Towards Flexible and Controllable Unknown Rejection

We sincerely thank Reviewer AfjX for your review and are grateful for the time you spent on our submission. Below, we provide a point-by-point rebuttal to clarify your concerns.

Q1. The novelty of this paper seems somewhat limited. The paper aims to address Unrecoverable, Inflexible, and Inefficient problems, but in my view, these issues are quite easily resolved by introducing LoRA.

We agree with you that equations (4) (AugMix) and (5) (OE) are well-known, existing methods, which have been clearly described in the submission. However, as the Remark on Page 5, our contribution is not to design a specific method for OOD generalization or a specific method for OOD detection. Differently, we focus on the unified unknown rejection in the wild, where the key difficulty is how to get a model that can detect both misclassified covariate-shifted and semantic-shifted inputs in a unified and flexible manner because many recent methods fail to achieve this goal (e.g., many methods perform worse than the CE baseline). In our paper, we demonstrate that combining existing OOD generalization and detection is not trivial, e.g., the straightforward multitask learning paradigm has limitations due to the objective conflict between the two objectives. The proposed TrustLoRA is verified to be quite effective and flexible in the spirit of reliability separation and integration, outperforming multitask learning paradigm. We would like to clarify our contributions as follows.

• Problem setting. To the best of our knowledge, we might be the first to call for (1) flexible and controllable unknown rejection that can control the uncertainty behavior of one's model, as deployment settings may evolve and change. Therefore, the topic of our paper is important, underexplored, interesting, and focused. (2) Rejecting both misclassified covariate-shifted and semantic-shifted samples in a unified manner, are both very important for model trustworthiness. There are only a few previous works that consider the recognition of covariate-shifts and rejecting semantic shifts, while they do not consider the rejection of misclassified covariate shifts.
• Problem analysis. (1) In Section 4.1, we show the trade-off between the two unknown rejection tasks and verify that many popular existing methods are less effective for the unified unknown rejection problem. As shown in Fig.4, we can clearly observe that when fine-tuning with OE to acquire OOD detection ability, it is harder to detect misclassified corrupted samples. As a result, the unified unknown rejection performance also becomes worse. (2) On page 9, we analyze why multi-task learning is not the final solution for the unknown rejection problem. We believe those observations are insightful and enlightening.
• Proposed method. While the proposed TrustLoRA framework may appear natural and simple once explained, (1) to the best of our knowledge, we are the first to introduce LoRA into unknown rejection or OOD detection fields. Besides, the focus on parameter efficiency is also novel since efficiency is typically overlooked in the domain of uncertainty estimation, and TrustLoRA potentially presents a low-cost method of adding better uncertainty estimation ability to an existing model. The proposed random projection-based LoRA further improves the efficiency. (2) TurstLoRA is a general and simple framework. Although we implement it based on AugMix and OE, it can be built on other existing methods. Experimentally, our method outperforms strong baselines and the multi-task learning approach.

The above contributions have been acknowledged by other Reviewers. In summary, this work (1) studied an under-explored and important problem with desirable property for real-world application (Reviewer Sug3, 8ZXm and 8Zzj); (2) provides insights, offers important motivation for future research (Reviewer 8Zzj), and opens up exciting possibilities for future expansions (Reviewer 8ZXm); (3) the proposed TrustLoRA for unknown rejection is innovative and interesting (Reviewer Sug3 and 8Zzj), allowing flexible adaptation to deployment situations.

Thank you for your response, and I respect your opinion, but I still disagree with your viewpoint. The contribution you mentioned is more of a framework-level contribution rather than a methodological one. I still believe that merely proposing such a so-called framework might be insufficient in terms of methodology. For instance, the methodological contribution of LoRA, which you cited, can be clearly articulated in terms of methodology .

Dear Reviewer AfjX,

Thanks for your feedback. In our previous response (Part 1/3), we summarize the contribution of this paper in three aspects, i.e., problem setting, problem analysis, and proposed method. Your major concern is about the proposed method.

AugMix and OE are failure-specific methods. In recent years, various failure-specific, especially OOD detection methods, have been proposed. However, they are useless or even harmful for real-world unknown rejection when facing both covariate and semantic shifts. We respectfully argue that investigating how to better use existing failure-specific techniques is meaningful and valuable, perhaps more important than proposing another failure-specific approach.

This work aims to design a unified framework to integrate different sources of reliability knowledge. We would like to emphasize that ensembling different failure-specific objectives is not trivial. For example, optimizing AugMix and OE in the multi-task learning manner is less effective (e.g., AugMix + OE is worse than OE) due to the existing conflicts (as shown in the Table below, CIFAR-100 with S1), and also suffers from unrecoverable, inflexible, and inefficient issues. In this work, we modify the original LoRA with random projection (which is more computation and memory efficient) and tune it with different failure-specific objectives sequentially. Finally, flexible and controllable unknown rejection can be achieved with LoRA arithmetic. Therefore, this work is not simple LoRA + AugMix + OE. We believe our work shows the good application of LoRA for the important and challenging unified unknown rejection problem.

