acceptodds
Under review as a conference paper at ICLR 2027

BOLD: Learning Optimization Objectives via Bilevel Optimization and Symbolic Distillation

Abstract

Many inverse problems are naturally formulated as optimization problems with stable solvers and interpretable objective functions. Recent deep learning approaches have mainly focused on learning optimizer parameters or replacing optimization pipelines with end-to-end black-box models, while the automatic discovery of optimization objectives remains comparatively underexplored. We propose BOLD, a framework for Bilevel Objective Learning and Symbolic Distillation. BOLD parameterizes optimization objectives as decomposable smooth components, trains them through finite-horizon bilevel optimization, and supports symbolic distillation of learned components into analytic expressions. We study three progressively broader instantiations: scalar-objective augmentation for finite-budget sparse recovery, learned local regularization for spring-system structure recovery, and full objective discovery for graph learning from smooth signals. Across these three settings, BOLD demonstrates complementary benefits of learning the objective itself: LP-ALISTA substantially improves finite-budget recovery over its ALISTA backbone; the learned spring-system regularizer can be distilled into an analytic penalty while largely preserving the neural teacher's support-recovery performance; and the learned graph objectives match or outperform classical Kalofolias-style baselines in most representative settings while revealing interpretable signal-distance and degree-shaping structure. Together, these results show that objective learning can improve downstream recovery while preserving an explicit optimization formulation and enabling symbolic interpretation.

Primary area:
optimization
Keywords:
Bilevel Optimization, Unrolled Optimization, Symbolic Regression, Inverse Problems

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_45CuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a framework for learning optimization objective through bilevel optimization. Instead of using a fixed objective function, the method learns neural objective components and trains them by differentiating through the optimization process. The authors also use symbolic regression to convert the learned objectives into interpretable mathematical expressions.

Critical strengths
  1. The idea of learning the optimization objective instead of the optimizer is interesting.
  2. The symbolic distillation part is interesting, as it makes the learned objective more interpretable.
Critical weaknesses
  1. I don't quite get the motivation for using bilevel optimization here. Usually, we use bilevel/differentiable optimization when we know some structure of the problem. However, in this paper, the goal is to learn the optimization objective itself. I don't see clear evidence showing why this is better than directly training a neural network to predict the solution. Could the authors add experimental comparisons with direct prediction and learning-to-optimize (L2O) methods? Physical constraints or physics-based regularization could also be included in these baselines for a fair comparison.
  2. I also feel that the comparison with LISTA/ALISTA is not fully fair. LISTA/ALISTA mainly learns parameters of the optimization updates, such as weights, step sizes, and thresholds, while BOLD can learn a more flexible neural regularizer. Therefore, the improvement might come from the additional modeling capacity rather than learning the objective itself. Could the authors provide a more controlled comparison with similar model capacity and computational budgets?
  3. All experiments are conducted on synthetic datasets. It would be helpful to include some real-world experiments to show the practical value of the proposed method.
Questions for authors

See weaknesses.

2 comments

We thank the reviewer for the constructive feedback. To further address the concerns, we conducted additional experiments.

1. Motivation for Bilevel Objective Learning

We agree that bilevel optimization often exploits known problem structure. If we understand the reviewer's concern correctly, it may relate to the fact that, in many existing applications, the inner objective has a largely prescribed form. A typical bilevel formulation is

