arxiv: 2605.09337 · v1 · submitted 2026-05-10 · 💻 cs.LG · math.OC

Recognition: no theorem link

Adversary-Robust Learning from Fully Asynchronous Directional Derivative Estimates

Anik Kumar Paul , Nibedita Roy , Nagesh Talagani , Swetha Ganesh , Gugan Thoppe , Alexandre Reiffers-Masson

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:22 UTC · model grok-4.3

classification 💻 cs.LG math.OC

keywords asynchronous distributed optimizationadversary-robust learningsign-based updatesdirectional projectionstwo-timescale methodsnonconvex convergencezeroth-order optimization

0 comments

The pith

Signed directional projections with two-timescale updates let fully asynchronous systems converge to stationary points despite adversaries.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops FAR-SIGN to let parameter-server systems run robust optimization when workers execute fully asynchronously and some may act adversarially. It replaces full gradients or function values with sign information taken along carefully chosen directions, then corrects the resulting bias through a two-timescale averaging rule. The approach works for both gradient-based and derivative-free versions and removes the need for any trusted reference dataset at the server. A reader would care because it removes synchronization barriers and extra data requirements while still guaranteeing almost-sure convergence on smooth nonconvex problems at near-optimal iteration complexity.

Core claim

FAR-SIGN achieves almost-sure convergence to the set of stationary points for smooth nonconvex objectives by combining sign-based updates along designed directions with a two-timescale mechanism that offsets bias from asynchrony and compression; the method admits first-order and zeroth-order realizations, runs without a private reference dataset, and attains rates of O(n^{-1/4+ε}) in the first-order case and O(n^{-1/6+ε}) in the zeroth-order case.

What carries the argument

Signed directional projections paired with a two-timescale bias-correction rule, which converts delayed and compressed information into unbiased progress toward stationarity.

If this is right

Distributed training can proceed without global synchronization barriers or a trusted reference dataset at the server.
Both first-order and zeroth-order versions converge almost surely on smooth nonconvex problems.
The first-order version reaches near-optimal rates of O(n^{-1/4+ε}) while the zeroth-order version reaches O(n^{-1/6+ε}).
Empirical wall-clock performance improves over robust aggregation baselines on image-classification tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Removing the reference-dataset requirement could simplify deployment in large-scale federated or edge-learning deployments.
The same two-timescale correction might be reusable with other forms of gradient compression beyond the sign operator.
The framework suggests a general template for converting asynchronous, compressed updates into provably convergent algorithms for nonconvex objectives.

Load-bearing premise

The two-timescale averaging rule offsets bias from delays and sign compression under the given adversary model without any trusted reference data.

What would settle it

A concrete run of FAR-SIGN on a smooth nonconvex test function in which the iterates fail to approach stationary points or the observed convergence rate is slower than O(n^{-1/4+ε}) under fully asynchronous execution with adversarial workers.

Figures

Figures reproduced from arXiv: 2605.09337 by Alexandre Reiffers-Masson, Anik Kumar Paul, Gugan Thoppe, Nagesh Talagani, Nibedita Roy, Swetha Ganesh.

read the original abstract

We propose FAR-SIGN (Fully Asynchronous Robust optimization via SIGNed directional projections) for adversary-resilient learning in parameter-server--worker systems. FAR-SIGN achieves robustness through sign-based updates along carefully designed directions and mitigates the resulting bias via a two-timescale mechanism. It admits both first-order and zeroth-order implementations and enables fully asynchronous execution without requiring a private reference dataset at the server. We establish almost-sure convergence of FAR-SIGN to the set of stationary points for smooth, nonconvex objectives. Moreover, we prove the near-optimal rate of $O(n^{-1/4+\epsilon})$ in the first-order setting and the standard $O(n^{-1/6+\epsilon})$ in the zeroth-order setting, where $n$ is the iteration count and $\epsilon>0$ can be chosen arbitrarily small. Experiments on MNIST show that FAR-SIGN outperforms robust aggregation-based methods in both accuracy and wall-clock time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

FAR-SIGN gives a reference-free async robust method with claimed rates, but the bias correction under arbitrary delays needs verification before the guarantees can be taken at face value.

read the letter

The main point is that FAR-SIGN runs fully asynchronously in a parameter-server setup, uses sign-based directional projections for adversary robustness, and applies a two-timescale mechanism to remove bias from asynchrony and sign compression. It does this without any private reference dataset at the server and claims almost-sure convergence to stationary points for smooth nonconvex objectives, plus rates of O(n^{-1/4+ε}) in the first-order case and O(n^{-1/6+ε}) in the zeroth-order case.

