HUANet: Hard-Constrained Unrolled ADMM for Constrained Convex Optimization

Binh Nguyen; Trinh Tran; Truong X. Nghiem

arxiv: 2604.13179 · v1 · submitted 2026-04-14 · 🧮 math.OC · cs.LG· cs.SY· eess.SY

HUANet: Hard-Constrained Unrolled ADMM for Constrained Convex Optimization

Trinh Tran , Binh Nguyen , Truong X. Nghiem This is my paper

Pith reviewed 2026-05-10 14:38 UTC · model grok-4.3

classification 🧮 math.OC cs.LGcs.SYeess.SY

keywords HUANetADMM unrollingconstrained convex optimizationhard-constrained networksdifferentiable correctionoptimality conditionsalgorithm accelerationdeep learning for optimization

0 comments

The pith

Unrolling ADMM iterations into a neural network with hard constraint correction and soft optimality penalties solves constrained convex optimization problems.

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

The paper seeks to build a neural network that follows the exact steps of the Alternating Direction Method of Multipliers rather than learning a direct mapping from inputs to outputs. At each unrolled step the network applies a correction that forces equality constraints to hold exactly, and training adds penalties based on first-order optimality conditions to steer the whole structure toward true solutions. A reader would care because many engineering and control tasks require solutions that obey hard rules and reach optimality, yet standard deep networks treat the problem as a black box and often violate those rules.

Core claim

HUANet unrolls the iterations of ADMM into a trainable neural network for solving constrained convex optimization problems. A hard-constrained neural network is embedded at each iteration, equality constraints are enforced by a differentiable correction stage placed at the network output, and first-order optimality conditions are added as soft constraints during training to promote convergence of the unrolled network.

What carries the argument

Unrolled ADMM iterations with an embedded hard-constrained neural network and a differentiable correction stage that enforces equality constraints at each step.

If this is right

The network accelerates standard ADMM while producing feasible points that satisfy equality constraints exactly.
Training with soft optimality penalties guides the unrolled network to fixed points that solve the original optimization problem.
The architecture avoids the black-box nature of end-to-end learning methods by embedding explicit algorithmic structure.
Numerical experiments demonstrate improved performance on a range of constrained convex problems compared with baseline approaches.

Where Pith is reading between the lines

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

The same unrolling-plus-correction pattern could be applied to other first-order methods such as proximal gradient descent.
In settings where constraints represent physical laws, the hard correction stage may reduce the need for post-processing or projection steps.
Because each layer follows a known optimization update, the trained network can be interpreted as an accelerated solver whose iteration count is learned from data.

Load-bearing premise

The differentiable correction stage remains numerically stable and the added soft optimality penalties during training cause the network to converge to solutions of the original constrained problem.

What would settle it

Train HUANet on a set of constrained convex test problems, then evaluate the final outputs on held-out instances and measure the maximum violation of the equality constraints together with the norm of the optimality residual; large persistent violations would show the claim does not hold.

Figures

Figures reproduced from arXiv: 2604.13179 by Binh Nguyen, Trinh Tran, Truong X. Nghiem.

**Figure 1.** Figure 1: A neural layer Fθp of HUANet for an ADMM iteration at step k. III. PROPOSED HUANET FRAMEWORK In this section, we present HUANet, a deep unrolled ADMM framework designed to efficiently solve the constrained optimization problem (1) across varying problem parameters λ. HUANet unfolds the ADMM iterations in (3) into N sequential neural layers, where each layer corresponds to an ADMM iteration and executes a… view at source ↗