\qquad x_\theta^\star\in\arg\min_x E_\theta(x;y),$$ where the inner objective may take the form $$E_\theta(x;y)=\frac12\|Ax-y\|^2+\lambda_\theta R(x).$$ In such settings, learning may focus on regularization weights, prior parameters, or parameters within a predefined regularizer family. BOLD takes a step further by aiming to learn the functional form of the objective itself, rather than only tuning parameters within a fixed functional form. However, learning an unknown objective does not imply that the optimization problem is entirely unstructured. BOLD retains the optimization variables, decomposable factor structure, and inner solver, while learning the functional forms of the scalar objective components. Bilevel optimization enables this by differentiating the recovery loss through the inner solver to update the objective parameters. In this way, the objective is learned directly from its downstream recovery performance. Unlike direct prediction, which learns an input-to-solution mapping, BOLD aims to discover an explicit, decomposable objective that can be independently optimized, analyzed, and further distilled into an interpretable analytic expression. For a more comprehensive comparison, we introduce a direct-prediction multilayer perceptron (MLP) baseline model in the experiments on real-world inverse problems. We also include a learning-based graph recovery baseline model L2G (Pu et al., NeurIPS 2021) for the graph learning experiment. ### 2. Controlled Comparison with Similar Model Capacity We agree that the original ALISTA baseline, with only $2K$ learnable parameters, has substantially lower modeling capacity than LP-ALISTA. This difference is partly by design: while ALISTA primarily learns layerwise step sizes and thresholds, LP-ALISTA introduces the gradient of a learned scalar potential to provide more expressive, state-dependent corrections to the update direction. However, BOLD does not claim that scalar-objective parameterization inherently yields higher reconstruction accuracy than an equally expressive learned update rule. Rather, its central goal is to discover interpretable analytic optimization objectives that retain competitive recovery performance. We constructed a capacity-matched LC-ALISTA variant that directly learns a correction field instead of deriving it from a scalar potential. Both methods share the ALISTA backbone, analytic terms, parameter count, and training protocol. | K / SNR | LC-ALISTA NMSE (dB) ↓ | LP-ALISTA NMSE (dB) ↓ | |---|---:|---:| | 8 / 20 dB | −18.170 | −18.530 | | 8 / 40 dB | −25.397 | −25.301 | | 24 / 20 dB | −20.840 | −20.802 | | 24 / 40 dB | −34.384 | −35.067 | The results show comparable performance between the two approaches. Constraining the correction field to be the gradient of a scalar potential does not inherently guarantee better recovery performance. Nevertheless, compared with directly predicting a correction direction, the scalar-potential formulation provides a more interpretable mathematical structure with greater potential for scientific insight.

3. New Real-World Inverse-Problem Experiments

We evaluated BOLD on two real-world inverse problems and introduced direct-prediction MLP baselines with the same number of trainable neural parameters as BOLD. All neural experiments used three training seeds, the original data splits, and validation-selected checkpoints.

Raman spectral unmixing.

We recover the proportions of five mixture components from measured Raman spectra. BOLD learns unary spectral factors and adjacent-channel coupling factors, and FCLS (Fully Constrained Least Squares) serves as a classical baseline.

MethodTest RMSE (pp) ↓
FCLS4.929
Direct-prediction MLP3.014 ± 0.339
BOLD1.916 ± 0.054
SR-BOLD2.064 ± 0.001

UWB localization.

We estimate 3D positions from measured time-difference-of-arrival (TDOA) observations and surveyed anchor geometry. BOLD learns an exchange-symmetric geometric objective, and LS (Least Squares) serves as a classical baseline by minimizing the squared TDOA residuals.

MethodTest RMSE (m) ↓
LS0.651
Direct-prediction MLP0.955 ± 0.104
BOLD0.440 ± 0.008
Compact SR-BOLD0.452 ± 0.004

Under these experimental settings, BOLD outperforms the tested parameter-matched MLP baselines and classical objective-based methods. Meanwhile, the compact symbolic objectives retain most of the recovery performance of the learned objectives. Further details regarding the actual inverse problem experiments are presented in the revised paper.

Interpretability.

Across three training seeds, the distilled Raman objectives consistently contain squared spectral residuals and log-rational penalties on adjacent-channel differences. The UWB objectives similarly recover TDOA residual penalties and baseline-dependent geometric corrections. These compact expressions retain much of the neural objectives' recovery performance and provide interpretable empirical structures that can be independently optimized. More details are provided in the revised manuscript.

Reviewer_iiAROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents BOLD, a framework for learning optimization objectives through bilevel optimization and subsequently distilling the learned objectives into interpretable symbolic expressions. Instead of learning an optimizer or directly predicting solutions, BOLD parameterizes objective functions using decomposable neural components and optimizes them through downstream task supervision. The authors demonstrate the framework on three problems: sparse recovery, spring-system structure recovery, and graph learning from smooth signals.

Critical strengths

Interesting and potentially impactful research direction. Learning what to optimize, rather than only learning how to optimize, is an appealing direction for inverse problems and scientific machine learning. I appreciate the authors' attempt to preserve an explicit optimization formulation while introducing data-driven flexibility. The possibility of discovering interpretable analytic objectives from supervised data is promising.