Referee Report

2 major / 2 minor

Summary. The paper proposes FAR-SIGN, a fully asynchronous robust optimization algorithm for parameter-server--worker systems that uses signed directional projections along carefully chosen directions combined with a two-timescale mechanism to mitigate bias from asynchrony and sign compression. It claims to achieve this without requiring a private reference dataset at the server. For smooth nonconvex objectives, the authors prove almost-sure convergence to the set of stationary points, with near-optimal rates O(n^{-1/4+ε}) in the first-order setting and O(n^{-1/6+ε}) in the zeroth-order setting (ε>0 arbitrary). Experiments on MNIST demonstrate improved accuracy and wall-clock time over robust aggregation baselines.

Significance. If the convergence and rate results hold under the stated model of fully asynchronous execution and arbitrary adversaries, the work would be significant: it removes the need for a reference dataset while delivering practical robustness and competitive rates. The two-timescale plus directional-projection design is a concrete technical contribution that could influence distributed learning systems. The MNIST experiments provide supporting evidence of wall-clock gains, though broader validation would strengthen the practical claim.

major comments (2)

[§4, Theorem 1] §4 (Convergence Analysis), specifically the bias-cancellation argument underlying Theorem 1 and the rate proofs: the two-timescale mechanism and directional projections are asserted to keep the effective gradient estimator unbiased (or with controllable bias) for arbitrary delay sequences and any subset of adversarial workers. However, the provided analysis sketch does not explicitly derive a bound that remains valid when delays are unbounded and adversary timing is completely arbitrary; if the proof implicitly relies on a positive fraction of synchronized honest workers or a delay bound, the central guarantee does not follow from the model stated in the abstract and §2. This is load-bearing for both the a.s. convergence and the stated rates.
[§3.2] §3.2 (Algorithm Description) and the zeroth-order implementation: the directional projection choice is presented as removing sign-induced bias without extra assumptions, but the paper does not provide an explicit comparison (e.g., via a counter-example or lemma) showing that the same bias term would remain uncontrolled under standard sign compression without the projection. This directly affects whether the O(n^{-1/6+ε}) rate is achieved under the fully asynchronous adversary model.

minor comments (2)

[§5] The experimental section reports only MNIST; adding at least one additional non-convex task (e.g., CIFAR-10 or a simple NLP model) would make the practical claims more robust.
[§3] Notation for the two timescales (e.g., the step-size sequences α_t and β_t) is introduced without an explicit table summarizing their relative scaling; a small table would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review. The two major comments raise important points about the clarity and completeness of the bias analysis. We address each below, indicating the revisions we will make to strengthen the presentation without altering the claimed results.

read point-by-point responses

Referee: [§4, Theorem 1] the bias-cancellation argument underlying Theorem 1 and the rate proofs: the two-timescale mechanism and directional projections are asserted to keep the effective gradient estimator unbiased (or with controllable bias) for arbitrary delay sequences and any subset of adversarial workers. However, the provided analysis sketch does not explicitly derive a bound that remains valid when delays are unbounded and adversary timing is completely arbitrary; if the proof implicitly relies on a positive fraction of synchronized honest workers or a delay bound, the central guarantee does not follow from the model stated in the abstract and §2.

