acceptodds
Under review as a conference paper at ICLR 2027

PatMask: Pluggable Adversarial Topology Masking for Multi-Agent Reinforcement Learning under Communication Disconnection

Abstract

Cooperative Multi-Agent Reinforcement Learning (MARL) relies on communication to mitigate limited local observations. However, existing methods are vulnerable to communication disconnection. To address this, we propose PatMask, a pluggable adversarial topology masking mechanism to enhance the robustness of existing MARL methods under diverse communication disconnections. Specifically, PatMask introduces an adversarial mask generator that adaptively disrupts critical communication links during training. We establish a contraction guarantee for robust policy evaluation under topology perturbations and develop an alternating adversarial training procedure. Extensive evaluations on various cooperative benchmarks show that PatMask can maintain high win rate and high coordination efficiency across various environments under controlled communication disconnection. Compared with baselines trained without additional disconnection augmentation, PatMask achieves 15.7-41.6 percentage-point higher win rates with only 41% of the training steps and less than 77% of the communication cost. Moreover, trained PatMask policies can generalize without retraining across the evaluated disconnection types and probabilities.

Primary area:
reinforcement learning
Keywords:
Multi-agent systems, reinforcement learning, communication topology

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qJ1ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PatMask combines learned adversarial topology masks with random masking when training communication-based MARL policies. The generator maximizes TD error and is removed at execution. Experiments cover SMAC, SMACv2, LBF, and several communication backbones under different simulated failure patterns.

Critical strengths

The mechanism can be added to existing topology-aware learners without an extra execution-time module. The GAT, GCN, and TGCNET experiments support this practical flexibility.

Table 6 provides useful evidence beyond comparison with unaugmented baselines: hybrid masking improves over random-only masking by 10.4 and 11.8 percentage points on two challenging tasks.

Critical weaknesses

Adversarial edge selection is not fully separated from perturbation severity. Generated masks remove 57–77% of links, whereas Appendix C.1 uses 30% random dropout; the random-mask rate in Table 6 is unspecified.

The mask implementation needs clarification. As written, Eq. 6 does not match the stated survival-probability interpretation, and the gradient path from soft masks through the neighbor-based GAT equations is unclear. These details are central to reproducing the method.

Questions for authors

What random-mask rate is used in Table 6, and does the advantage persist when random and learned masks remove the same number of links?

Is Eq. 6 exactly the implemented sampler? Please specify how masks enter the attention computation and how gradients reach the generator.

2 comments

Thanks very much for your valuable comments. Our responses and revisions are provided as follows. We hope that we have addressed all your concerns. We appreciate your dedicated work on the review and will be grateful if you could raise your rating in response to our rebuttal.

W1/Q1. 1) About dropout rate sensitivity for baselines: Thanks for your suggestion of the sensitivity study of the dropout rates for each baseline method. In the revision, we train the baselines with random communication masking under various link-drop probabilities and test them under various disconnection modes. The results are shown in Tables 7 and 8 and discussed in Section 5.5 “Ablation Study->Random training masking” in the main text. Table 7 shows the results of different baselines trained with different random link-drop rates and tested under the mixed disconnection. Training with random link drops can directly improve each baseline's performance, but the dropout rate does not ensure a consistent magnitude of the improvement. PatMask, which combines adversarial training with random masking, outperforms all baselines across various dropout rates. Comparing each baseline at its best tested rate, TGCNET, trained with a dropout rate of 0.5, is the strongest at 90.5%, while PatMask surpasses it by 4.7 pp, achieving 95.2%. We further compare PatMask with the baselines trained with a random link drop under different test-time disconnection modes, with results shown in Table 8. PatMask outperforms the baselines with random masking in all three disconnection modes. These results indicate the benefit of adversarial training in PatMask over the random dropout of links. The learned mask does not randomly drop links but identifies specific links to retain or drop during training to tolerate diverse communication disconnections during execution.

Table 7. Effect of random training mask on MMM2, varying the training link-drop rate. Results are test win rates (%) of training baselines with various random dropout rates (q) on the communication links on MMM2 and testing under the mixed disconnection with p=0.5p=0.5 independently for nodes and links. q=0q=0 disables link dropout during training and corresponds to the default baselines..