We hope the above responses and our previous responses have addressed your concerns. As other Reviewers mentioned, this work focuses on an under-explored and important problem (Sug3, 8ZXm, 8Zzj), provides valuable insights, offers important motivation for future research (8Zzj, 8ZXm), and the introduction of LoRA into unknown rejection field is innovative and interesting (Sug3, 8Zzj, 8ZXm). If you still have some specific concerns, please let us know. We look forward to and appreciate your feedback.

Q2. The failure rejection setting in this study does not align with requirements for real-world environments... including ID test data would better align the study with realistic scenarios.

Thank you for your valuable suggestion. In the initial submission, we did not include clean ID in Table 1&2 in order to clearly reflect the unknown rejection ability under covariate and semantic shits. Following your suggestion:

• We added a clear description in Datasets and implementation part on Page 6.
• To comprehensively present and discuss the results when clean ID is involved, we (1) include a new Figure 7 about the MisD performance on clean ID and (2) report the unified unknown rejection performance evaluated follow your instruction in Table 7 on Page 10 in the revised version. We also report the results below (Table 7) and find that our method performs best. We believe that the newly added results and discussion would better align the study with realistic scenarios.

Q3. Specify what each red cross and green check mark represents.

Thank you for your comment. We have added the meaning of the red cross and green check in the caption of Figure 1.

Q4. Are there other advantages besides efficiency of training B matrix of LoRA? Does training both A and B affect performance?

For LoRA, tuning only B matrix mainly aim to further improve the efficiency of LoRA without affecting the performance. In below, we compare the original LoRA and our random projection based LoRA, and find that they perform similar. We have added those results in Table 12 in the Appendix.

Q5. Linear combination of LoRA weights for approximating a family of mappings.

We provide Figure 9 in the Appendix of the revised paper to show that TrustLoRA can approximate a family of mappings, where scaling $\alpha$ can control the preference between MisD under covariate shits and OOD detection. Figure 6 shows that we can achieve selective reliability forgetting with LoRA negation. Those results imply the effectiveness of the simple linear combination of LoRA weights once we compress the reliability knowledge in LoRA module.

Q6. Typos.

Thank you for your careful reading, and we corrected the typo in the revised paper.

We sincerely thank Reviewer 8Zzj for your review and are grateful for the time you spent on our submission. We are glad for the acknowledgment that our work provides insights into the unknown rejection task and the proposed approach is interesting, effective and practical to tackling a critical challenge in real-world applications. Below we would like to give detailed responses to each of your comments.

Q1. Insufficient review of prior research on unknown rejection.. discussions and comparative experiments with prior works [A, B, C, D] are necessary.

Thank you for this comment and suggested references. Following your suggestion, we added discussions and comparisons with those prior works (OpenMix [C], SURE [D], RCL [A], FS-KNN [B]) on unknown rejections in the revised version.

• In the Introduction section on Page 2, we discuss our work with them and add a new paragraph in the Related work section in the Appendix, providing detailed discussion.
• Experimental comparison have been conducted and added to Tables 1&2 on Pages 7 and 8 of the revised version, and we summarize the results below. It is observed that TrustLoRA can outperform those strong baselines.

Reference

[A] Zhu et al. “RCL: Reliable continual learning for unified failure detection”, CVPR 2024.

[B] Cen et al. “The devil is in the wrongly-classified samples: Towards unified open-set recognition”, ICLR 2023.

[C] Zhu et al. “OpenMix: Exploring outlier samples for misclassification detection”, CVPR 2023.

[D] Li et al. “SURE: Survey recipes for building reliable and robust deep networks”, CVPR 2024.

We sincerely appreciate Reviewer 8ZXm for the positive recommendation as well as the valuable suggestions. We really appreciate your kind words that our work is effective, efficient and flexible for unknown rejection in real-world applications. Below we would like to give detailed responses to each of your comments.

Q1. Dependence on Auxiliary Data, and how to mitigate this dependency.

(1) We agree with you that reliance on outlier exposure is a notable limitation which is also true for other OE-based OOD detection methods. Without auxiliary data, it could be quite difficult to separate covariate shifts and semantic shifts, whose distribution often overlaps and this challenging problem is rarely considered by previous work. While auxiliary data is important, we find that our method consistently outperforms other auxiliary-based methods such as OpenMix, SCONE, and OE (as shown in Table 1&2), which demonstrates the superiority of TrustLoRA.

(2) To mitigate this dependency on auxiliary data, we could consider the setting in SCONE [A], which assumes that the unlabeled covariate and semantic shifts are available at inference time and can be used to finetune the model directly. Besides, we can also explore the paradigm of human-in-the-loop to acquire few shot outliers via system feedback or active learning. Our proposed framework can be easily extended to those settings, and more importantly, is much more efficient than existing methods since we only tune the LoRA module.