Authors: We appreciate the referee drawing attention to the need for explicit verification of the bias bound. The full proof in Appendix A.2 derives the bound on the effective estimator bias using a two-timescale supermartingale argument that holds for any sequence of finite (but possibly unbounded) delays and arbitrary adversarial subsets, without assuming a positive fraction of synchronized honest workers. The directional projections are constructed so that their conditional expectation aligns with the true gradient independently of delay realizations. We will expand the sketch in §4 to include the key steps of this argument (specifically, the decomposition in Eq. (12) and the application of Lemma 4), making the independence from delay bounds explicit. This is a clarification only; the stated a.s. convergence and rates remain unchanged. revision: partial
Referee: [§3.2] the directional projection choice is presented as removing sign-induced bias without extra assumptions, but the paper does not provide an explicit comparison (e.g., via a counter-example or lemma) showing that the same bias term would remain uncontrolled under standard sign compression without the projection. This directly affects whether the O(n^{-1/6+ε}) rate is achieved under the fully asynchronous adversary model.

Authors: We agree that an explicit comparison would improve the exposition. In the revised manuscript we will add a short lemma (new Lemma 2 in §3.2) and a brief counter-example remark showing that plain sign compression without the directional projection produces a non-vanishing bias term whose magnitude depends on the delay distribution; under fully asynchronous execution this bias prevents the O(n^{-1/6+ε}) rate from holding. The specific projection directions in FAR-SIGN cancel this term in expectation, which is what enables the stated zeroth-order rate. We will also cross-reference the new lemma in the rate proof in §4. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; rates follow from standard stochastic approximation

full rationale

The paper states that almost-sure convergence to stationary points and the rates O(n^{-1/4+ε}) (first-order) and O(n^{-1/6+ε}) (zeroth-order) are established for smooth nonconvex objectives under the FAR-SIGN algorithm. These claims rest on the two-timescale mechanism and directional projections mitigating asynchrony and sign bias. No equations or sections are provided that reduce the convergence result to a self-definition, a fitted parameter renamed as prediction, or a load-bearing self-citation whose own justification is internal. The derivation is presented as applying standard stochastic approximation tools to the proposed updates, which is self-contained against external benchmarks and does not match any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on standard assumptions of smoothness and non-convexity of the objective plus the existence of suitable directional projections that the two-timescale mechanism can correct; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)

domain assumption The objective function is smooth and nonconvex.
Explicitly stated as the setting for which almost-sure convergence to stationary points is proved.

pith-pipeline@v0.9.0 · 5484 in / 1387 out tokens · 48125 ms · 2026-05-12T04:22:50.119367+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

201 extracted references · 201 canonical work pages · 2 internal anchors

[1]

International conference on machine learning , pages=

The hidden vulnerability of distributed learning in byzantium , author=. International conference on machine learning , pages=. 2018 , organization=

work page 2018
[2]

Foundations and Trends in Optimization , volume=

Gradient-based algorithms for zeroth-order optimization , author=. Foundations and Trends in Optimization , volume=. 2025 , publisher=

work page 2025
[3]

Proceedings of the 41st International Conference on Machine Learning , pages=

Revisiting zeroth-order optimization for memory-efficient LLM fine-tuning: a benchmark , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

work page
[4]

arXiv preprint arXiv:1906.06629 , year=

Robust federated learning in a heterogeneous environment , author=. arXiv preprint arXiv:1906.06629 , year=

work page arXiv 1906
[5]

IEEE Transactions on Information Theory , volume=

Data encoding for byzantine-resilient distributed optimization , author=. IEEE Transactions on Information Theory , volume=. 2020 , publisher=

work page 2020
[6]

Foundations of Computational Mathematics , volume=

Random gradient-free minimization of convex functions , author=. Foundations of Computational Mathematics , volume=. 2017 , publisher=

work page 2017
[7]

IEEE Transactions on Information Theory , volume=

Optimal rates for zero-order convex optimization: The power of two function evaluations , author=. IEEE Transactions on Information Theory , volume=. 2015 , publisher=

work page 2015
[8]

Foundations and Trends in Machine Learning , year=

Distributed optimization and statistical learning via the alternating direction method of multipliers , author=. Foundations and Trends in Machine Learning , year=

work page
[9]

Proceedings of the IEEE , year=

Consensus and cooperation in networked multi-agent systems , author=. Proceedings of the IEEE , year=

work page
[10]

Langley , title =

P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

work page 2000
[11]

Brockett, R. W. , journal =. Least squares matching problems , volume =

work page
[12]

