Optimal Transport Model Distributional Robustness

Distributional robustness is a promising framework for training deep learning models that are less vulnerable to adversarial examples and data distribution shifts. Previous works have mainly focused on exploiting distributional robustness in the data space. In this work, we explore an optimal transport-based distributional robustness framework in model spaces. Specifically, we examine a model distribution within a Wasserstein ball centered on a given model distribution that maximizes the loss. We have developed theories that enable us to learn the optimal robust center model distribution. Interestingly, our developed theories allow us to flexibly incorporate the concept of sharpness awareness into training, whether it's a single model, ensemble models, or Bayesian Neural Networks, by considering specific forms of the center model distribution. These forms include a Dirac delta distribution over a single model, a uniform distribution over several models, and a general Bayesian Neural Network. Furthermore, we demonstrate that Sharpness-Aware Minimization (SAM) is a specific case of our framework when using a Dirac delta distribution over a single model, while our framework can be seen as a probabilistic extension of SAM. To validate the effectiveness of our framework in the aforementioned settings, we conducted extensive experiments, and the results reveal remarkable improvements compared to the baselines.

Paper

References (72)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer X3P46/10 · confidence 3/52023-06-16

Summary

This paper explores the relationship between OT-based distributional robustness and sharpness-aware minimization (SAM), and shows that SAM is a special case of the proposed framework when a Dirac delta distribution is used over a single model. To demonstrate the effectiveness of the framework, the authors perform several experiments and show improvements over the baselines.

Strengths

1. This paper is theoretically sound. 2. The experiment verifies the effectiveness of the proposed framework.

Weaknesses

The practical methods stated in Section 4.2 are not easy to follow. It would be better if the authors can provide a pseudocode to help the readers understand the methods easily.

Questions

1. Theorem 4.2 holds for $p=+\infty$, can the results be extended to more general settings? 2. To make the OP problem solvable, the authors resort to the entropy-regularized version of OP. However, when establishing the connection between SAM and OT-based model distribution robustness, the authors use the original version of OP. Can the authors elaborate more on this?

Rating

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

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

2 fair

Contribution

3 good

Limitations

The practical methods stated in Section 4.2 are not easy to follow.

Reviewer WK2o5/10 · confidence 4/52023-06-29

Summary

This paper proposes an optimal-transport-based distribution robustness framework on model space. The proposed theory considers the worst-case loss w.r.t a Wasserstein distance from the model distribution, which can be considered a probabilistic extension of sharpness-aware minimization (SAM).

Strengths

- The paper is well-written and easy to follow. - The proposed approach shows empirical improvements compared to SGD, SAM, DGVB, etc.

Weaknesses

- The theoretical and empirical results are not surprising to me. Moreover, they do not conduct generalization analysis, which is important in learning theory. - I do not agree with the claim that this paper is the first work that considers distribution robustness within the model space. The sharpness-related learning problem has already been well addressed in PAC Bayes theory. The counterpart of the worst-case awareness studied in this paper can be found in [1]. And other Wasserstein or more general theories can be referred to [2] and [3]. - The single-model, ensemble, and BNN cases are also common algorithm realizations in the PAC Bayes community. >[1] Jiang, Yiding, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. "Fantastic Generalization Measures and Where to Find Them." In *International Conference on Learning Representations*. >[2] Amit, R., Epstein, B., Moran, S. and Meir, R., Integral Probability Metrics PAC-Bayes Bounds. In *Advances in Neural Information Processing Systems*. >[3] Haddouche, Maxime, and Benjamin Guedj. "Wasserstein PAC-Bayes Learning: Exploiting Optimisation Guarantees to Explain Generalisation." (2023).

Questions

I suggest the author fully discuss the aforementioned related works and improve their methods or add comparisons to corresponding algorithms PAC Bayes community.

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

No, I suggest the authors add a section to discuss the limitation of the proposed algorithm, e.g., the increment of computation complexity of BNN and the multiple particle setting.

