution samples and OOD samples by using regions of disjoint latent space (equ 6-7). This can lead to disjoint support set by enforcing a distance preserving loss (equ 8). In order to transfer auxiliary task to real OOD detection, it suggests to use a contrastive learning to bring together generated in-dist samples and real in-dist samples. With these two settings, the OOD detector can learn from generated OOD samples.

Strengths

This discussion is complete and clear. The idea of crafting the auxiliary task using generative model(s) is promising in general. By dividing the samples into high and low density regions, this method bypasses the issue of mistaken OOD samples.

Weaknesses

The major concern is the performance of the generator (or possibly I miss that piece of information). While section 3 is convincing, in practice (section 4) how to guarantee high MoG density region has high concentration of in-distribution samples. This may happen when a generator does not perform well.

Questions

$\bullet$ Can the author change the term predictor to OOD detector? $\bullet$ While it is a good idea to generate in-dist and OOD samples using high MoG density region, does it also make the task too easy? I imagine that high MoG region samples may look like in-dist samples, while the rest may look like noise. Does the predictor can perform better if it can learn from more difficult (harder to discern) generated in-dist and generated OOD samples? $\bullet$ Does this method consider the case when high MoG density region still have some OOD samples? $\bullet$ The performance gain of the (OOD) predictor may be comparable to the generative model used for auxiliary task. Can the auxiliary generative model(s) be used for OOD detection? Can you provide its (or their) performance? If the generative model(s) performs worse, what could contribute to the advantage of the predictor (maybe the contrastive learning loss or maybe the auxiliary task)? Can you identify where the advantage comes from?

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

3 good

Limitations

The paper states there is a limitation discussion. Where is it?

Summary

The paper propose to fix the mistaken OOD generation issue in generative model based approach to out-of-distribution data detection, where the mistaken OOD generation means generated OOD data have semantics of ID data. To fix this issue, auxiliary task-based OOD learning (ATOL) is proposed, which is claimed to have the effect of satisfying two key conditions, i.e., auxiliary ID and auxiliary OOD data have disjoint supports in the input space, and auxiliary OOD data are reliable. To achieve this goal, ATOL adds auxiliary task learning loss and ID distribution alignment loss to the real task learning loss. The empirical study shows the non-trivial improvement of OOD detection performance when using ATOL. Ablation study further confirms the effectiveness of each loss in ATOL.

Strengths

The motivation to fix the mistaken OOD generation makes sense to me as shown in Fig. 1. The two new losses in the proposed ATOL are reasonable and are clearly presented. I have to say that I like the empirical studies in the paper since they are quite comprehensive and strong, though many interesting results are presented in the supplementary as a result of page limits.

Weaknesses

The drawback and strength of generative model based OOD detection and its comparison with other approaches like scoring or regularized training is not fully discussed in the paper, e.g., performance and efficiency. In Table 17, the benefit of ATOL is not quite clear when compared with ReAct and CSI.

Questions

1. What is M(z) in Equation 6 and 7? Is it density function of MoG? 2. It is kind of misleading to make the ATOL result bold in Tab. 17 since it is not the best result in terms of AUROC. The best one should be CSI. 3. What is the proportion in Fig. 1b and how is it computed? I tried to search for this information in the paper but it seems that there is no such information. I believe this information is quite important to motivate this paper.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Limitations are not discussed in the main paper (please correct me if I am wrong about this).

Summary

The paper tries to overcome the impact of directly applying incorrect OOD data on the OOD model through auxiliary tasks, thereby improving the performance of OOD tasks. The theoretical part is hard to follow, and the experimental part proves the effectiveness of the theory.

Strengths

1. The paper introduces an auxiliary OOD detection task to combat mistaken OOD generation. 2. The proposed method requires a small additional calculation cost. 3. Experiments show the effectiveness of the proposed method.

Weaknesses

1. The proof of C2 is not clear enough to allow me to clearly understand why achieving C2 can better utilize OOD data. If the model is strong enough or fully trained, it can still confuse incorrect OOD data even if formula 5 is met. Therefore, I judge that the training steps of the algorithm should not be too many and the model should not be too large. Does the author have an explanation for this aspect? 2. There is a writing error in part b of Formula 10. 3. Many aspects of the experiment followed the settings of reference [31], but why not include them in the comparison? 4. The paper is really hard to follow, please polish the paper carefully.

Questions

There are too many Mathematical notations. It is suggested to sort out a table to make it clearer.

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

Yes.

Summary

One of the techniques for detecting OOD instances is to train a model on OOD data. However, that task is not easy due to difficulty inherent with collecting such OOD data. Rather than collecting such data, this paper proposes instead to generate it, and to train an auxiliary task to improve the OOD detection capabilities of deep learning networks. The proposed approach is called ATOL, for Auxiliary Task-based OOD Learning, and proposes to address one fundamental flaw in existing data-generation based detection methods: the collection of OOD instances from ID data that can mistakenly be labeled as OOD data while being in reality ID.

Strengths

This paper addresses an interesting problem that I think is quite overlooked in the research community. Indeed, many studies focus on collecting OOD instances without necessarily evaluating the possible side effects of the collected data. This paper identified that instances that are collected and labeled as OOD instances can in fact be ID instances, which could lead to poor generalization performance on ID data, and poor OOD detection capabilities. The approach the authors devised to address the problem is sound and quite intuitive, and the evaluation is quite strong.

Weaknesses

Although this paper addresses the OOD detection problem from a data-generation perspective, I would have very much liked to see how their approach fair with other techniques like distance-based OOD detection methods. Some interesting distance-based OOD detection mechanisms have been proposed in the recent past. For instance, CIDER [1] has achieved state-of-the-art performance OOD detection capabilities that shouldn't be overlooked by the researchers approaching OOD detection from a data-generation perspective. This would help educate mainstream readers more on exactly what techniques to pursue to robustify their models against OOD samples. A more fundamental limitation of this study is the fact that it heavily relies on a mixture of Gaussian to decide on what latents to consider as OOD and which ones to consider as ID. As MoGs can be sensitive to outliers, have a rather limited expressive power, their accuracy needs to be presented in the paper to showcase their effectiveness in helping collect the data to train the auxiliary task on. [1]: How To Exploit Hyperspherical Embeddings For OOD detection? https://arxiv.org/pdf/2203.04450.pdf

Questions

Based on the limitations I raised above, I would suggest the authors to perform a comparative study with some of the distance-based OOD detection approaches like CIDER. Additionally, evaluating the performance of the MoGs they used could help validate further the effectiveness of their approach.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes

Thank you for your reply!

Thank you for your reply! First I completely agree with other reviewers that this draft is not very well written because the main points are not clearly answered. I would like to summarize my idea here and confirm with the authors. The main point is that, if the auxiliary generative models could achieve C1 which needs the disjoint support set for gen-in vs gen-OOD, and C2 which needs the gen-in is close to real-in. Then this framework will work. Moreover, based on what the author replies, then how to make sure the sets of gen-in and gen-OOD disjoint while gen-OOD more difficult could make the results better. Can the author(s) confirm the above summary? Thank you!

After Rebuttal

Thanks for the response, I will keep my score.