**Figure 2.** Figure 2: Training process of HUANet. Black and red arrows [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Solve time of HUANet, OSQP, and Clarabel on the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Average solve time of HUANet and ADMM. nx = 10 4 5 6 7 8 9 Time (seconds) ×10 5 nx = 100 0.0 0.2 0.4 0.6 0.8 1.0 ×10 2 Clarabel OSQP HUANet [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Solve time of HUANet, OSQP, and Clarabel on the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Solve time of HUANet, Clarabel, and SCS on the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

read the original abstract

This paper presents HUANet, a constrained deep neural network architecture that unrolls the iterations of the Alternating Direction Method of Multipliers (ADMM) into a trainable neural network for solving constrained convex optimization problems. Existing end-to-end learning methods operate as black-box mappings from parameters to solutions, often lacking explicit optimality principles and failing to enforce constraints. To address this limitation, we unroll ADMM and embed a hard-constrained neural network at each iteration to accelerate the algorithm, where equality constraints are enforced via a differentiable correction stage at the network output. Furthermore, we incorporate first-order optimality conditions as soft constraints during training to promote the convergence of the proposed unrolled algorithm. Extensive numerical experiments are conducted to validate the effectiveness of the proposed architecture for constrained optimization problems.

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

HUANet unrolls ADMM with a hard constraint correction layer and soft optimality penalties, but the construction does not yet show it reaches the true KKT point of the original problem.

read the letter

The key point is that HUANet unrolls ADMM iterations and adds a differentiable correction to enforce equality constraints hard at each step, plus soft penalties on first-order optimality conditions during training. This setup aims to keep the network feasible while encouraging it to mimic the original optimizer. The architecture is a direct response to the common problem that plain unrolled networks ignore constraints. By baking in a correction stage, the method guarantees satisfaction of the equalities no matter what the network predicts inside. Training with the optimality conditions as soft constraints is a sensible way to regularize the learned parameters toward the true solution. If the experiments back this up, it could be a practical tool for embedding optimization into larger models where constraints cannot be violated. The soft spots center on whether the finite unrolled network actually reaches the same point as classical ADMM. The stress-test concern holds: because the number of iterations is fixed and the correction is part of the learned map, the network could converge to a different fixed point where the soft penalty is small but the original optimality gap remains. Without a proof that the learned iteration stays contractive or numerical evidence that increasing the unroll depth recovers the true solution, the claim that it promotes convergence is not yet supported. The abstract mentions extensive experiments, but the lack of reported numbers, baselines, or stability checks leaves the practical performance unclear. This work is for people building neural networks that solve constrained convex problems, especially in control and signal processing. A reader looking for new ways to enforce hard constraints inside learned optimizers will find the correction layer worth examining. It deserves a serious referee because the idea is concrete and the gap it targets is real, even though more evidence on convergence is needed before it can be recommended.

Referee Report

2 major / 2 minor

Summary. The paper claims to introduce HUANet, a hard-constrained unrolled ADMM network for constrained convex optimization. It unrolls ADMM iterations, embeds a neural network with a differentiable correction stage to enforce equality constraints exactly at each iteration's output, and adds first-order optimality conditions as soft constraints during training to promote convergence. Effectiveness is validated through numerical experiments.

Significance. If the results hold, HUANet would offer a principled way to accelerate constrained optimization solvers using deep learning while maintaining hard constraint satisfaction and approximating optimality conditions. This hybrid approach could be significant for applications in control, signal processing, and machine learning where both speed and constraint adherence are critical. The explicit incorporation of ADMM structure and optimality conditions distinguishes it from purely data-driven methods.

major comments (2)

[Unrolled architecture and training (likely §3)] The incorporation of first-order optimality conditions as soft constraints is intended to promote convergence, but no analysis is provided demonstrating that the finite unrolled network with the differentiable correction stage converges to the true KKT points of the original problem rather than a different fixed point that balances the soft penalty. Since classical ADMM convergence is only asymptotic, this is a load-bearing gap for the central claim.
[Numerical experiments (likely §4 or §5)] Although the abstract states that extensive numerical experiments validate the architecture, the provided text lacks any quantitative results, comparisons to baselines such as standard ADMM or other unrolled methods, error bars, or metrics on constraint violation and optimality gap. This prevents verification of the claimed effectiveness.

minor comments (2)

[Abstract] The abstract is clear but could specify the types of constrained problems tested or key performance gains observed.
[Notation] Ensure consistent use of symbols for the ADMM variables and network parameters throughout the manuscript.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments and the recommendation for major revision. We address each major comment point by point below, indicating whether revisions will be made.

read point-by-point responses

Referee: The incorporation of first-order optimality conditions as soft constraints is intended to promote convergence, but no analysis is provided demonstrating that the finite unrolled network with the differentiable correction stage converges to the true KKT points of the original problem rather than a different fixed point that balances the soft penalty. Since classical ADMM convergence is only asymptotic, this is a load-bearing gap for the central claim.