Reviewer NFva6/10 · confidence 3/52023-07-06

Summary

This paper explores the optimal transport distributional robustness framework in the model parameters spaces, and provides practical algorithms to solve the corresponding optimization problems. This flexible framework can be used to train a single model, ensemble models and Bayesian neural networks in a sharpness-aware fashion. The authors additionally establish that the sharpness-aware minimization serves as a special instance within their proposed framework. The experimental results demonstrate the usefulness of this framework.

Strengths

This paper considers an important question of neural networks training. It is clearly written and easy to follow. The idea of extending the form of SAM to the DRO framework is natural but very meaningful, since it provides more well-developed tools to understand and improve the generalization of training models.

Weaknesses

1. In the theory, this paper shows that SAM is equivalent to the optimization problem in the OT-MDR framework when using a particular distance metric and picking the center distribution as a Dirac delta distribution. In the experiments, the results show (1) OT-MDR performs better than SAM for both single model and ensemble models cases; (2) The loss landscape found by OT-MDR is more flatten than the one found by SAM. This probably implies the training algorithm of OT-MDR is better than directly using SGD to solve SAM. But the paper doesn’t provide much intuitive understanding or analysis of these observations. 2. Most of the experimental results don’t provide standard deviation, which makes the results less solid.

Questions

Q1. In SAM [1], their experiments show that using 2-Wasserstein distance is better than using \infty-Wasserstein distance. For the current paper, in theorem 4.3, it shows that choosing center distribution as a Dirac delta in OT-MDR with 2-Wasserstein distance is equivalent to SAM (using 2-Wasserstein distance). But based on theorem 4.2, the algorithm solves the minimization problem of OT-MDR with \infty-Wasserstein distance. And the algorithm of OT-MDR still outperforms the SAM (with 2-Wasserstein distance). Does this probably imply the performance of OT-MDR can be further improved when using algorithms that solves the minimization problem of OT-MDR with 2-Wasserstein norm? Here are a few more delicate methods [2,3] to solve the general distributional robustness learning problem. [1] Foret, P., Kleiner, A., Mobahi, H., & Neyshabur, B. (2020). Sharpness-aware minimization for efficiently improving generalization. [2] Wang, G., & Chizat, L. (2022). An exponentially converging particle method for the mixed nash equilibrium of continuous games. [3] Trillos, C. G., & Trillos, N. G. (2023). On adversarial robustness and the use of Wasserstein ascent-descent dynamics to enforce it. Q2. In figure 1, the difference of accuracy for using different numbers of particles is actually small. Since this figure doesn’t provide the information standard deviation, it might be difficult to conclude that using 2 particles gives the best results. It would be better if the author could provide different random seeds results of this experiment. It will probably provide more insight. Q3. As mentioned in the weakness, the experimental results in table 1, 3 and 4 don't contain standard deviation. Could the author provide more results on those experiments?

Rating

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

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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

In general, I like the idea and the presentation of this paper. The authors could probably further improve this paper through more experiments and add some analysis of the observations from comparing SAM and OT-MDR.

Area Chair a4y92023-08-20

Reminder from AC

Dear reviewer, The author-reviewer discussion period ends in 2 days. Please review the authors' rebuttal and engage with them if you have additional questions or feedback. Your input during the discussion period is valued and helps improve the paper. Thanks, Area Chair

Reviewer NFva2023-08-21

I would like to thank the authors for their responses. Along with previous recommendations, the experimental results should be updated with standard deviations. Overall I would like to keep the score.

Authorsrebuttal2023-08-22

Thank you again for taking the time to review our paper and give us valuable comments. Due to the rebuttal time constraints, for the standard deviation, we focus on experiments with OT-MDR as reported in the PDF file. We will definitely add full experiments and results to our revised version.

Reviewer P1BN7/10 · confidence 3/52023-07-13

Summary

This paper proposes a novel distributional robustness optimization problem on the space of model parameters (OT-MDR). By formulating the uncertainty set with Wasserstein distance and introducing an entopy regularization, the explicit solution to the primal problem is derived. A practical algorithm is proposed based on sampling techniques from SGLD. The algorithm is extended beyond point estimation to accomodate emsembling and bayesian learning. The dual of OT-MDR turns out as a stochastic relaxation of the objective of SAM, bridging sharpness aware optimization and distributional robustness. Empirical results on Cifar validates the advantage of proposed method over baseline optimizers.