Critical weaknesses

Learning regularizers and energy functions through bilevel optimization has been extensively studied. Similarly, symbolic regression and algorithm unrolling are established techniques. The contribution of BOLD appears to lie primarily in integrating these ideas into a unified framework. The authors should more explicitly distinguish their approach from existing learned regularization and variational optimization methods.

While the classical optimization baselines are reasonable, comparisons with more recent learned regularizers, learned proximal operators, or neural energy-based formulations would help establish the advantages of learning explicit decomposable objectives.

Besides, although the chosen benchmarks provide useful controlled evaluations, demonstrating the framework on a real-world inverse problem would considerably strengthen the paper, particularly given its motivation of discovering practical and interpretable optimization objectives.

Questions for authors

What are the main practical advantages of learning a scalar objective compared with learning a direct update rule or a proximal operator, assuming comparable model capacity and training budgets?

Does symbolic distillation provide measurable computational advantages at inference time, beyond interpretability?

1 comment

We thank the reviewer for the constructive suggestions. To further address the concerns, we conducted additional experiments.

1. Distinction from Learned Variational and Proximal Methods

We agree that learning regularizers and variational objectives through bilevel optimization is a well-established research direction. BOLD does not claim novelty from learning scalar energy functions alone. Instead, its core contribution lies in learning shared, decomposable objective components and distilling them into independently executable symbolic objectives, selected based on downstream solver performance.

2. Real-World Inverse-Problem Experiments and More Baselines

We evaluated BOLD on two real-world inverse problems and introduced direct-prediction MLP baselines with the same number of trainable neural parameters as BOLD. All neural experiments used three training seeds, the original data splits, and validation-selected checkpoints. We also added a learned graph-recovery baseline L2G (Pu et al., NeurIPS 2021) in our graph learning experiments.

Raman spectral unmixing.

We recover the proportions of five mixture components from measured Raman spectra. BOLD learns unary spectral factors and adjacent-channel coupling factors, and FCLS (Fully Constrained Least Squares) serves as a classical baseline.

MethodTest RMSE (pp) ↓
FCLS4.929
Direct-prediction MLP3.014 ± 0.339
BOLD1.916 ± 0.054
SR-BOLD2.064 ± 0.001

UWB localization.

We estimate 3D positions from measured time-difference-of-arrival (TDOA) observations and surveyed anchor geometry. BOLD learns an exchange-symmetric geometric objective, and LS (Least Squares) serves as a classical baseline by minimizing the squared TDOA residuals.

MethodTest RMSE (m) ↓
LS0.651
Direct-prediction MLP0.955 ± 0.104
BOLD0.440 ± 0.008
Compact SR-BOLD0.452 ± 0.004

Under these experimental settings, BOLD outperforms the tested parameter-matched MLP baselines and classical objective-based methods. Meanwhile, the compact symbolic objectives retain most of the recovery performance of the learned objectives. Further details regarding the actual inverse problem experiments are presented in the revised paper.

Interpretability.

Across three training seeds, the distilled Raman objectives consistently contain squared spectral residuals and log-rational penalties on adjacent-channel differences. The UWB objectives similarly recover TDOA residual penalties and baseline-dependent geometric corrections. These compact expressions retain much of the neural objectives' recovery performance and provide interpretable empirical structures that can be independently optimized. More details are provided in the revised manuscript.

3. Q1: Scalar Objectives vs. Learned Update Rules

The original ALISTA baseline has only 2K2K learnable parameters, resulting in a substantial capacity difference from BOLD. We understand that the comparison the reviewer has in mind may be between directly learning a correction vector field hϕ(x;y)h_\phi(x;y) and using a correction derived from a scalar potential, ∇xPθ(x;y)\nabla_x P_\theta(x;y), as in LP-ALISTA.

This effectively removes the constraint that the correction field must be the gradient of a scalar potential. We therefore constructed a version of LC-ALISTA with the same parameter count as LP-ALISTA. The results are shown below, where Nx=2000N_x=2000

K / SNRLC-ALISTA NMSE (dB) ↓LP-ALISTA NMSE (dB) ↓
8 / 20 dB−18.170−18.530
8 / 40 dB−25.397−25.301
24 / 20 dB−20.840−20.802
24 / 40 dB−34.384−35.067

The results show comparable performance between the two approaches. Constraining the correction field to be the gradient of a scalar potential does not inherently guarantee better recovery performance. Its advantage lies in retaining competitive performance while providing an explicit mathematical object that can be inspected, analyzed, and further distilled into analytic expressions.