International Conference on Machine Learning and Data Mining in Pattern Recognition , pages=

Optimization for large-scale machine learning with distributed features and observations , author=. International Conference on Machine Learning and Data Mining in Pattern Recognition , pages=. 2017 , organization=

work page 2017
[13]

IET Control Theory & Applications , volume=

Combining federated learning and control: A survey , author=. IET Control Theory & Applications , volume=

work page
[14]

arXiv preprint arXiv:2103.12840 , year=

A survey of distributed optimization methods for multi-robot systems , author=. arXiv preprint arXiv:2103.12840 , year=

work page arXiv
[15]

IEEE Transactions on Signal Processing , volume=

Distributed signal processing and optimization based on in-network subspace projections , author=. IEEE Transactions on Signal Processing , volume=. 2020 , publisher=

work page 2020
[16]

Abstraction and Control for Groups of Robots , volume =

Belta, Calin and Kumar, Vijay , journal =. Abstraction and Control for Groups of Robots , volume =

work page
[17]

The computer journal , volume=

An automatic method for finding the greatest or least value of a function , author=. The computer journal , volume=. 1960 , publisher=

work page 1960
[18]

IEEE Signal Processing Magazine , volume=

A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications , author=. IEEE Signal Processing Magazine , volume=. 2020 , publisher=

work page 2020
[19]

2013 , publisher=

Algorithms for minimization without derivatives , author=. 2013 , publisher=

work page 2013
[20]

A Distributed Optimization Framework for Localization and Formation Control: Applications to Vision-Based Measurements , volume =

Tron, Roberto and Thomas, Justin and Loianno, Giuseppe and Daniilidis, Kostas and Kumar, Vijay , journal =. A Distributed Optimization Framework for Localization and Formation Control: Applications to Vision-Based Measurements , volume =

work page
[21]

Distributed Subgradient Methods for Multi-Agent Optimization , volume =

Nedic, Angelia and Ozdaglar, Asuman , journal =. Distributed Subgradient Methods for Multi-Agent Optimization , volume =

work page
[22]

and Spletzer, John R

Derenick, Jason C. and Spletzer, John R. , journal =. Convex optimization strategies for coordinating large-scale robot formations , volume =

work page
[23]

Notarstefano, Giuseppe and Notarnicola, Ivano and Camisa, Andrea , journal =

work page
[24]

and Spesivtsev, Leonid and Pappas, George J

Zavlanos, Michael M. and Spesivtsev, Leonid and Pappas, George J. , booktitle =

work page
[25]

Kia, Solmaz S. and. IEEE Control Systems Magazine , number =

work page
[26]

Distributed online optimization in dynamic environments using mirror descent , volume =

Shahin Shahrampour and Ali Jadbabaie , journal =. Distributed online optimization in dynamic environments using mirror descent , volume =

work page
[27]

2019 12th Asian Control Conference (ASCC) , title=

Y. 2019 12th Asian Control Conference (ASCC) , title=. 2019 , volume=

work page 2019
[28]

2009 , publisher=

Convex Optimization Theory , author=. 2009 , publisher=

work page 2009
[29]

Online Learning and Online Convex Optimization , title=

S. Online Learning and Online Convex Optimization , title=. 2012 , volume=

work page 2012
[30]

IEEE Transactions on Automatic Control , title=

Y. IEEE Transactions on Automatic Control , title=. 2018 , volume=

work page 2018
[31]

An Introduction to Optimization on Smooth Manifolds , publisher=

Boumal, Nicolas , year=. An Introduction to Optimization on Smooth Manifolds , publisher=

work page
[32]

IEEE Transactions on Control of Network Systems , title=

M. IEEE Transactions on Control of Network Systems , title=. 2017 , volume=

work page 2017
[33]

Graph Theoretic Methods in Multiagent Networks , year =

Mehran Mesbahi and Magnus Egerstedt , edition =. Graph Theoretic Methods in Multiagent Networks , year =

work page
[34]