Strengths

1. I might be unfamiliar with literature on sharpness aware optimization, but from my perspective this paper is the first to reveal an intrinsic connection between distributional robustness and local sharpness. Theorem 3.4 is intriguing since it might inspire further research to apply abundant literature on distributional robustness (and adversarial robustness which is highly relative) to solving sharpness aware optimization. 1. The proposed method is a stochatstic extension of SAM but it introduces limited computational burden. On the other hand, the algorithm well extends to bayesian learning because of its stochastic formulation of model parameters.

Weaknesses

This paper proves OT-MDR to be more general than SAM but the advantage of OT-MDR in the case of single model and emsembling is theoretically vague. From Theorem 3.4, it seems that OT-MDR is equivalent to SAM for single model except that OT-MDR softly expands the uncertainty ball around the centering model, by allowing parameters to go beyond the radius $\rho$. The gap is insufficient to account for the empiricial success of OT-MDR for single model. Additionally, is it possible that enlarging the radius $\rho$ for SAM would make its loss landscape approach OT-MDR in Figure 2?

Questions

1. The practical algorithm is based on SGLD which shows that stochastic gradient descent could converge to the posterior distribution of model parameters. However, this paper adopts two-step sampling such that the sampled parameters might deviate much from the distribution in Theorem 4.2. Is this a significant gap between the empirical algorithm and theory?

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

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

4 excellent

Contribution

3 good

Limitations

na

Area Chair a4y92023-08-20

Reminder from AC

Dear reviewer, The author-reviewer discussion period ends in 2 days. Please review the authors' rebuttal and engage with them if you have additional questions or feedback. Your input during the discussion period is valued and helps improve the paper. Thanks, Area Chair

Reviewer X3P42023-08-10

Thank you for your response. The response addresses my concerns. I recommand accept of this paper.

Reviewer WK2o2023-08-17

Thanks for the authors' responses. I have been a little busy recently. So sorry for the late reply. * I am afraid that I cannot agree with your discussion on [1]. I mentioned [1] since it interprets that PAC-Bayes can capture sharpness in the expected sense compared to conventional worst-case sharpness in Sec. 4.3 (Success of Sharpness-Based Measures). Moreover, it uses the PAC-Bayesian framework to construct generalization bounds for worst-case sharpness bound (Eq.(50)) presented in Sec.C.3 (Flatness-based Measures). * Use PAC-Bayes bound to derive a similar optimization objective in your paper. - A typic Wasserstein-based PAC-Bayes Bound has the following form: $\mathbb{E}\_{\theta \sim \mathbb{Q}\_{\phi}} \mathcal{L}\_{\mu}(\theta) \leq \mathbb{E}\_{\theta \sim \mathbb{Q}\_{\phi} }\mathcal{L}\_{S}(\theta) + \sqrt{c_1(m, \delta)\cdot W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}}) + c_2(m,\delta)}$, where $c_1, c_2$ are some sample complexity functions, $\tilde{\mathbb{Q}}$ is prior distribution of on the model space and $\mathbb{Q}_{\phi}$ is the posterior. - Based on the bound, we often set the optimization objective as $\min_{\phi} \{\mathbb{E}\_{\theta \sim \mathbb{Q}\_{\phi} }\mathcal{L}\_{S}(\theta) + \lambda W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}})\}$. - In PAC-Bayes community, $\tilde{\mathbb{Q}}$ is often some known prior. But if we try to find the optimal prior, we can have $\min_{\phi} \min_{\tilde{\mathbb{Q}}} \\{\mathbb{E}\_{\theta \sim \mathbb{Q}\_{\phi} }\mathcal{L}_{S}(\theta) + \lambda W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}})\\}$. - If we set a constraint for $W_1(\mathbb{Q}_{\phi}, \tilde{\mathbb{Q}}) \leq \rho$, we can have the following objective: $\min\_{\phi}\min\_{\lambda\geq 0}\\{\lambda \rho + \max\_{\tilde{\mathbb{Q}}} \\{\mathbb{E}\_{\theta \sim \mathbb{Q}\_{\phi} }\mathcal{L}\_{S}(\theta) - \lambda W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}})\\}\\}\ $, which is the same as your eq(14). These are what I expected from your discussion. And that's why I said the results are not surprising to me and I do not think the paper is the first work that considers distribution robustness within the model space. However, since DRO and PAC-Bayes are considered different domains, and the paper has other experimental contributions, I will discuss them with AC and other reviewers to decide whether I will modify the score.