Methodq=0q=0.3q=0.5q=0.7
GTDE+dropout@q36.9 ± 8.160.9 ± 3.668.4 ± 6.057.6 ± 6.8
TGCNET+dropout@q58.1 ± 13.082.6 ± 6.390.5 ± 5.787.6 ± 6.8
T2MAC+dropout@q55.8 ± 22.670.8 ± 17.275.6 ± 16.477.4 ± 15.7
GACG+dropout@q42.7 ± 17.861.5 ± 14.767.1 ± 13.565.2 ± 13.8
PatMask (adversarial and random masking)95.2 ± 1.695.2 ± 1.695.2 ± 1.695.2 ± 1.6

Table 8. Effect of random training mask on MMM2, varying the test communication disconnection modes. Results are test win rates (%) of training baselines with a random link-drop rate of 0.5 and testing under various disconnection modes on MMM2. Node-only and link-only disconnections each use a failure probability of p=0.5, and the mixed disconnection applies both failures independently.

MethodNodeLinkMixed
GTDE+dropout@q=0.570.6 ± 4.988.6 ± 4.468.4 ± 6.0
TGCNET+dropout@q=0.591.6 ± 4.591.8 ± 4.190.5 ± 5.7
T2MAC+dropout@q=0.580.8 ± 12.879.6 ± 12.175.6 ± 16.4
GACG+dropout@q=0.572.0 ± 10.571.6 ± 9.867.1 ± 13.5
PatMask94.6 ± 1.494.3 ± 1.395.2 ± 1.6

2) Clarifying the random masking setting and its sensitivity in Table 6: Random training mask in Table 6 uses a link-drop probability of 0.5. Note that Table 6 was not mainly intended to compare the approaches with random masking and adversarial masking. We mainly aim to show the orchestration effect of both masking (Hybrid vs. other masking settings). As we can expect and the results show, random masking can directly improve performance (Random-only vs. No Masking). However, in practice, test-time disconnection probability is usually unknown in advance. Therefore, it can be common to use unaligned training-time and test-time dropout rates, especially when the test-time dropout rate may shift. We are pleased to see that the adversarial masking not only indicates a better replacement for random masking, but can also collaborate with random masking to further improve performance (Adversarial-only and Hybrid vs. Random-only). In the revision, we add clarification of the link-drop probability setting in the caption of Table 6 and this benefit of PatMask in Section 5.5.``Ablation Study->Masking Strategy''.

To better compare the influence of the dropout rate of random masking, we add a sensitivity study of the dropout rates of the random masking when it complements the baselines in Section 5.5."Ablation Study->Random Training Masking". Please refer to the above response to W1/Q1.1 "About dropout rate sensitivity for baselines''.

W2/Q2. Clarifying the implementation of Eq. 6: I clarify the questions about implementation details one by one.

  1. Is Eq. 6 the implemented sampler, and is it proper to interpret πij\pi_{ij} as a survival probability?

Yes, Eq. 6 describes the implemented sampler. However, we agree that ``survival probability'' was an inaccurate interpretation of πij\pi_{ij}. πij\pi_{ij} is a learned gating parameter, with larger values favoring link retention, but its numerical value is not the probability that the sampled binary mask retains the link. In the revision, we correct the interpretation of this terminology and explicitly distinguish the soft sample from the binary forward mask and its straight-through gradient estimator.

  1. How do masks enter attention computation?

During attention computation, removed links receive zero attention weight, and attention is normalized over the remaining neighbors. Masks remove inter-agent communication links while preserving each agent’s access to its own features.

  1. How do gradients reach the generator?

During the generator update, as illustrated by the red arrows in Figure 1, the TD-loss gradient propagates through the joint value, message aggregation, and attention computation to the generator via the differentiable mask relaxation. The learner’s parameters are held fixed, while gradients with respect to the mask inputs remain enabled.

Reviewer_WFkTOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes PatMask, a pluggable adversarial topology masking mechanism to improve the robustness of multi-agent reinforcement learning under communication disconnection. It combines a learned adversarial mask generator with random masking through alternating adversarial training. The paper also establishes a contraction guarantee for robust policy evaluation under topology perturbations. Experiments on SMAC, SMACv2, and LBF show improved win rates and generalization across different communication disconnection types and probabilities without retraining.