4. Q2: Does Symbolic Distillation Improve Inference Efficiency?

We measured inference latency using the same CPU and 50-step solver configurations, excluding common initialization costs.

CPU benchmarkNeural BOLDCompact SR-BOLD
Raman60.4 ms43.6 ms
UWB58.9 ms27.2 ms

These representative results suggest that compact optimization objectives derived through symbolic distillation can be computationally more efficient than neural-network-based objectives. However, the primary value of symbolic distillation lies in the interpretability of the resulting analytical expressions rather than computational efficiency. Even with a highly compact objective, inherently sequential iterative optimization is generally slower than the GPU-parallelized inference of a direct-prediction neural network.

Reviewer_z9mmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes BOLD: learn components of an optimization objective as shared per-coordinate / per-edge / per-node MLPs, train them by unrolling K steps of the inner solver and backpropagating a supervised task loss, then distill the MLPs into analytic expressions with symbolic regression, selecting candidates by re-running the solver rather than by pointwise fit. Three synthetic instantiations: a learned scalar potential added to ALISTA for sparse recovery, a learned edge penalty for spring-system structure recovery, and a fully learned objective for graph learning from smooth signals compared against Kalofolias.

Critical strengths
  • Table 8 is a good point: picking the symbolic formula by rollout instead of by fit matters a lot (F1 0.62 vs 0.94). I hadn’t seen this made explicit before.
  • Honest paper. Burn-in ablation shows the NaN case, limitations are stated, code is included. Easy to read.
Critical weaknesses
  • I don’t see what is new beyond swapping Kalofolias’s two scalars for MLPs. L2G (Pu et al., NeurIPS 2021) already learns this objective through an unrolled solver with ground-truth graphs; it isn’t cited. Fang et al. 2024 (learned scalar priors used via their gradient) is cited but not discussed.
  • The gain is small. Table 2: +0.014+0.014 F1 on average over Kalofolias, negative in 2/8 settings. The only real wins are BA at N=100N=100, where the log-degree prior is simply wrong for hub graphs (a learned scalar α would likely do the same). SR-BOLD, the actual selling point, is about +0.01+0.01 F1 and usually worse than the teacher on edge/degree error. One seed, no error bars.
  • Missing baselines everywhere: NA-ALISTA / GLISTA (cited, not run), NRI (cited, not run), L2G, and a learned-(α,β\alpha,\beta) Kalofolias through the same solver.
  • “Interpretable” is a stretch. The App. C formulas are long PySR strings with random constants and no common structure across settings. The contextual teacher doesn’t distill (Table 9, 12), and transfer to BA is worse than Kalofolias (Table 13).
Questions for authors
  • Add Kalofolias with α, β learned through the same unrolled projected GD and loss. Does the MLP beat two scalars?
  • Compare to L2G.
  • Seeds / error bars for Tables 1 and 2.
  • How was Kalofolias tuned? Edge ℓ1=1.49\ell_1 = 1.49 at BA+Tikhonov N=100N=100 looks like a missed grid.
  • Why does SR-LP-ISTA beat its teacher at NpN_p = 20?
  • Is PθP_θ ever used as an energy, or only its gradient?
6 comments

We sincerely thank the reviewer for the detailed and constructive feedback. We have conducted additional experiments and will incorporate these results into the revised manuscript.

1. Performance, Novelty, and Motivation of BOLD

BOLD and Kalofolias exhibit mixed performance across graph-learning settings, and SR-BOLD sometimes underperforms its neural teacher. However, BOLD does not aim for absolute performance superiority over carefully designed baselines, but rather to learn explicit, decomposable objectives that retain competitive recovery performance and can be distilled into interpretable analytic expressions.