Authorsrebuttal2023-08-18

Response to the reviewer's further derivations

We thank the reviewer for providing detailed comments. While we value your input, we respectfully present our counterarguments and responses to your points and derivations as outlined below. - We agree with the reviewer up to $\\text{min}\_{\\phi}\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)+\\lambda W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)$. - We consider $\\text{min}\_{\\phi}\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)$ **(Eq. 1)**. - However, the next one $\\text{min}\_{\\phi}\\min\_{\\tilde{\\mathbb{Q}}}\\left\\{ \\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)+\\lambda W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)\\right\\}$ **(Eq. 2)** does not make any further progress and does not have the flavor of distributional robustness. The reason is that Eq. (1) and Eq. (2) are equivalent because one can easily verify that if $\\mathbb{Q}^*$ is an optimal solution of Eq. (1) then $\\left(\\mathbb{Q}^*, \\mathbb{Q}^* \\right)$ is an optimal solution of Eq. (2). - We now look into $\\min\_{\\phi}\\min\_{\\lambda\\geq0}\\left\\{ \\lambda\\rho+\\max\_{\\tilde{\\mathbb{Q}}}\\left\\{ \\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)-\\lambda W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)\\right\\} \\right\\} $ **(Eq. (3))**. It is very obvious that $\\max\_{\\tilde{\\mathbb{Q}}}\\left\\{ \\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)-\\lambda W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)\\right\\} =\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right) $, obtaining when $\\tilde{\\mathbb{Q}}=\\mathbb{Q}\_{\\phi}$, hence the OP in (3) can be simplified to $\\min\_{\\phi}\\min\_{\\lambda\\geq0}\\left\\{ \\lambda\\rho+\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)\\right\\} $. This can be further minimized to $\\min\_{\\phi}\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)$ **Eq. (1)**, obtaining with $\\lambda = 0$. - Therefore, Eq. (3) does not have any flavor of distributional robustness. Certainly, it cannot be equivalent to our Eq. (14) which has a clear flavor of distributional robustness. Furthermore, in our derivations for both main theories and practical methodologies, we depart from Eq. (5), which is a stronger and more comprehensive formulation than Equation (14). To summarize, we cannot see any flavor of distributional robustness within your derivations in derivations Eqs (1, 2, 3), although they look interesting. Additionally, the papers [1], [2], and [3] you referenced do not explicitly address or mention distributional robustness in the model space. We believe that it should not require the readers to infer the things that **were not mentioned explicitly in the papers**.

Reviewer WK2o2023-08-18

Sorry for the typo, it should be $\min\_{\phi}\min\_{\lambda\geq 0}\\{\lambda \rho + \max\_{\tilde{\mathbb{Q}}} \\{\mathbb{E}\_{\theta \sim \tilde{\mathbb{Q}} }\mathcal{L}\_{S}(\theta) - \lambda W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}})\\}\\}\$. When $W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}}) \leq \rho$, we have $\lambda =0$, and the objective becomes $\min\_{\phi}\max\_{\tilde{\mathbb{Q}}:W_1(\tilde{\mathbb{Q}}, \mathbb{Q}) \leq \rho} \mathbb{E}\_{\theta \sim \tilde{\mathbb{Q}} }\mathcal{L}\_{S}(\theta) $ as your eq.(13). When $W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}}) > \rho$, we have $\lambda \rightarrow \infty$, the objective becomes $\min_{\tilde{\mathbb{Q}}} W_1(\tilde{\mathbb{Q}}, \mathbb{Q}\_\phi)$. [1,2,3] do not mention distribution robustness, however, the sharpness-awareness has been mentioned and the related flat minima has been discussed in your paper. I do agree that maybe different domain has different terminology. So I will discuss these with AC and other reviewers.