Critical strengths
  1. The paper addresses an important problem. PatMask is simple and can be easily integrated into existing MARL methods.
  2. The experiments are comprehensive and the results are strong. The method performs well across different tasks and communication failure settings.
Critical weaknesses
  1. The claimed communication efficiency improvement is not well supported. The paper reports training-time mask sparsity but does not clearly explain how communication costs are measured.
  2. The improvement over dropout-based baselines is relatively small. PatMask achieves 95.2% on MMM2, compared to 91.2% for TGCNET with random dropout. Since adversarial masking is the main novelty of this work, this modest improvement makes the overall contribution less convincing.
Questions for authors
  1. What link-drop probability is used for random masking in Table 6? Could the performance gap between random and adversarial masking be caused by different masking rates?
  2. How is the reported 77% communication cost calculated? Does it refer to training or execution?
  3. In Table 10, Appendix C.1, PatMask only slightly outperforms TGCNET with random dropout. How do the authors explain this small improvement, given that adversarial masking is the main contribution?
2 comments

Thanks very much for your valuable comments. Our responses and revisions are provided as follows. We hope that we have addressed all your concerns well. We appreciate your dedicated work on the review and will be grateful if you could raise your rating in response to our rebuttal.

W1/Q2. Clarifying the communication cost: Thank you for identifying this imprecision. The reported 77% refers to the mean fraction of communication links masked during adversarial training in the 1o_10b_vs_1r scenario, as shown in Table 2 in Section 5.2. It corresponds to 23% retained links and characterizes the generated training topologies, rather than measured communication cost as the initial manuscript can mislead. The generator and the corresponding mask is removed during execution. In the revision, we remove the numerical communication-cost description from the Abstract and Introduction and describe these results explicitly as training-time mask sparsity in Section 5.2.

W2/Q3. About performance gain against TGCNET with random dropout: Considering that the 91.2% accuracy TGCNET achieves on MMM2 has left little margin for improvement, PatMask’s additional improvement to 95.2%, 4 percentage points, can be reasonably considered a valuable improvement. The 91.2% accuracy is the compound effect of TGCNet’s mechanism of dynamic directed communication and training-time graph coarsening, along with the additional random dropout. To assess PatMask’s adversarial masking contribution more directly, Table 6 compares masking strategies within the same backbone. Adversarial-only training achieves 93.8% versus 87.5% for random-only training on MMM2, and 90.9% versus 83.7% on 27M_vs_30M. Hybrid training further improves these results to 97.9% and 95.5%, respectively. These comparisons support the value of learned perturbations alongside random failure exposure under the tested settings.

Q1. Clarifying the random masking setting and its sensitivity in Table 6: Random training mask in Table 6 uses a link-drop probability of 0.5. Note that Table 6 was not mainly intended to compare the approaches with random masking and adversarial masking. We mainly aim to show the orchestration effect of both masking (Hybrid vs. other masking settings). As expected and the results show, random masking can directly improve performance (Random-only vs. No Masking). However, in practice, test-time disconnection probability is usually unknown in advance. Therefore, it can be common to use unaligned training-time and test-time dropout rates, especially when the test-time dropout rate may shift. We are pleased to see that the adversarial masking not only indicates a better replacement for random masking, but can also collaborate with random masking to further improve performance (Adversarial-only and Hybrid vs. Random-only). In the revision, we add clarification of the link-drop probability setting in the caption of Table 6 and this benefit of PatMask in Section 5.5.``Ablation Study->Masking Strategy''.

To better compare the influence of the dropout rate of random masking, we add the sensitivity study of the dropout rates of the random masking when it complements the baselines in Section 5.5.``Ablation Study->Random Training Masking''. Table 7 shows the results of different baselines trained with different random link-drop rates and tested under the mixed disconnection. Training with random link drops can directly improve each baseline's performance, but the dropout rate does not ensure a consistent magnitude of the improvement. PatMask, which combines adversarial training with random masking, outperforms all baselines with random masking across various dropout rates. Comparing each baseline at its best tested rate, TGCNET, trained with a dropout rate of 0.5, is the strongest at 90.5%, while PatMask surpasses it by 4.7 pp, achieving 95.2%. These results indicate that the learned mask does not randomly drop links but identifies specific links to retain or drop during training to tolerate diverse communication disconnections during execution.