2004 , isbn =

Shnayder, Victor and Hempstead, Mark and Chen, Bor-rong and Allen, Geoff Werner and Welsh, Matt , title =. 2004 , isbn =. doi:10.1145/1031495.1031518 , booktitle =

work page doi:10.1145/1031495.1031518 2004
[35]

2016 American Control Conference (ACC) , title=

S. 2016 American Control Conference (ACC) , title=. 2016 , volume=

work page 2016
[36]

2008 47th IEEE Conference on Decision and Control , title=

A. 2008 47th IEEE Conference on Decision and Control , title=. 2008 , volume=

work page 2008
[37]

Sundhar and A

Ram, S. Sundhar and A. Nedi \'c and Veeravalli, V. V. Incremental stochastic subgradient algorithms for convex optimization. SIAM Journal on Optimization. 2009. doi:10.1137/080726380

work page doi:10.1137/080726380 2009
[38]

Sundhar Ram and A

S. Sundhar Ram and A. Nedi \'c and Veeravalli, V. V. Distributed Stochastic Subgradient Projection Algorithms for Convex Optimization. Journal of Optimization Theory and Applications. 2010. doi:10.1007/s10957-010-9737-7

work page doi:10.1007/s10957-010-9737-7 2010
[39]

Introductory lectures on stochastic optimization , author=

work page
[40]

Joint and Separate Convexity of the

Bauschke, Heinz and Borwein, Jonathan (Jon) , year =. Joint and Separate Convexity of the

work page
[41]

and Johnson, Charles R

Horn, Roger A. and Johnson, Charles R. , title =. 2012 , isbn =

work page 2012
[42]

Williams, David , biburl =

work page
[43]

, title =

Meyn, Sean and Tweedie, Richard L. , title =. 2009 , isbn =

work page 2009
[44]

, title =

Polyak, B.T. , title =

work page
[45]

1976 , type =

Rudin, Walter , title =. 1976 , type =

work page 1976
[46]

IEEE Signal Processing Magazine , title=

A. IEEE Signal Processing Magazine , title=. 2020 , volume=

work page 2020
[47]

Proceedings of the 25th International Conference on Machine Learning , pages =

Duchi, John and Shalev-Shwartz, Shai and Singer, Yoram and Chandra, Tushar , title =. Proceedings of the 25th International Conference on Machine Learning , pages =. 2008 , isbn =

work page 2008
[48]

Bubeck, Sébastien , title =

work page
[49]

and Ghosh, Joydeep , title =

Banerjee, Arindam and Merugu, Srujana and Dhillon, Inderjit S. and Ghosh, Joydeep , title =. J. Mach. Learn. Res. , month = dec, pages =. 2005 , issue_date =

work page 2005
[50]

T. T. IEEE Control Systems Letters , title=. 2019 , volume=

work page 2019
[51]

, biburl =

Kolmogorov, Andrey N. , biburl =. Foundations of the Theory of Probability , url =

work page
[52]

Distributed Optimization for Control , volume =

Nedic, Angelia and Liu, Ji , year =. Distributed Optimization for Control , volume =. Annual Review of Control, Robotics, and Autonomous Systems , doi =

work page
[53]

Johansson

Tao Yang and Xinlei Yi and Junfeng Wu and Ye Yuan and Di Wu and Ziyang Meng and Yiguang Hong and Hong Wang and Zongli Lin and Karl H. Johansson. A survey of distributed optimization. Annual Reviews in Control. 2019. doi:https://doi.org/10.1016/j.arcontrol.2019.05.006

work page doi:10.1016/j.arcontrol.2019.05.006 2019
[54]

D. K. IEEE Transactions on Smart Grid , title=. 2017 , volume=

work page 2017
[55]

IEEE Transactions on Information Theory , title=

R. IEEE Transactions on Information Theory , title=. 1975 , volume=

work page 1975
[56]

The Annals of Mathematical Statistics , number =

Herbert Robbins and Sutton Monro , title =. The Annals of Mathematical Statistics , number =. 1951 , doi =

work page 1951
[57]

IEEE 14th International Conference on Control and Automation (ICCA) , title=

Y. IEEE 14th International Conference on Control and Automation (ICCA) , title=. 2018 , volume=

work page 2018
[58]