Authors: We acknowledge that the manuscript does not contain a formal convergence analysis proving that the finite unrolled network reaches the exact KKT points of the original problem. The soft first-order optimality conditions are used as training penalties to encourage the learned parameters to produce iterates closer to optimality, while the differentiable correction layer enforces hard equality constraint satisfaction at every step. Classical ADMM converges only asymptotically, and our finite unrolling with learned components is intended as a practical approximation rather than an exact solver. We will revise the manuscript to add an explicit discussion of this limitation, the potential for other fixed points, and the role of the soft penalties, but we do not claim a full theoretical guarantee. revision: partial
Referee: Although the abstract states that extensive numerical experiments validate the architecture, the provided text lacks any quantitative results, comparisons to baselines such as standard ADMM or other unrolled methods, error bars, or metrics on constraint violation and optimality gap. This prevents verification of the claimed effectiveness.

Authors: The full manuscript contains a numerical experiments section with quantitative comparisons to standard ADMM, other unrolled networks, error bars from repeated trials, and metrics including constraint violation (identically zero due to the hard correction layer) and optimality gaps. We apologize if the version sent to the referee omitted or insufficiently highlighted these results. In the revision we will ensure all quantitative tables, figures, and metrics are clearly presented and expanded where helpful for verification. revision: yes

standing simulated objections not resolved

Rigorous proof that the finite unrolled network with soft optimality penalties converges to the true KKT points rather than an alternative fixed point.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper proposes an unrolled ADMM network with a differentiable correction stage for hard equality constraints and soft first-order optimality penalties in the loss. This is an independent algorithmic construction relying on standard ADMM iteration maps and neural network training procedures. No claimed result reduces by definition to a fitted parameter, self-citation chain, or renamed input; the method remains self-contained with external validation via numerical experiments on constrained convex problems.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

Based solely on the abstract, no explicit free parameters, axioms, or invented entities are stated beyond standard convex optimization assumptions and the differentiability of the correction operator.

axioms (2)

domain assumption The original problem is a constrained convex optimization problem for which ADMM is applicable.
Stated in the abstract as the target class of problems.
domain assumption The correction stage that enforces equality constraints is differentiable.
Required for end-to-end training via backpropagation.

invented entities (1)

HUANet architecture no independent evidence
purpose: Unrolled ADMM network with hard constraint enforcement
New trainable model introduced in the paper.

pith-pipeline@v0.9.0 · 5443 in / 1337 out tokens · 21419 ms · 2026-05-10T14:38:26.988918+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

[1]

D. P. Bertsekas,Constrained optimization and Lagrange multiplier methods. Academic press, 2014

work page 2014
[2]

Spatially temporally distributed informative path planning for multi-robot systems,

B. Nguyen, L. Nguyen, T. X. Nghiem, H. La, J. Baca, P. Rangel, M. C. Montoya, and T. Nguyen, “Spatially temporally distributed informative path planning for multi-robot systems,” in2025 American Control Conference (ACC). IEEE, 2025, pp. 3429–3434

work page 2025
[3]

Algorithms for fitting the constrained lasso,

B. R. Gaines, J. Kim, and H. Zhou, “Algorithms for fitting the constrained lasso,”Journal of Computational and Graphical Statistics, vol. 27, no. 4, pp. 861–871, 2018

work page 2018
[4]

Distributed quadratic programming problems of power systems with continuous and discrete variables,

S.-S. Lin, S.-C. Hornget al., “Distributed quadratic programming problems of power systems with continuous and discrete variables,” IEEE Transactions on Power Systems, vol. 28, no. 1, pp. 472–481, 2012

work page 2012
[5]

Distributed optimization with local domains: Applications in mpc and network flows,

J. F. Mota, J. M. Xavier, P. M. Aguiar, and M. P ¨uschel, “Distributed optimization with local domains: Applications in mpc and network flows,”IEEE Transactions on Automatic Control, vol. 60, no. 7, pp. 2004–2009, 2014

work page 2004
[6]