Table 7. Effect of random training mask on MMM2, varying the training link-drop rate. Results are test win rates (%) of training baselines with various random dropout rates (q) on the communication links on MMM2 and testing under the mixed disconnection with p=0.5p=0.5 independently for nodes and links. q=0q=0 disables link dropout during training and corresponds to the default baselines..

Methodq=0q=0.3q=0.5q=0.7
GTDE+dropout@q36.9 ± 8.160.9 ± 3.668.4 ± 6.057.6 ± 6.8
TGCNET+dropout@q58.1 ± 13.082.6 ± 6.390.5 ± 5.787.6 ± 6.8
T2MAC+dropout@q55.8 ± 22.670.8 ± 17.275.6 ± 16.477.4 ± 15.7
GACG+dropout@q42.7 ± 17.861.5 ± 14.767.1 ± 13.565.2 ± 13.8
PatMask (adversarial and random masking)95.2 ± 1.695.2 ± 1.695.2 ± 1.695.2 ± 1.6

Thank you for the detailed responses. I have no further questions and have raised my soundness and confidence scores.

Reviewer_BAbJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies method for robust learning in Multi-Agent Reinforcement Learning under Communication Disconnection. The authors assume that under attacks some communication toopologies might be changed and thus learning with the fixed topology during training is no longer useful. The proposed approached is therefore consider masking of some edges in the communication topology during training to make the learning more robust. Masking is not random but based of prediction of which edge contribute most to the learning and so masking is done on edges that impact the learning a lot. Experiments compared against baseline approaches show that the proposed masking approach help the model learn more robust model and perform better under different communication disconnection.

Critical strengths

The paper was well written, the proposed approach was supported with theoretical results even the convergence is not guaranteed. The comparison to baseline support the usefulness of learning with adversarial masking even there needs stronger experiments to judge the difference between adversarial masking and random masking.

Critical weaknesses

The work is a based on an assumption that random communication can be disconnected and thus the learning is designed to deal with such situation. That limits the application of the work in practice because the probability of disconnection is not known in advance and also the chance that may happen is highly dependent on applications. Such artificial difficulty may lead to a new problem to solve but whether it is a common issue in practice is unknown.

The idea is similar to dropout in ML learning and thus benchmarking with careful dropout is necessarily to support the proposed adversarial masking. I saw in the experiments the baselines are not trained with drop-out and the appendix does compare to dropout results but the chosen dropout hyperparameters were fixed. Therefore I would recommend to do the following experiments: carefully finetune hyperparameters of the dropout for each baseline methods and report new results of the baseline with dropout

Questions for authors

Please look at the weaknesses.

3 comments

Thanks very much for your valuable comments. Our responses and revisions are provided as follows. We hope that we have well addressed all your concerns and welcome further discussion if necessary. We appreciate your dedicated work on the review and will be grateful if you could raise your rating in response to our rebuttal.

W1. About practical consideration of the problem: Thank you for raising the concern about the application of PatMask in practical situations. We need to clarify that PatMask does not require the deployment disconnection probability to be known in advance in agent execution, as already specified in the Introduction and Section 5.1.Experimental Setup. Communication masks are only used during training to augment the system to tolerate potential disconnection of various conditions, while agent execution uses whatever messages are available, without the generator or a failure-probability estimate. We agree that failure frequency and structure depend on the application. While we can enumerate all real-life disconnection modes, we deem it reasonable to evaluate the robustness of PatMask under various disconnection types and probabilities during execution in our experiments. The results in Section 5.3 demonstrate that PatMask can generalize to various execution-time disconnection conditions without retraining the model. After the agents are trained under harsh communications created by the mask generator, the information about the disconnection type or probability does not need to be provided to the deployed agents.

We have also already discussed the limitation of the synthetic communication disconnection modes in Appendix D.