We appreciate the reviewer's observation regarding the possible mismatch between the log-degree prior and hub-rich BA graphs. This highlights a key motivation for BOLD: hand-designed objectives encode distributional assumptions that may not generalize across settings. BOLD instead learns objective components from downstream recovery supervision, potentially revealing which objective structures are effective for different data distributions. We will also clarify the relationship to L2G (Pu et al., 2021) and Fang et al. (2024) in the revised manuscript.

2. Additional Baselines and Multi-Seed Evaluations

We have completed three-seed evaluations for Tables 1 and 2 and revised the manuscript. We have also added Full L2G (Pu et al., NeurIPS 2021) as a baseline. Preliminary results suggest that L2G performs relatively well on NgN_g=20 RGG graphs but less favorably on BA graphs, while exhibiting severe performance degradation at NgN_g=100. Whether this reflects optimization instability, protocol differences, or method limitations remains unclear. We are further investigating the reproduction and plan to add NA-ALISTA and NRI comparisons to strengthen baseline coverage across the main tasks.

3. Interpretability and Symbolic Complexity

Some expressions in Appendix C are unnecessarily complex, partly due to unsimplified PySR outputs, excessive numerical precision, and insufficient emphasis on the complexity–performance trade-off. Our new experiments on Raman spectral unmixing and UWB localization show that compact symbolic objectives can retain recovery performance close to their neural teachers, with diminishing benefits from increased complexity. Across three random seeds, the distilled expressions exhibit recurring, interpretable structures.

For Raman spectral unmixing, compact SR-BOLD achieves a test RMSE of 2.064 ± 0.001 percentage points, compared with 1.916 ± 0.054 for its neural teacher, 3.014 ± 0.339 for a parameter-matched direct-prediction MLP, and 4.929 for classical FCLS. Across three seeds, the symbolic factors consistently contain squared spectral residuals, wavenumber-dependent corrections, and log-rational penalties on adjacent-channel differences. Representative factors are

ϕ(e,ν)=99.99e2−6.39ν2e,\phi(e,\nu)=99.99e^2-6.39\nu^2e,

ψ(q)=55.13log⁡(1+q2)−51.28q21+q2.\psi(q)=55.13\log(1+q^2)-\frac{51.28q^2}{1+q^2}.

For UWB localization, compact SR-BOLD achieves a test RMSE of 0.452 ± 0.004 m, compared with 0.440 ± 0.008 m for its neural teacher, 0.899 ± 0.019 m for a parameter-matched direct-prediction MLP, and 0.651 m for classical LS, respectively. Across three seeds, the symbolic expressions consistently contain squared TDOA residuals and baseline-dependent geometric corrections, such as

ρ(r,t,bˉ)=1.024r2+(899.4bˉ−1066.7)t.\rho(r,t,\bar b)=1.024r^2+(899.4\bar b-1066.7)t.

We have also identified analytic objectives that balance complexity and recovery performance in the graph-learning experiments.

For NgN_g=20, RGG+Tikhonov, a compact objective achieves test F1 = 0.9481, with

ρ(w,z)=8.857w(z−0.6084),\rho(w,z)=8.857w(z-0.6084),

η(d)=4.086tanh⁡4(d+0.3397).\eta(d)=4.086\tanh^4(d+0.3397).

Here, the edge factor introduces different preferences for increasing or decreasing edge weights depending on the signal distance between nodes, while the node factor provides a bounded, saturating degree penalty.

For NgN_g=20, BA+Tikhonov, another compact objective achieves test F1 = 0.9580, with

ρ(w,z)=8.973w(z−1.049),\rho(w,z)=8.973w(z-1.049),

η(d)=3.039d−7.536tanh⁡(d).\eta(d)=3.039d-7.536\tanh(d).

The derivative of its degree factor is

η′(d)=3.039−7.536sech⁡2(d).\eta'(d)=3.039-7.536\operatorname{sech}^2(d).

Unlike the classical log-degree prior, whose derivative is −α/d-\alpha/d, this learned degree factor has a finite derivative at zero. It encourages degree growth at low degrees, while its marginal contribution gradually transitions into a penalty as the degree increases.

4. Learning Kalofolias Parameters: A Controlled Comparison

We introduced a bilevel-learned Kalofolias baseline, in which the parameters α\alpha and β\beta are trained using the same supervised loss and 200-step projected-gradient descent (PGD) solver as BOLD.