Constrained portfolio optimisation: The state-of-the-art markowitz models,

Y . Jin, R. Qu, and J. Atkin, “Constrained portfolio optimisation: The state-of-the-art markowitz models,” inInternational Conference on Operations Research and Enterprise Systems, vol. 2. SCITEPRESS, 2016, pp. 388–395

work page 2016
[7]

Distributed optimization and statistical learning via the alternating direction method of multipliers,

S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of multipliers,”Foundations and Trends in Information Re- trieval, vol. 3, no. 1, pp. 1–122, 07 2011

work page 2011
[8]

A survey on some recent developments of alternating direction method of multipliers,

D.-R. Han, “A survey on some recent developments of alternating direction method of multipliers,”Journal of the Operations Research Society of China, vol. 10, no. 1, pp. 1–52, 2022. [Online]. Available: https://doi.org/10.1007/s40305-021-00368-3

work page doi:10.1007/s40305-021-00368-3 2022
[9]

Interior-point methods for opti- mization,

A. S. Nemirovski and M. J. Todd, “Interior-point methods for opti- mization,”Acta Numerica, vol. 17, pp. 191–234, 2008

work page 2008
[10]

arXiv preprint arXiv:2405.12762 (2 024)

P. J. Goulart and Y . Chen, “Clarabel: An interior-point solver for conic programs with quadratic objectives,”arXiv preprint arXiv:2405.12762, 2024

work page arXiv 2024
[11]

Nocedal and S

J. Nocedal and S. J. Wright,Numerical optimization. Springer, 2006

work page 2006
[12]

Learning to optimize: A primer and a benchmark,

T. Chen, X. Chen, W. Chen, H. Heaton, J. Liu, Z. Wang, and W. Yin, “Learning to optimize: A primer and a benchmark,”Journal of Machine Learning Research, vol. 23, no. 189, pp. 1–59, 2022

work page 2022
[13]

Learning optimal solutions for ex- tremely fast ac optimal power flow,

A. S. Zamzam and K. Baker, “Learning optimal solutions for ex- tremely fast ac optimal power flow,” in2020 IEEE international conference on communications, control, and computing technologies for smart grids (SmartGridComm). IEEE, 2020, pp. 1–6

work page 2020
[14]

Fast explicit machine learning-based model predictive control of nonlinear pro- cesses using input convex neural networks,

W. Wang, H. Zhang, Y . Wang, Y . Tian, and Z. Wu, “Fast explicit machine learning-based model predictive control of nonlinear pro- cesses using input convex neural networks,”Industrial & Engineering Chemistry Research, vol. 63, no. 40, pp. 17 279–17 293, 2024

work page 2024
[15]

DC3: A learning method for optimization with hard constraints,

P. L. Donti, D. Rolnick, and J. Z. Kolter, “DC3: A learning method for optimization with hard constraints,” inInternational Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=V1ZHVxJ6dSS

work page 2021
[16]

FSNet: Feasibility-seeking neural network for constrained optimization with guarantees,

H. T. Nguyen and P. L. Donti, “FSNet: Feasibility-seeking neural network for constrained optimization with guarantees,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id= oum1txoy1D

work page 2025
[17]

Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,

V . Monga, Y . Li, and Y . C. Eldar, “Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,”IEEE Signal Processing Magazine, vol. 38, no. 2, pp. 18–44, 2021

work page 2021
[18]

Learning fast approximations of sparse coding,

K. Gregor and Y . LeCun, “Learning fast approximations of sparse coding,” inProceedings of the 27th international conference on international conference on machine learning, 2010, pp. 399–406

work page 2010
[19]

Learning to learn by gradient descent by gradient descent,

M. Andrychowicz, M. Denil, S. Gomez, M. W. Hoffman, D. Pfau, T. Schaul, B. Shillingford, and N. De Freitas, “Learning to learn by gradient descent by gradient descent,”Advances in neural information processing systems, vol. 29, 2016

work page 2016
[20]