W2. About baselines with finetuned dropout: Thanks for your suggestion of finetuning the dropout for each baseline method. In the revision, we train the baselines with random communication masking under various link-drop probabilities test them under various disconnection modes. The results are shown in Tables 7 and 8 and discussed in Section 5.5 “Ablation Study->Random training masking” in the main text. Table 7 shows the results of different baselines trained with different random link-drop rates and tested under the mixed disconnection. Training with random link drops can directly improve each baseline's performance, but the dropout rate does not ensure consistent magnitude of the improvement. PatMask, which combines adversarial training with random masking, outperforms all baselines across various dropout rates. Comparing each baseline at its best tested rate, TGCNET, trained with a dropout rate of 0.5, is the strongest at 90.5%, while PatMask surpasses it by 4.7 pp, achieving 95.2%. We further compare PatMask with the baselines trained with a random link drop under different test-time disconnection modes, with results shown in Table 8. PatMask outperforms the baselines with random masking in all three disconnection modes. These results indicate the benefit of adversarial training in PatMask over the random dropout of links. The learned mask does not randomly drop links but identifies specific links to retain or drop during training to tolerate diverse communication disconnections during execution.

Table 7. Effect of random training mask on MMM2, varying the training link-drop rate. Results are test win rates (%) of training baselines with various random dropout rates (q) on the communication links on MMM2 and testing under the mixed disconnection with p=0.5p=0.5 independently for nodes and links. q=0q=0 disables link dropout during training and corresponds to the default baselines..

Methodq=0q=0.3q=0.5q=0.7
GTDE+dropout@q36.9 ± 8.160.9 ± 3.668.4 ± 6.057.6 ± 6.8
TGCNET+dropout@q58.1 ± 13.082.6 ± 6.390.5 ± 5.787.6 ± 6.8
T2MAC+dropout@q55.8 ± 22.670.8 ± 17.275.6 ± 16.477.4 ± 15.7
GACG+dropout@q42.7 ± 17.861.5 ± 14.767.1 ± 13.565.2 ± 13.8
PatMask (adversarial and random masking)95.2 ± 1.695.2 ± 1.695.2 ± 1.695.2 ± 1.6

Table 8. Effect of random training mask on MMM2, varing the test communication disconnection modes. Results are test win rates (%) of training baselines with a random link-drop rate of 0.5 and testing under various disconnection modes on MMM2. Node-only and link-only disconnections each uses a failure probability of p=0.5, and the mixed disconnection applies both failures independently.

MethodNodeLinkMixed
GTDE+dropout@q=0.570.6 ± 4.988.6 ± 4.468.4 ± 6.0
TGCNET+dropout@q=0.591.6 ± 4.591.8 ± 4.190.5 ± 5.7
T2MAC+dropout@q=0.580.8 ± 12.879.6 ± 12.175.6 ± 16.4
GACG+dropout@q=0.572.0 ± 10.571.6 ± 9.867.1 ± 13.5
PatMask94.6 ± 1.494.3 ± 1.395.2 ± 1.6
Reviewer_BAbJOpenReview ↗

Some more related works comparison are needed, more careful benchmarking is required.

Thank you for the clarification and additional comparison with random dropout. I think the proposed idea is similar to the following works and thus additional comparisons and discussion is required.

  • Robust Multi-agent Communication Based on Decentralization-Oriented Adversarial Training https://arxiv.org/pdf/2504.21278 The proposed idea DMAC (2025) is very similar to the PatMask idea where " Specifically, we train an adversary DMAC Adv which can dynamically identify and mask the critical communication channels, and then apply the adversarial samples generated by DMAC Adv to the adversarial learning of the communication policy to force the policy in exploring other potential communication schemes and ransition to a decentralized structure"

  • Learning multi-agent communication via graph contrastive learning, https://www.sciencedirect.com/science/article/abs/pii/S0031320325007538?utm_source=chatgpt.com proposed methods for importance-based edge masking, but primarily to learn better representations rather than simulate worst-case failures but still can be considered as a potential baseline.

I saw that in the appendices, even the authors have compared their proposed methods with another older related work like ADMAC (2024) but my concern is that the significant gap is very large, while the reported baselines number is around 10% the PatMask results is 97%. The experiments was done on 1 environment and it is unclear whether the comparison is fair with proper implementation of the baselines and HPO. The authors should work carefully on the benchmark setting to make sure that the large gaps in the performance is explainable.

Thanks for recognizing our effort in addressing the baselines with fine-tuned dropouts. As for the new concerns raised:

1) Relationship to DMAC. DMAC is related to our study in that they share the general idea of using an adversarial learned mask in MARL, and we explicitly acknowledge it as prior work on adversarial communication-channel masking in the revision. However, our study distinguish with DMAC in several major aspects, including the following design choices and their supporting evidence.