(3) Inspired by your valuable comment, we conduct experiments to demonstrate the robustness of our method. Specifically, the RandomImage is used as auxiliary outliers following existing work. In the following Table, we show that TrustLoRA is robust to other auxiliary outliers like TIN597 [B]. We have added those results in the revised paper (Table 6 on Page 10).

Reference

[A] Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. ICML 2023

[B] OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detection. ArXiv:2306.09301

Q2. Minor comments L239: For semantic shifts.

Thank you for your careful reading, and we corrected the typo in the revised paper.

Q3. How does gradient interference in multi-task learning impact the simultaneous optimization of covariate shift adaptation and semantic shift detection tasks? How could the proposed TrustLoRA mitigates this?

(1) Conflicts existed. When optimizing both covariate shift adaptation and semantic shift detection in a multi-task learning manner, there exist remarkable conflicts between pulling covariate-shifted samples close to class centers while pushing semantic-shifted samples away from class centers. This is the reason for the poor performance of existing methods for the joint unknown rejection task. Theoretically, the Bayes-optimal reject rule for detecting misclassified covariate shifts is based on maximum class-posterior probability ${\max}_{y \in \mathcal{Y}}\mathbb{P}(y|**x**)$, while detecting semantic shifts is based on density ratio $p(**x**|\text{in}) / p(**x**|\text{out})$, as follows:

Proposition: (1) Bayes-optimal reject rule for detecting misclassified covariate shifts. For the risk of detecting misclassified covariate shifts $R_{cov-MisD}(h,r)$, the optimal solution $r^*$ of minimizing $R_{cov-MisD}(h,r)$ is:

where $c \in (0,1)$ is the reject cost. (2) Bayes-optimal reject rule for detecting semantic shifts. For the risk of detecting misclassified covariate shifts $R_{OOD}(h,r)$, the optimal solution $r^*$ of minimizing $R_{OOD}(h,r)$ is:

Thank you for the authors' responses. I also reviewed the responses to other reviews, and I do not find any concerns significant enough to affect the score from my side, so I will keep it as is. I believe that auxiliary data, when combined with current generative AI methods, holds strong practical potential. I appreciate the good application of an existing idea like LoRA during investigating an important problem and would like to give it high contribution for that.

We sincerely appreciate Reviewer Sug3 for the review and are grateful for the time you spent with our submission. We are glad for the acknowledgement that our approach is innovative and addresses a gap for real-world application requirements in dynamic environments. We wish to address your concerns by giving detailed responses to each of your comments as follows:

Q1. Although Proposition 3.1 is presented in the main paper, its proof is not provided.

As described in our theoretical analysis, Proposition 3.1 is a corollary of Theorem 4 in (Dimitriadis et al., 2023). Following your suggestion, we provide the proof below and add it to the Appendix (Section B) of the revised version.

Proof. Denote $\sigma$ the ReLU non-linearity $\sigma(x) = \max(0, x)$. From the universal approximation theorem [A], for any $\epsilon > 0$, there exists $Q \in \mathbb{N}$, $M \in \mathbb{R}^{(D+2) \times Q}$, $C \in \mathbb{R}^Q$, $M' \in \mathbb{R}^{Q \times D'}$ such that： $\forall x \in \mathcal{X}, \forall n \in \{1,...,N\}, ~~\|f_n(x) - g(x, \alpha) \| \leq \epsilon,$ where $g(x, \alpha) = M'\sigma(M(x, \alpha) + C)$. Define two matrices $R \in \mathbb{R}^{D \times (2D+2)}$ and $S \in \mathbb{R}^{(2D+2) \times D}$ as follows: $R_{i,j} = \begin{cases} 1, & \text{if } j = 2i - 1 \\ -1, & \text{if } j = 2i \\ 0, & \text{otherwise} \end{cases}$ and $~~~S_{i,j} = \begin{cases} 1, & \text{if } i = 2j - 1 \text{ or } (i > 2u \text{ and } j = u) \\ -1, & \text{if } i = 2j \\ 0, & \text{otherwise} \end{cases}$ and let $W_k = (0,...,0,k), k=1,2$. Then, with ${x} \in \mathbb{R}^D$, we have $\forall \alpha \geq 0, \quad S^\top\sigma(R^\top x + (1 - \alpha) W_1 + \alpha W_2) = ({x}, \alpha).$ We can learn a ReLU multi-layer perceptron $f(x, M,C,M',R,S,W)=M'\sigma(MS^\top\sigma(R^\top {x} + W) + C)$. Then with $\theta_{\text{pre}} = (M, C, M', R, S, 0)$, $\theta_i = (0,0,0,0,0, W_i), i=1,2$, we have