acceptodds
Withdrawn from ICLR 2027

Auxiliary Location Inference for Recurrent Sparse Field Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse field reconstruction with uncertain sensor locations must distinguish prediction errors from spatial misalignment, often with only some vector components observed. We develop calibrated auxiliary location inference with sensor-block cross-fitting and incorporate its probabilities consistently into a kernel correction. On three cylinder cases, the recurrent implementation reduces aggregate NRMSE by 3.40–5.25% over the original auxiliary procedure, retaining 1.63–2.29% in a second observation batch. Matched reference comparisons and component ablations separate the roles of the auxiliary field, likelihood adjustment and stage assignment. We then connect location probabilities to the implemented field response through the observation operator, working covariance, solve and postprocessing. Controlled interventions identify two breaks in this link: physical closure can reverse a raw-field gain, and held-sensor scores can favor corrections that increase full-field error. Cross-physics evaluations show how the measured gains depend on the scene and reference quality.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific machine learning, sparse field reconstruction, sensor location uncertainty, neural operators, data assimilation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CHP3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies recurrent sparse-field reconstruction when sensor locations are uncertain and only some vector components are observed. It introduces a cross-fitted, calibrated auxiliary location-weight generator, inserts its probabilities consistently into a shared kernel correction, and studies stage assignments in which the original generator A or new generator B is used for model/configuration selection and final fitting.

Critical strengths
  • The distinction between location-probability quality and final-field quality is conceptually important.
  • Cross-fitting seems to be useful in addressing the reuse bias.
Critical weaknesses
  • The public neural comparisons assess complete systems and cannot isolate the contribution of auxiliary location inference, since the adapters do not share AB’s kernel correction and selection machinery. If I understand it correctly, the paper partly addresses this with matched reference-weight and GP-moment comparisons, plus a limited conditional-mixture diagnostic and a NIGP-inspired adaptation. I would like a fuller recurrent comparison against averaging per-location reconstructions, and a comparison that changes only the uncertain-input treatment while holding the downstream solver and selection procedure fixed.
  • The evidence for selecting AB over BB is weak. AB was chosen using development-exposed cylinder cases, and its small advantage over BB in the original window reverses in the later, nonoverlapping window: BB wins five of six case–backbone groups, and AB fails the paper’s preregistered later-window criterion.
  • The probability-to-field analysis validates how changes in location weights propagate through the reconstruction, including tests of frozen response-accuracy predictions. It would be more useful if the authors can validate a prospective prediction of reconstruction benefit. The failure explanations rely on target-assisted diagnostics, and the paper offers no demonstrated observation-only criterion for deciding when AB will outperform the alternatives. This limits the analysis’s value for choosing a method on a new task.
Questions for authors
  • Where is the matched baseline that exactly or by Monte Carlo marginalizes over the supplied discrete location support while sharing AB’s references, kernel, candidate grid, selector, closure, cap, and recurrence?
  • What auxiliary-inference claim, if any, is supported by the gaps to RecFNO, FLRONet, Senseiver, and PalSB? These are complete-system comparisons with different machinery and training assets, and the public adapters’ convergence sensitivity remains unresolved.
  • What are confidence intervals when physical cases or trajectories are the sampling units, without treating horizons, observation seeds, backbones, conditions, or overlapping windows as independent?
  • How much selection optimism is introduced by choosing among 109 configurations from sparse held-sensor scores, and where is a nested-selection or multiplicity-matched control?
  • On a physical panel untouched during method development, how does frozen AB compare with its matched alternatives? Separately, how does it behave, or detect failure, when the true sensor site lies outside the supplied candidate support?
Reviewer_49cqOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors reconstruct a physical field from sparse sensors whose locations are uncertain. A neural operator predicts a reference field, a kernel correction fits the sensors, and probabilities over the candidate sensor locations enter this correction. The contribution is a modification of an existing estimator of these location probabilities and a recipe for which of the two estimators to use in which stage of the pipeline.

Critical strengths

I had a hard time following this paper. It reads as an engineering report on one reconstruction pipeline rather than a paper written for a general machine learning audience, so I cannot say much in its favor.

Critical weaknesses
  • The problem is not motivated. Outside the related-work section, the main text cites no prior work at all, so the general claims in the introduction are asserted without support, and no application is named.

  • It is unclear what the methodological contribution is.

  • The evidence is thin and the abstract overstates it. The main results are gains of a few percent on three cases that were used during development, reported without confidence intervals.

  • The computational cost is also not characterized beyond the timing of a first step, so it is unclear how the method scales.

  • The paper is very hard to read. Many sentences cannot be parsed, for example "held-sensor scores can favor corrections that increase full-field error" in the abstract, terms are used before they are defined, and the appendices run to 37 pages. Half of the main text is the experiments section while the introduction and related work together take less than a fifth. This space would be better spent on motivating the problem and giving a general reader the domain context.

Questions for authors
  • What is the general lesson for a reader who does not use this pipeline?
Reviewer_HwxaOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses an important problem in spatiotemporal modeling systems where the location of sensors may be uncertain. In this context, this paper proposes an algorithm to learn the auxiliary location associated with a set of measurements. Not only do they propose algorithmic techniques, the paper also introduces a range of different evaluation methodologies to test the robustness of their approach and existing approaches using controlled interventions (that impact sensor readings), sensor hold outs and also perform component ablations to separate the roles of the auxiliary field, likelihood adjustment and stage assignment.

Critical strengths
  1. Interesting problem formulation
  2. Robust evaluation methodology to test for a lot of simple but important hypotheses which typical ML papers typically skip. I appreciate the fact that the authors have carefully thought about some of these cases.
Critical weaknesses

(i) I view the problem to be fairly narrow in scope for an ICLR paper submission and ICLR audience. (ii) The algorithmic contribution and novelty is also quite limited from the viewpoint of a conventional ICLR paper. (iii) Evaluation (while the methodology is good) is also limited both in terms of comparison techniques and in terms of datasets. This is an area where there is extensive methodologies and the bar for evaluation is quite high.

Questions for authors

I encourage the authors to rethink the scope of the problem and the associated evaluations in a future version of the paper if the goal is to publish in a top-tier ML conference.

a. Expand the scope of the problem formalism. I view this as a narrow part of a much larger problem. Think of a large-scale scenario where location information is lacking or uncertain in a large spatiotemporal dataset. How would you use your algorithm to estimate the auxiliary location?

b. Expand the scope of your algorithm: Once you estimate the auxiliary locations in a large-scale dataset, how does it help in larger scale inference in that problem domain. What is the "impact" of your solution.

c. Expand the scope of your related work analysis and baselines: For a paper in such a crowded space, the onus is on the authors to back up their algorithms with realistic comparisons and baselines.

d. Try to demonstrate theoretical guarantees on the location inference of your algorithms under the requisite assumptions which need careful thought.

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

Related papers

Discussion (0)

Sign in to comment.