First, PatMask explicitly combines learned adversarial masks with random masks. This combination provides training exposure to both learner-dependent disruptions and additional randomly sampled topologies. Its contribution is directly supported by our existing ablation:

Training masksMMM227M_vs_30M
Random-only87.583.7
Adversarial-only93.890.9
Hybrid97.995.5

Under mixed disconnection with p=0.5, hybrid training improves over adversarial-only training by 4.1 and 4.6 percentage points, respectively. Adversarial-only training also improves over random-only training by 6.3 and 7.2 points. Thus, the results support contributions from both learned masking and its combination with random masking. Our empirical argument extends beyond the observation that adversarial channel masking is useful.

Second, the adversary uses a different learning objective and optimization procedure. DMAC formulates masking as a multi-agent RL problem, using an adversarial reward that balances reduced team return against the number of masked channels. PatMask instead trains a shared mask generator through a differentiable relaxation, maximizing the host learner’s TD error while alternating with learner updates. This directly couples mask generation to the learner’s current value-estimation difficulty and avoids introducing a separate adversarial value-learning problem.

Third, our evaluation specifically examines transfer across communication-availability failures. DMAC’s reported attack evaluations concern learned adaptive and heuristic message perturbations. Our evaluation considers link, node, mixed, observation-conditioned, and regional disconnections. These results support PatMask’s intended use under heterogeneous availability loss.

2) Relationship to MAGE: We agree that MAGE is relevant but explicitly different. Its masking preferentially removes less important edges to preserve informative communication, whereas PatMask learns disruptive masks that increase the learner’s TD error, training coordination under missing information. Similar to the discussion above about DMAC, our ablation supports this failure-oriented training: adversarial-only masking improves over random-only masking by 6.3/7.2 percentage points, and hybrid masking increases these gains to 10.4/11.8 points on MMM2/27M_vs_30M. These results support PatMask’s design, while a direct comparison with MAGE remains an evaluation limitation.

In the revision, we explicitly discuss the relationship of DMAC and MAGE with our study in the related work section.

3) Interpretation of the large gaps compared with ADMAC. ADMAC estimates message reliability and reduces the influence of potentially harmful messages. PatMask instead removes communication links during training and updates the coordination policy under the resulting missing information. Thus, reliability-based aggregation and topology-failure training address different requirements. The former (ADMAC) requires identifying untrustworthy information, while the latter (PatMask) requires learning effective decisions when information is unavailable. ADMAC’s poor performance may reflect a mismatch between its message-reliability defense and our mixed link/node-disconnection setting, where reweighting received messages cannot compensate for missing coordination information.

The approximately 10% vs. 95% comparison measures differences between complete methods under mixed link/node disconnection. Our controlled experiments provide a more informative decomposition that helps to better explain such a significant performance gap.

i) Failure exposure contributes substantially: random masking raises the same backbone from 38.5% to 87.5% on MMM2.

ii) Learned masks provide an additional benefit: adversarial-only training reaches 93.8%, and hybrid training reaches 97.9%.

iii) The advantage persists against dropout-trained baselines: the strongest tested TGCNET dropout configuration reaches 90.5±5.7%90.5\pm5.7\%, compared with PatMask’s 95.2±1.6%95.2\pm1.6\% in that comparison.

We understand that multiple experiment runs are important for result significance. We have compared PatMask’s performance with other recent related baselines in the main results in a setting of multiple runs. The single-run ADMAC (and other methods) results are provided as supplementary evidence of PatMask’s advantage in the appendix.

Comments on the paper

1 comment

Thanks for the valuable review comments, which can help us improve the final manuscript. As the discussion period is ending soon, we would like to summarize what we have done in the revision to facilitate your judgment of our response.

  1. We added evaluation results and analysis of baseline training at various random dropout rates, and tested them under various disconnection modes. PatMask still shows a significant performance advantage over these baselines across various dropout rates and test-time disconnections. Revisions can be found in Section 5.5. "Ablation Study->Random training masking," and Tables 7 and 8.

  2. We have clarified the random masking setting in Table 6 and the implementation details around Eq. 6.

  3. We have clarified the concerns about a practical issue of disconnection probability mismatch between training and test.

  4. We have further discussed the distinction of our work with some related studies.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.