acceptodds
Withdrawn from ICLR 2027

Allocation Entropy for Single-Pass Risk Triage in Neural Operators

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The same predicted field can receive a different risk decision when its allocation is read in another basis. Allocation Uncertainty Decomposition (AUD) makes this representation effect measurable in one pass: it reads threshold entropy from wavelet, Fourier, and branch–trunk components, and energy dispersion from selective updates, without changing the predicted field or training a risk head.For threshold adapters, entropy equals conditional uncertainty in component membership under fixed-threshold logistic perturbations; held-out data connect that structural quantity to physical error. On five frozen Haar-trained models, switching the readout to db2 changes of accept/reject decisions at coverage. Crossed training finds EM matching gains and a Darcy preference for Haar under either training basis. Four electromagnetic adapters rank in-domain error with five-seed mean correlations of –; newly trained MWNO and FNO models retain correlations near – on advection–diffusion.AUD improves screening over random selection in an independent label-budget test, while inexpensive input controls screen more strongly. At batch 96 on one RTX\ 4090, prediction plus AUD takes ms/sample versus for ten-pass Monte Carlo dropout. These results identify representation choice as a measurable axis of single-pass risk triage.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Uncertainty Quantification, Scientific Machine Learning, Selective Prediction, Out-of-Distribution Detection, Conformal Prediction

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

This work proposes AUD, a training-free score that flags neural-operator predictions that are likely to be wrong in the normal forward pass. It works by reading threshold entropy from wavelet, Fourier, and branch–trunk components, and energy dispersion from selective up-dates, without changing the predicted field or training a risk head.

Critical strengths

None

Critical weaknesses

The paper is impossible to follow and needs a major rewrite:

  • The writing is extremely dense and hard to parse. Almost every sentence is qualified several times making hard to see what is actually being claimed
  • Many terms are not introduced (“allocation”, “adapter”, “zero-head readout”, “routed”, “signal ledger”) and standard concepts are relabeled making the text even harder to follow
  • The abstract leads with results before explaining the method, it does not explain what this paper is really about
  • The results sections are mostly numbers without interpretation leaving the reader to interpret by himself
Questions for authors

None

Reviewer_zBb4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes the allocation entropy score to be able to separate unreliable predictions of a neural operator. First, soft masks are obtained by thresholding the components of the operator's internal representation. The score is then the average binary entropy of these masks. The score can be used in deployment for efficiency, where a neural operator produces many cheap predictions, by deciding which ones should be sent to an expensive solver for verification. The fact that the score can be computed in the forward pass without trainable parameters supports the use case. The authors discuss the importance of the choice of basis for the score. They argue that changing the basis influences which predictions are accepted and how much error the accepted ones have. Across multiple operator architectures and PDE tasks, the score beats a random baseline but does not correlate with the test error consistently. The results show that the statistics of the input field select predictions more strongly than the score does.

Critical strengths
  1. The experiments are set up faithfully and the results are reported objectively. Three of the experiments were specified in advance. The paper reports the results that go against the score, in particular that statistics of the input field select predictions better than the allocation entropy does.

  2. Experiments that change only the score are useful. Reading the score from a random network or in another wavelet basis isolates what the score contributes to performance.

Critical weaknesses
  1. The score's utility is not showcased, as it loses to almost every cheaper alternative it is evaluated against. On the electromagnetic scattering task, a ridge regression on three statistics of the input field ranks errors with Spearman correlation 0.58. The allocation entropy reaches 0.46, and adding it to the regression does not prove to be useful. Note that the regression inputs are independent of the model. Two other statistics of the same soft masks beat the allocation entropy on average. A single input statistic and a ridge regression on input statistics beat it in every model (Table 17). On the new advection-diffusion task, the roughness of the input field beats it in all three sets of models (Table 28). A head with ~4000 parameters that predicts the log variance is faster than the score and ranks errors with correlation 0.62 against 0.46 (Table 2). The results question why the score should be used instead of simple statistics that are both cheaper and better at the task.

  2. The results on how the score depends on the wavelet basis do not support the conclusion drawn from them. On five models on the electromagnetic task, the authors find that computing their score in the db2 basis instead of the Haar basis changes 22.5% of the decisions to accept or reject (at 80% coverage). However, while the score in the Haar basis has a Spearman correlation of 0.51 with the test error, that in the db2 basis has only 0.04. The latter is an uninformative score. On the twenty smaller models, the db2 score is again uninformative, even on models trained with db2 masks (Table 23). In summary, the paper compares a working score with an uninformative one.

  3. The paper is difficult to read. The sentence structures are overly complex with an adjective-heavy presentation (such as "validation-fixed two-sided OOD score" or "budget-integrated retained error"). Several non-standard coinages appear in the text. On the electromagnetic task, the correlation appears as 0.46, 0.51, or 0.67, depending on which of three sets of trained models is used. Only the appendix clarifies which number corresponds to what setting, and even there it is stated that the protocol behind the 0.67 result cannot be recovered (Appendices F and H).

Questions for authors

On the electromagnetic scattering and Maxwell tasks, do the ranking results hold for better predictors? The predictors shown have 60 to 90% relative error.

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

This paper proposes Allocation Uncertainty Decomposition (AUD), which provides a single-pass allocation readout that is shown to be sometimes useful for risk triage. Because this is a single-pass method, it is computationally cheaper than multi-pass methods like Monte Carlo dropout. However, the predictive uncertainty signal provided appears to be small.

Critical strengths
  1. The experiment design in this paper is clean, and multiple adaptors are presented.
  2. The paper is clear about the negative results.
  3. The proposed method is computationally cheap compared to established uncertainty estimators like Monte Carlo Dropouts.
Critical weaknesses
  1. The intuition that allocation entropy can contain predictive uncertainty that's not in the final prediction itself is poorly motivated.
  2. The performance seems weak compared to very simple baselines, such as the variance head in table 2. I am further concerned that in table 17 it seems like model-agnostic policies such as roughness is beating AUD entropy, which would suggest that the a difficulty heuristics that is completely blind to the model can outperform the predictive uncertainty provided by AUD entropy.
Questions for authors
  1. Why should allocation ambiguity indicate predictive uncertainty? (see W1)
  2. What is the benefit of the proposed method compared to the simple baselines that outperform it? (see W2)
  3. I am a bit confused by why the sign of entropy needs fitting at all, usually we think about higher entropy as higher uncertainty because of information theory.

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

Related papers

Discussion (0)

Sign in to comment.