Optimization Letter , volume =

Jueyou Li and Guoquan Li and Zhiyou Wu and Changzhi Wu , title =. Optimization Letter , volume =

work page
[59]

SIAM Review , volume =

Blair, Charles , title =. SIAM Review , volume =. 1985 , doi =

work page 1985
[60]

Operation Research Letter , month = may, pages =

Beck, Amir and Teboulle, Marc , title =. Operation Research Letter , month = may, pages =. 2003 , issue_date =

work page 2003
[61]

IEEE Transactions on Information Theory , title=

G. IEEE Transactions on Information Theory , title=. 2015 , volume=

work page 2015
[62]

2015 IEEE 6th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP) , title=

M. 2015 IEEE 6th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP) , title=. 2015 , volume=

work page 2015
[63]

J. C. 49th Annual Allerton Conference on Communication, Control, and Computing (Allerton) , title=. 2011 , address=

work page 2011
[64]

IEEE Transactions on Signal and Information Processing over Networks , title=

M. IEEE Transactions on Signal and Information Processing over Networks , title=. 2019 , volume=

work page 2019
[65]

E. C. IEEE Journal of Selected Topics in Signal Processing , title=. 2015 , volume=

work page 2015
[66]

G. S. 2015 53rd Annual Allerton Conference on Communication, Control, and Computing (Allerton) , title=. 2015 , volume=

work page 2015
[67]

Countering Feedback Delays in Multi-Agent Learning , volume =

Zhou, Zhengyuan and Mertikopoulos, Panayotis and Bambos, Nicholas and Glynn, Peter W and Tomlin, Claire , booktitle =. Countering Feedback Delays in Multi-Agent Learning , volume =

work page
[68]

IEEE 56th Annual Conference on Decision and Control (CDC) , title=

Z. IEEE 56th Annual Conference on Decision and Control (CDC) , title=. 2017 , volume=

work page 2017
[69]

, title =

Duflo, Marie and Wilson, Stephen S. , title =. 1997 , isbn =

work page 1997
[70]

IEEE Transactions on Automatic Control , title=

K. IEEE Transactions on Automatic Control , title=. 2017 , volume=

work page 2017
[71]

IEEE Transactions on Automatic Control , title=

P. IEEE Transactions on Automatic Control , title=. 2013 , volume=

work page 2013
[72]

IEEE Journal of Selected Topics in Signal Processing , title=

K. IEEE Journal of Selected Topics in Signal Processing , title=. 2011 , volume=

work page 2011
[73]

Journal of the Operational Research Society , keywords =

Kelly, Frank and Maulloo, Aman and Tan, David , biburl =. Journal of the Operational Research Society , keywords =

work page
[74]

S. H. IEEE/ACM Transactions on Networking , title=. 1999 , volume=

work page 1999
[75]

1986 , publisher=

Principles of Mathematical Analysis , author=. 1986 , publisher=

work page 1986
[76]

and Juditsky, A

Nemirovski, A. and Juditsky, A. and Lan, G. and Shapiro, A. , title =. SIAM Journal on Optimization , volume =. 2009 , doi =

work page 2009
[77]

Improved Algorithms for Convex-Concave Minimax Optimization , volume =

Wang, Yuanhao and Li, Jian , booktitle =. Improved Algorithms for Convex-Concave Minimax Optimization , volume =

work page
[78]

Distributed

Hu, Guoqiang and Pang, Yipeng and Sun, Chao and Hong, Yiguang , journal=. Distributed. 2022 , volume=

work page 2022
[79]

Distributed Generalized Nash Equilibrium Seeking: An Operator-Theoretic Perspective , year=

Belgioioso, Giuseppe and Yi, Peng and Grammatico, Sergio and Pavel, Lacra , journal=. Distributed Generalized Nash Equilibrium Seeking: An Operator-Theoretic Perspective , year=

work page
[80]

Structured Prediction via the Extragradient Method , volume =

Taskar, Ben and Lacoste-Julien, Simon and Jordan, Michael , booktitle =. Structured Prediction via the Extragradient Method , volume =

work page

Showing first 80 references.