ADMM-CSNet: A deep learning approach for image compressive sensing,

Y . Yang, J. Sun, H. Li, and Z. Xu, “ADMM-CSNet: A deep learning approach for image compressive sensing,”IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 3, pp. 521– 538, 2018

work page 2018
[21]

ADMM-ESINet: A deep unrolling network for EEG extended source imaging,

K. Liu, H. Jiang, H. Yang, J. Zhang, Z. Gu, Z. Yu, Y . Zhang, B. Xiao, and W. Wu, “ADMM-ESINet: A deep unrolling network for EEG extended source imaging,”IEEE Journal of Biomedical and Health Informatics, 2025

work page 2025
[22]

Boyd and L

S. Boyd and L. Vandenberghe,Convex optimization. Cambridge university press, 2004

work page 2004
[23]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inInternational Conference on Learning Representations, 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

work page 2019
[24]

Osqp: An operator splitting solver for quadratic programs,

B. Stellato, G. Banjac, P. Goulart, A. Bemporad, and S. Boyd, “Osqp: An operator splitting solver for quadratic programs,”Mathematical Programming Computation, vol. 12, no. 4, pp. 637–672, 2020

work page 2020
[25]

SCS: Splitting conic solver v1.1.0,

“SCS: Splitting conic solver v1.1.0,” https://github.com/cvxgrp/scs, 2015

work page 2015

[1] [1]

D. P. Bertsekas,Constrained optimization and Lagrange multiplier methods. Academic press, 2014

work page 2014

[2] [2]

Spatially temporally distributed informative path planning for multi-robot systems,

B. Nguyen, L. Nguyen, T. X. Nghiem, H. La, J. Baca, P. Rangel, M. C. Montoya, and T. Nguyen, “Spatially temporally distributed informative path planning for multi-robot systems,” in2025 American Control Conference (ACC). IEEE, 2025, pp. 3429–3434

work page 2025

[3] [3]

Algorithms for fitting the constrained lasso,

B. R. Gaines, J. Kim, and H. Zhou, “Algorithms for fitting the constrained lasso,”Journal of Computational and Graphical Statistics, vol. 27, no. 4, pp. 861–871, 2018

work page 2018

[4] [4]

Distributed quadratic programming problems of power systems with continuous and discrete variables,

S.-S. Lin, S.-C. Hornget al., “Distributed quadratic programming problems of power systems with continuous and discrete variables,” IEEE Transactions on Power Systems, vol. 28, no. 1, pp. 472–481, 2012

work page 2012

[5] [5]

Distributed optimization with local domains: Applications in mpc and network flows,

J. F. Mota, J. M. Xavier, P. M. Aguiar, and M. P ¨uschel, “Distributed optimization with local domains: Applications in mpc and network flows,”IEEE Transactions on Automatic Control, vol. 60, no. 7, pp. 2004–2009, 2014

work page 2004

[6] [6]

Constrained portfolio optimisation: The state-of-the-art markowitz models,

Y . Jin, R. Qu, and J. Atkin, “Constrained portfolio optimisation: The state-of-the-art markowitz models,” inInternational Conference on Operations Research and Enterprise Systems, vol. 2. SCITEPRESS, 2016, pp. 388–395

work page 2016

[7] [7]

Distributed optimization and statistical learning via the alternating direction method of multipliers,

S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of multipliers,”Foundations and Trends in Information Re- trieval, vol. 3, no. 1, pp. 1–122, 07 2011

work page 2011

[8] [8]

A survey on some recent developments of alternating direction method of multipliers,

D.-R. Han, “A survey on some recent developments of alternating direction method of multipliers,”Journal of the Operations Research Society of China, vol. 10, no. 1, pp. 1–52, 2022. [Online]. Available: https://doi.org/10.1007/s40305-021-00368-3

work page doi:10.1007/s40305-021-00368-3 2022

[9] [9]

Interior-point methods for opti- mization,

A. S. Nemirovski and M. J. Todd, “Interior-point methods for opti- mization,”Acta Numerica, vol. 17, pp. 191–234, 2008

work page 2008

[10] [10]

arXiv preprint arXiv:2405.12762 (2 024)