Authorsrebuttal2023-08-19

Thanks for your acute discussion which certainly benefits our future research. However, we still maintain our viewpoint that [1, 2, 3] **did not discuss** distribution robustness on a model space. [1] is an empirical paper to study what factors influence the generalization ability. The Wasserstein-based PAC-Bayes bounds developed in [2,3] have the common form of $\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{\\mu}\\left(\\theta\\right)\\leq\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)+\\sqrt{c\_{1}(m,\\delta)W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)+c\_{2}(m,\\delta)},$ which suggests us the following OP $\\text{min}\_{\\phi}\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)+\\lambda W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)$ Basically, similar to other KL-based PAC-Bayes bounds, this states that we aim to learn the model distribution $\\mathbb{Q}\_\\phi$ that minimizes the empirical loss and stays as closest as possible to the prior $\\tilde{\\mathbb{Q}}$. Moreover, if one considers the prior $\\tilde{\\mathbb{Q}}$ as a variable to minimize: $\\text{min}\_{\\phi}\\min\_{\\tilde{\\mathbb{Q}}}\\left\\{ \\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)+\\lambda W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)\\right\\} $, this is only equivalent to $\\text{min}\_{\\phi}\\mathbb{E}\_{\\theta\\sim\\mathbb{Q}\_{\\phi}}\\mathcal{L}\_{S}\\left(\\theta\\right)$. Your subsequently suggested OP $\\min\_{\\phi}\\min\_{\\lambda\\geq0}\\left\\{ \\lambda\\rho+\\max\_{\\tilde{\\mathbb{Q}}}\\left\\{ \\mathbb{E}\_{\\theta\\sim\\tilde{\\mathbb{Q}}}\\mathcal{L}\_{S}\\left(\\theta\\right)-\\lambda W\_{1}\\left(\\mathbb{Q}\_{\\phi},\\tilde{\\mathbb{Q}}\\right)\\right\\} \\right\\} $ can be reduced to $\\min\_{\\phi}\\max\_{\\tilde{\\mathbb{Q}}}\\mathbb{E}\_{\\theta\\sim\\tilde{\\mathbb{Q}}}\\mathcal{L}\_{S}\\left(\\theta\\right)$ if we search $\\tilde{\\mathbb{Q}}: W\_{1}\\left(\\tilde{\\mathbb{Q}},\\mathbb{Q}\_{\\phi}\\right) \\leq \\rho$. It is obvious that the subsequently suggested OP is not equivalent to the suggested OP from PAC-Bayes bounds and we cannot see the pathway to connect them. More importantly, our main contribution in this paper is to theoretically turn the distributional robustness over a model space into practical methods that can be widely applied to several situations. Furthermore, we also point out the connection between distributional robustness and sharpness-aware minimization which we believe would benefit future research in developing new methods for improving the model generalization ability. Last but not least, your thoughtful discussion hints us a further question for our future research: _investigation of the connection between the WS-based PAC-Bayes bounds and the WS-based distributional robustness_. We really appreciate it. Finally, thanks for introducing us to the excellent papers [1, 2, 3]. We will definitely discuss them in the revised version.

Reviewer WK2o2023-08-20

Response

Thanks for the authors' reply. * I want to clarify some unclear or incorrect points in my previous responses. In fact, the prior $\tilde{\mathbb{Q}}$ should be data-free in PAC-Bayes, and the theoretical optimal should be $\tilde{\mathbb{Q}} = \mathbb{E}\_S \mathbb{Q}\_{\phi}$. Then we need $\min_{\phi} \\{\mathbb{E}\_{\theta \sim \mathbb{Q}\_{\phi} }\mathcal{L}_{S}(\theta) + \lambda W_1(\mathbb{Q}\_{\phi}, \tilde{\mathbb{Q}})\\}$. The robustness reflects on controlling the distance of the posterior from the prior (stability) and simultaneously minimizing the empirical risks. * However, you are correct that this objective cannot be straightforwardly connected to the second objective, which reflects the DRO. And considering these are in two different research domains, I raised my score accordingly.

Authorsrebuttal2023-08-20

Thank you for acknowledging our contribution and valuable discussion. We will definitely add more detail in the revised version.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC