acceptodds
Under review as a conference paper at ICLR 2027

When Does Derivative Supervision Help? A Study of Label Construction and Noise

Abstract

Derivative labels describe local structure that value observations alone may not provide. Their usefulness, however, depends on how that structure is observed. We study label construction and loss balancing under a shared neural training procedure, then connect the results to controlled noise experiments. Across discontinuous-payoff problems, construction contributes more to value and derivative-error rankings than balancing, while the preferred construction reflects the information retained by the label. Varying value and derivative labels separately shows that the advantage of smoothed finite-difference (fuzzy) labels comes from their derivatives rather than from smoothing the value labels. Matched noise experiments reveal where useful derivative supervision becomes harmful to value prediction. A finite-design risk decomposition explains the trade-off between value-noise reduction, derivative-noise cost and approximation-bias change; its Fourier specialization shows how representation and sampling shape this balance. Combining a measured clean-label benefit with a last-layer estimate of the noise cost predicts whether derivative supervision still helps on new configurations for similar targets. Scientific extensions demonstrate value-prediction gains in a stochastic-volatility barrier problem and Burgers learning with a Fourier neural operator. Together with comparisons of higher-order labels, derivatives outside the value training region, and molecular objectives, these results establish a practical priority: determine what derivative information is supplied and how reliably it constrains the prediction before optimizing its weight.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Derivative supervision, Differential machine learning, Sobolev training, Loss balancing, Label construction, Noisy labels, Regularization, Automatic Adjoint Differentiation, Greeks

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_YWDKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates when derivative supervision improves neural network predictions, focusing on derivative-label construction, loss balancing, and noise sensitivity. The authors compare different derivative estimators and weighting strategies, develop a finite-design risk decomposition to explain the effects of approximation bias and noise, and propose a last-layer-based rule to predict the usefulness of derivative supervision. Extensive experiments are conducted across synthetic functions and scientific applications, including financial models, PDEs, molecular modeling, and atmospheric reconstruction.

Critical strengths

Understanding when and why derivative supervision improves prediction is an important problem in scientific machine learning. While derivative-based supervision has been widely used in applications such as PDE learning, differential machine learning, and molecular modeling, its effectiveness is not always guaranteed. This paper addresses a meaningful question by investigating the conditions under which derivative information improves or degrades prediction performance.

Critical weaknesses
  1. The PDE experiments do not sufficiently distinguish additional derivative information from the effect of a modified training objective.

In the Burgers FNO experiment, derivative labels are constructed by applying finite-difference operators to the reference numerical solution. The same operators are applied to the predicted field. Consequently, the derivative loss can be expressed as

[ \mathcal L_g =|D_h\hat U-D_hU|^2 =|D_h(\hat U-U)|^2. ]

Therefore, when the solution is accurately predicted at every grid point, the corresponding finite-difference derivatives are automatically recovered. The constructed derivative labels do not provide independent information beyond the complete discrete solution field. Instead, they introduce an additional structured penalty on the prediction error, which can be interpreted as a discrete Sobolev-type loss or a frequency-dependent reweighting of the error.

Although the authors include a matched-observation experiment comparing derivative supervision with additional neighboring value observations, this does not fully distinguish the benefit of derivative information from the optimization effects induced by the modified loss.

As a result, the PDE experiments primarily demonstrate the effectiveness of a particular derivative-weighted training objective, rather than establishing the value of independently acquired derivative information. This distinction should be made more explicit, particularly given the paper's emphasis on derivative-label construction.

  1. Insufficient consideration of spatial resolution, observation sparsity, and PDE solution regularity.

The reliability of finite-difference derivative labels depends strongly on the underlying grid resolution, observation noise, and regularity of the target solution. Even for smooth solutions, finite-difference approximations introduce discretization errors, while numerical differentiation can amplify observation noise.

However, the PDE experiments primarily rely on numerical solution fields from PDEBench, where neighboring grid values are available for derivative construction. This assumption may not hold in practical settings with sparse or irregularly sampled observations, where the required finite-difference stencils cannot be directly constructed.

Furthermore, the Burgers experiments consider a viscous equation, for which solutions possess relatively favorable regularity under suitable conditions. The situation becomes substantially different for inviscid Burgers equations, where shock discontinuities may develop and classical pointwise derivatives may cease to exist. In such cases, derivative supervision based on squared pointwise gradient errors may not be an appropriate formulation.

These issues raise questions about the broader applicability of the PDE results. The paper would benefit from a systematic investigation of resolution dependence, sparse observations, and the regularity assumptions underlying derivative supervision.

  1. Limited connection between the theoretical analysis and practical derivative-label construction.

The theoretical risk decomposition is informative under its stated assumptions, but the exact analysis primarily concerns fixed feature representations and specified noise models. In contrast, finite-difference derivative labels in PDE applications are constructed from the same numerical observations as the value labels, potentially introducing correlated errors, discretization bias, and resolution-dependent uncertainty.

Although the paper discusses general label covariance, it does not establish a sufficiently direct connection between these practical sources of derivative-label error and the proposed predictive framework.

Moreover, the last-layer prediction rule requires an empirically measured clean-label benefit, obtained by training a derivative-supervised model with clean labels. Thus, the proposed rule evaluates how an already measured benefit deteriorates under additional noise, rather than providing a general criterion for determining whether an arbitrary derivative-label construction is reliable or useful in the first place.

This limits the practical scope of the claimed predictive capability.

  1. The manuscript suffers from excessive experimental accumulation, unclear organization, and presentation issues.

The paper contains a large number of experiments spanning different tasks, architectures, derivative constructions, weighting strategies, and evaluation metrics. However, the overall narrative is difficult to follow, and the logical connections between different experimental sections are not always sufficiently clear.

The presentation often relies on extensive numerical comparisons, error ratios, and rankings, without enough intuitive visualization of the underlying learning phenomena. For instance, the PDE experiments would benefit from direct visual comparisons of reference solutions, model predictions, and spatial error distributions, allowing readers to understand which structures are improved by derivative supervision.

Some figures also contain noticeable formatting issues. In Figure 3, the subfigure title "a Burgers FNO" is repeated twice in the left panel. Such errors suggest that the figures and overall manuscript require more careful proofreading and refinement.

Additionally, Table 1 aggregates a wide variety of heterogeneous experiments and evaluation metrics into a single dense table, making it difficult to extract the main conclusions. The extensive reporting of numerical results, combined with verbose descriptions, further obscures the central scientific contributions.

I recommend substantially streamlining the experimental presentation, improving figure quality, and establishing a clearer progression from the central research questions to the theoretical analysis and experimental evidence.

Questions for authors
  1. In the Burgers FNO experiments, derivative labels are constructed directly from numerical solutions using finite differences. How do the authors distinguish the benefit of derivative supervision from that of a discrete Sobolev-type or frequency-weighted loss?

  2. How sensitive are the PDE results to grid resolution and observation sparsity? Would the conclusions still hold when only sparse or irregularly sampled observations are available?

  3. How does the approach generalize to PDE solutions with limited regularity, such as inviscid Burgers equations with shocks, where classical pointwise derivatives may not exist?

  4. Since the proposed prediction rule requires measuring the clean-label benefit beforehand, how can it be applied when only noisy or biased derivative estimates are available?

  5. Could the authors improve the organization and presentation of the experiments, particularly by providing more intuitive visualizations of the predicted solutions and the effects of derivative supervision?

Reviewer_WPXZOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies when derivative supervision improves value prediction, separating the effects of label construction, loss balancing, and derivative-label noise. It conducted controlled comparisons of derivative estimators, crossed experiments disentangling derivative information from value-label smoothing, and a finite-design risk decomposition separating approximation-bias changes, value-noise reduction, and derivative-noise costs. It also proposes a conditional usefulness predictor combining a measured clean-label benefit with a frozen-last-layer noise-cost estimate. It conducted experiments on discontinuous-payoff problems, synthetic noise sweeps, higher-order and support experiments, and scientific applications involving PDE fields, molecular energies and forces, and atmospheric reconstruction.

Critical strengths
  1. The motivation is clear. This topic has cross-disciplinary impact and addresses a fundamental research question.

  2. The study provides a coherent, qualified account of both benefits and failures.

  3. The experiments are comprehensive. Strong controls isolate what derivative supervision contributes.

Critical weaknesses
  1. The headline comparison between construction and balancing is sensitive to task composition. Six of the nine core blocks are dimensions of one basket family, and several construction comparisons include almost-everywhere-zero pathwise labels that omit the relevant boundary sensitivity.

  2. The usefulness predictor is more conditional than its headline accuracy might suggest. The 88.3% accuracy concerns seed-averaged condition means within three familiar target families, after measuring the clean-label benefit for each design using independently evaluated clean models. It therefore does not directly establish usefulness for a single new training run or for settings without accurate clean references.

  3. The code is planned for release at camera-ready, so the reproducibility remains unclear.

  4. The presentation could be improved: several key assumptions are not sufficiently explicit or justified, and the organization and formatting make the paper difficult to follow.

Questions for authors
  1. Could the authors compare LRM and fuzzy derivatives with identical raw value labels, using a small validation-selected sweep of fixed derivative weights that includes zero?

  2. Does applying the predictor require access to clean derivative labels and an accurate estimate of the derivative-noise scale?

1 comment
Reviewer_WPXZOpenReview ↗

response to rebuttal

Thank you for the clarifications and additional experiments. The comparison using identical raw value labels and validation-selected weights directly addresses my main concern, and the revised manuscript makes the clean-reference requirement of the noise-tolerance analysis clear. I have decided to maintain my original rating of 3 (Weak Accept).

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

The paper is an empirical study of training networks on function values together with their derivatives at the same inputs. It asks when the derivative labels help. The core comparison is a network trained on values alone against the same network trained on values plus derivatives, run mostly on synthetic targets where exact derivatives exist and label quality can be controlled. On Monte Carlo option payoffs with discontinuities, the authors compare three ways of building the derivative label and several loss weightings, and find the construction matters more than the weighting. On smooth synthetic targets they add noise to exact derivatives and locate the level at which the supervision stops helping; a bias–variance calculation for linear least squares is used to interpret this, and a rule derived from it predicts the sign of the effect on new synthetic configurations. The same comparison is then applied to PDE fields, molecular forces, option surfaces and atmospheric reconstruction with task-specific models.

Critical strengths

The paper answers two practical questions in derivative training. First, when a payoff is discontinuous, the standard trick is to smooth it before differentiating, which changes both value and derivative labels. The paper shows the gain comes from the derivatives, which in practice means you can keep the unbiased value labels and smooth only to get the derivatives.

Second, finite-difference derivatives are built from neighboring grid points, so it was reasonable to suspect that a derivative loss amounts to training on more of the field. The paper shows that a model given those same points as plain value labels still does a third worse, so the derivative term adds something beyond denser data.

Critical weaknesses

Most of the findings are expected. When a step payoff is differentiated along sampled paths, the derivative label is zero almost everywhere. Training on it tells the network the expected payoff is constant, which is false, so the fit ends up worse than with values alone, and changing the loss weight does not help. Noisy derivative labels stop helping once the noise is large enough, and the linear calculation that locates the threshold is the standard bias–variance expansion. Training molecular models on energies and forces together is established practice, and the weighting the paper tests is the SchNetPack default. The option-payoff comparison builds on Glasserman and Karmarkar (2025). They already compared the same three label types on the same contracts, so the novelty here is the loss weighting and the crossed control. The new tool, a rule that predicts whether added derivative noise will tip the result, requires the clean derivative model to be trained first. It therefore answers a narrower question than the title asks and it is tested only on synthetic families. The closing advice, to check what a derivative label contains before tuning its weight, is sensible but is not a finding.

The main text does not accurately reflect the applied results in the appendix. For the barrier with likelihood-ratio labels and one of the option-surface folds, derivative supervision is several times worse than training on values alone (Tables 8 and 53), while the main text describes the first as “does not improve” and does not mention the second. On the Burgers MLP the result flips with the checkpoint rule from several times worse to better on average (Table 45), described in the main text as less consistent gains. The paper’s analysis implies that with exact value labels any benefit must come from how the network approximates the target rather than from averaging out label noise. What that means for a trained network is not examined.

Questions for authors
  • With exact value labels the decomposition attributes any benefit to the approximation term. Is there an experiment that separates finite-sample constraint geometry from optimization effects in the network case?
  • Why does the validation monitor flip the Burgers MLP result (Table 45)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.