SettingLearned αβ-PGD200Grid-PGD200BOLD-PGD200
Ng20, RGG+Tikhonov0.9080.9060.938 ± 0.004
Ng100, BA+Heat0.8160.8280.901 ± 0.003

The learned two-scalar baseline performs similarly to its grid-tuned counterpart, while BOLD achieves higher F1 scores in both tested settings. This supports the additional value of learning more expressive objective components.

The Grid-PGD200 baseline uses the classical Kalofolias objective but not its original primal–dual solver. The latter is specifically suited to the analytic structure of the Kalofolias objective and is less straightforward to differentiate through in a bilevel optimization setting. We therefore use PGD for both BOLD and this controlled comparison, while Table 2 of the paper reports the classic results of Kalofolias using the original primal–dual solver.

5. Additional Technical Clarifications

How were the Kalofolias hyperparameters tuned?

For each setting, α=1\alpha=1, while β\beta and the support threshold are selected from 19 and 13 candidates, respectively, spanning 10−510^{-5}–10410^4 and 10−1210^{-12}–0.50.5, using validation macro F1.

For Ng=100N_g=100, BA+Tikhonov, a finer search over β\beta and support thresholds reduces test Edge L1 from 1.490 to 1.104 under the original validation F1-based selection criterion, although the error remains relatively high. Switching to validation Edge L1-based selection further reduces test Edge L1 to 0.553, but lowers test F1 to 0.893 (vs. 0.922 under the original protocol). This suggests that the high weighted edge error may be largely attributable to the trade-off induced by the hyperparameter selection criterion.

Why can SR-LP-ISTA outperform its neural teacher?

This is also related to the selection criterion. Symbolic candidates are selected by downstream F1 rather than pointwise fidelity to the teacher. At Np=20,K=100N_p=20,K=100, SR-LP-ISTA improves precision (0.530 → 0.642) but reduces recall (0.793 → 0.687), resulting in higher F1 (0.661 vs. 0.630). This reflects a precision–recall trade-off.

Is the learned scalar function used as an energy or only through its gradient?

This depends on the BOLD instantiation. In graph learning, the learned edge and node factors define a complete scalar energy, directly optimized by PGD, while their distilled expressions form an explicit analytic objective. In sparse recovery, the exact gradient of PθP_\theta contributes to the ALISTA update, which is not standard proximal-gradient descent on a single fixed energy.

We thank the reviewer again for the valuable suggestions. These clarifications and additional results will be incorporated into the revised manuscript.

Thanks for the response and the extra experiments. The seeds and the two real-data tasks help. My main concerns are still open.

  • The comparison in point 4 is not controlled. Table 2 reports Kalofolias at 0.9580.958 F1 (Ng=20Ng=20, RGG+Tikhonov) and 0.8610.861 (Ng=100Ng=100, BA+Heat) with its own solver. The PGD-200 versions in your new table get 0.9060.906–0.9080.908 and 0.8160.816–0.8280.828, so 200 PGD steps lose 0.030.03–0.050.05 F1 on the same objective. BOLD then beats a baseline that has not converged. Against the Table 2 numbers, BOLD is below the two-scalar objective in the first setting (−0.020−0.020, five times your seed SD) and above it in the second (+0.040+0.040). Please run the two-scalar baseline to convergence in all eight settings.
  • With seeds, SR-BOLD’s SD is 0.030.03–0.050.05 F1 in five of eight settings, about ten times the teacher’s, and every bold SR-BOLD F1 is within one SD of its teacher. So “can further improve F1” is not supported. Since the claim is now interpretability and not accuracy, the key question is whether the three seeds give the same functional form in the graph settings. The Raman and UWB formulas are compact and stable, but Appendix C for graphs is unchanged.
  • L2G is still not cited in the revision, and the reproduction is inconclusive. I would need at least the Ng=20Ng=20 numbers in the paper and a paragraph on what BOLD adds over learning this objective through an unrolled solver.
  • NA-ALISTA/GLISTA and NRI are still missing. For Raman, the dataset paper’s own unmixing method is the natural baseline, not only FCLS, Huber and a parameter-matched MLP.

I am keeping my score for now. A converged two-scalar baseline across all settings, L2G in the paper, and per-seed symbolic forms for the graph experiments would change my view.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.