acceptodds
Withdrawn from ICLR 2027

SnareNet: Flexible Repair Layers for Neural Networks with Hard Constraints

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural networks are increasingly used as fast surrogate models across various domains, but unconstrained predictions can violate physical, operational, or safety requirements. We propose SnareNet, a feasibility-controlled architecture to learn mappings whose outputs must satisfy input-dependent constraints. SnareNet appends a differentiable repair layer that navigates in the constraint map’s range space, steering iterates toward feasibility and producing a repaired output that satisfies constraints to a user-specified tolerance. We stabilize end-to-end training by adaptive relaxation, a new training paradigm that snares the neural network at initialization and shrinks it into the feasible set, enabling early exploration and strict feasibility later in training. On optimization learning and trajectory planning benchmarks, SnareNet consistently attains improved objective quality while satisfying constraints more reliably than prior work, and it is the first to enforce non-convex constraints at medium-to-high precision robustly across instances.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Constraints, Constrained Deep Learning, Enforcement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HCQMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SnareNet, an end2end neural-based framework to ensure strict feasibility. The work uses a differentiable repair layer to make the output of neural network within the feasible region. Meanwhile, it proposes a technique called adaptive relaxation (AdaRel), to guide the early exploration of the neural network by relaxing feasibility requirements during the initial training phase. Experimental results show that SnareNet works well in constraint optimization problems and trajectory planning benchmarks.

Critical strengths
  1. This paper studies the constraint satisfactory in convex and non-conex problem, which is important, and through regularized adaptive Newton update to implement repair layer. The paper is very well-writen and easy to follow.

  2. The theoratical analysis shows the convergence of SnareNet in linear, convex, and non-convex constraint problem.

Critical weaknesses
  1. The same repair technique seems already used in a published paper, HardNet++: Nonlinear Constraint Enforcement in Neural Networks [1], which uses the formula yk+1HardNet++=yk−Jc,k⊤(Jc,kJc,k⊤+ϵI)−1[r(yk)]y_{k+1}^{\text{HardNet++}} = y_k-J_{c,k}^\top(J_{c,k}J_{c,k}^\top+\epsilon I)^{-1}[r(y_k)]. Additionally, the proposed training method AdaRel is more like an engineering optimization.

  2. Another main concern is about the scalability. The results in App. H show the memory bottleneck in large problems, which is predictive since it uses an iterative matrix computation and also as the analysis in Sec. 4.4. Although the authors gives some explanation, the reviewer think is a key problem, since many real-world physical problems, such as the optimal power flow problem in electrical systems mentioned in the author's introduction, often involve thousands of variables and constraints.

[1] Goertzen, Andrea, et al. "HardNet++: Nonlinear Constraint Enforcement in Neural Networks." IEEE Control Systems Letters (2026).

Questions for authors
  1. Could the authors clarify the difference between SnareNet and HardNet++ and compare their performance?
  2. Could the authors further explan that why iterative methods like conjugate gradient got divergence in SnareNet?
  3. Could the AdaRel method be generalized to other iterative-based repairation method? Additionally, could the authors analyze the sensitivity of λ\lambda in convex/non-convex QCQPs?
Reviewer_G2oUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a feasibility-controlled neural network architecture whose output must satisfy input-dependent constraints. The authors also present a new training scheme that stabilizes end-to-end training using adaptive relaxation, which induces the neural network output into the feasible set. This stabilization technique encourages early exploration and strict feasibility satisfaction. The main idea is to devise a differentiable repair layer that projects the neural network output onto the intersection of the feasible regions of the nonlinear constraints and the box constraints using Gauss-Newton method.

Critical strengths
  • The motivation that unconstrained or soft-constrained predictions can violate physical, operational, or safety requirements is solid.
  • Problem formulation and notations are easy to follow.
  • The proposed method covers non-convex constraints, which is more general than convex constraints.
  • The experiment section provides comprehensive empirical evaluations, which are easy to understand the strength of the method in comparison with other baseline methods.
Critical weaknesses

Readability

  • It is hard to understand what Figure 1-(b) is telling about adaptive relaxation.
  • Training time in Fig. 3 is hard to read.

Correctness

  • Referring the update rule in Section 4.1 as a Newton iteration is a misnomer. Technically, the update rule is Gauss-Newton step, since it does not involve exact second derivatives (Hessian).

Contribution

  • Despite the empirical efficacy, the adaptive relaxation approach appears to be an engineering effort.
  • It is hard to recognize the benefit of the adaptive relaxation method in Table 2.
Questions for authors
  • Please see the weaknesses section.
  • Why is pronunciation of y^\hat{y} instructed in Section 2?
  • Why does DC3 collapse in Section 5.2?
  • What are the feasibility tolerance values for other baseline methods in Section 5?
  • How does the proposed method compare to the QCQP-net [1] in the QCQP problem class?
  • How does the proposed method compare to the differentiable QP layers in convex QP problem class?

[1] Zeng, Sihan, et al. "QCQP-Net: Reliably learning feasible alternating current optimal power flow solutions under constraints." 6th Annual Learning for Dynamics & Control Conference. PMLR, 2024.

Reviewer_BH9HOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SnareNet, a differentiable repair layer for neural networks with input-dependent constraints. The layer uses regularized Newton updates to reduce constraint violations. The authors also introduce adaptive relaxation, which gradually tightens constraints during training. The paper provides convergence analysis for linear and certain convex constraints and evaluates the method on optimization learning and simulated control tasks.

Critical strengths
  1. SnareNet improves objective quality and constraint satisfaction on several tested problems. Reporting both average and maximum constraint violations is useful, since good average performance can hide failures on individual instances.

  2. The convergence results clarify the behavior under linear and convex constraints. The experiments on regularization and warm-up strategies also help explain how the repair process affects the final solution quality.

Critical weaknesses

My main concern is that SnareNet makes only a small change to an established framework. DC3 and FSNet already combine neural predictions with differentiable feasibility correction and end-to-end training. SnareNet retains this basic approach, with the main additions being a regularized Jacobian-based repair step and a gradually tightened constraint schedule. In my view, these are relatively straightforward changes to the repair solver and training procedure, rather than a substantial new methodological contribution.

The paper reports improvements on the tested benchmarks, but these results do not sufficiently establish the value of the proposed changes over well-configured existing methods. Although the authors study regularization and warm-up strategies, it remains unclear whether similar gains could be obtained by using a stronger repair solver and the same relaxation schedule within an existing framework. Comparisons under matched training and computational budgets would help answer this question. Given the small methodological difference, I would expect stronger evidence of a clear advantage in solution quality, feasibility, or efficiency to justify the contribution.

Questions for authors

none

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

Related papers

Discussion (0)

Sign in to comment.