P. J. Goulart and Y . Chen, “Clarabel: An interior-point solver for conic programs with quadratic objectives,”arXiv preprint arXiv:2405.12762, 2024

work page arXiv 2024

[11] [11]

Nocedal and S

J. Nocedal and S. J. Wright,Numerical optimization. Springer, 2006

work page 2006

[12] [12]

Learning to optimize: A primer and a benchmark,

T. Chen, X. Chen, W. Chen, H. Heaton, J. Liu, Z. Wang, and W. Yin, “Learning to optimize: A primer and a benchmark,”Journal of Machine Learning Research, vol. 23, no. 189, pp. 1–59, 2022

work page 2022

[13] [13]

Learning optimal solutions for ex- tremely fast ac optimal power flow,

A. S. Zamzam and K. Baker, “Learning optimal solutions for ex- tremely fast ac optimal power flow,” in2020 IEEE international conference on communications, control, and computing technologies for smart grids (SmartGridComm). IEEE, 2020, pp. 1–6

work page 2020

[14] [14]

Fast explicit machine learning-based model predictive control of nonlinear pro- cesses using input convex neural networks,

W. Wang, H. Zhang, Y . Wang, Y . Tian, and Z. Wu, “Fast explicit machine learning-based model predictive control of nonlinear pro- cesses using input convex neural networks,”Industrial & Engineering Chemistry Research, vol. 63, no. 40, pp. 17 279–17 293, 2024

work page 2024

[15] [15]

DC3: A learning method for optimization with hard constraints,

P. L. Donti, D. Rolnick, and J. Z. Kolter, “DC3: A learning method for optimization with hard constraints,” inInternational Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=V1ZHVxJ6dSS

work page 2021

[16] [16]

FSNet: Feasibility-seeking neural network for constrained optimization with guarantees,

H. T. Nguyen and P. L. Donti, “FSNet: Feasibility-seeking neural network for constrained optimization with guarantees,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id= oum1txoy1D

work page 2025

[17] [17]

Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,

V . Monga, Y . Li, and Y . C. Eldar, “Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,”IEEE Signal Processing Magazine, vol. 38, no. 2, pp. 18–44, 2021

work page 2021

[18] [18]

Learning fast approximations of sparse coding,

K. Gregor and Y . LeCun, “Learning fast approximations of sparse coding,” inProceedings of the 27th international conference on international conference on machine learning, 2010, pp. 399–406

work page 2010

[19] [19]

Learning to learn by gradient descent by gradient descent,

M. Andrychowicz, M. Denil, S. Gomez, M. W. Hoffman, D. Pfau, T. Schaul, B. Shillingford, and N. De Freitas, “Learning to learn by gradient descent by gradient descent,”Advances in neural information processing systems, vol. 29, 2016

work page 2016

[20] [20]

ADMM-CSNet: A deep learning approach for image compressive sensing,

Y . Yang, J. Sun, H. Li, and Z. Xu, “ADMM-CSNet: A deep learning approach for image compressive sensing,”IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 3, pp. 521– 538, 2018

work page 2018

[21] [21]

ADMM-ESINet: A deep unrolling network for EEG extended source imaging,

K. Liu, H. Jiang, H. Yang, J. Zhang, Z. Gu, Z. Yu, Y . Zhang, B. Xiao, and W. Wu, “ADMM-ESINet: A deep unrolling network for EEG extended source imaging,”IEEE Journal of Biomedical and Health Informatics, 2025

work page 2025

[22] [22]

Boyd and L

S. Boyd and L. Vandenberghe,Convex optimization. Cambridge university press, 2004

work page 2004

[23] [23]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inInternational Conference on Learning Representations, 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

work page 2019

[24] [24]

Osqp: An operator splitting solver for quadratic programs,

B. Stellato, G. Banjac, P. Goulart, A. Bemporad, and S. Boyd, “Osqp: An operator splitting solver for quadratic programs,”Mathematical Programming Computation, vol. 12, no. 4, pp. 637–672, 2020

work page 2020

[25] [25]

SCS: Splitting conic solver v1.1.0,

“SCS: Splitting conic solver v1.1.0,” https://github.com/cvxgrp/scs, 2015

work page 2015