acceptodds
Under review as a conference paper at ICLR 2027

DOES PREDICTIVE UNCERTAINTY GUIDE US TO THE MISSING INFORMATION? A CONTROLLED AUDIT UNDER PARTIAL EEG OBSERVATION

Abstract

Predictive uncertainty can identify difficult predictions, but can it also identify which unavailable measurements matter? We audit this stronger interpretation through controlled channel withholding in EEG: the history, future target, horizon, and retained channel count stay fixed while channel identity changes. A standard heteroscedastic predictor trained on 1,600 hours produces sample-dependent, error-associated uncertainty. Yet its measurement-level interpretation breaks down in a specific way. Teacher-state consequence rankings reproduce across disjoint subject halves with correlation approximately , and uncertainty rankings with – across three constructions, while their aggregate cross-half alignment is weak or negative. Both rankings therefore contain population structure, but that structure does not reliably align; many individual mask-pair orderings also remain unresolved. To characterize what the score tracks, we measure Local Predictive-State Dispersion, a model-defined structural probe that covaries positively with uncertainty. An analytic oracle shows that interpreting this probe as exact conditional ambiguity is regime-dependent. Together, these results establish a controlled boundary between recognizing prediction difficulty and identifying missing-measurement consequences, motivating validation against the specific semantic claim made about an uncertainty score.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
predictive uncertainty, missing information, EEG, uncertainty quantification, semantic validity

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_VJb3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper conducts an empirical study using EGG dataset to audit a deep learning network (Conv1d + a four-layer transformer). Their results reveal that 1) there is correlation between uncertainty score and model error, and 2) uncertainty score can predict the masked channel in EGG data. The authors also conduct additional checks showing that 1) the masked channel would make the prediction different and 2) the results can not be interpreted by conditional ambiguity (through a toy example with 8 different possible oracles).

Critical strengths
  • The study is thorough. The authors provide evidence to exclude the confounders of their conclusions.
Critical weaknesses
  • The empirical results are not new. The two main takeaways (in my opinion) have all been studied by prior works [1, 2, 3].
  • The scope of the paper is somehow limited. The authors only study one model and one dataset. It's hard to think whether the results can be generalized to other fields (especially the misinformation is only modeled as masked channels in EGG type of data).

[1] Xiong, Miao, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. "Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms." In International Conference on Learning Representations, vol. 2024, pp. 23650-23678. 2024.

[2] Gadgil, Soham, Ian Covert, and Su-In Lee. "Estimating conditional mutual information for dynamic feature selection." In International Conference on Learning Representations, vol. 2024, pp. 34979-35010. 2024.

[3] Hou, Bairu, Yujian Liu, Kaizhi Qian, Jacob Andreas, Shiyu Chang, and Yang Zhang. "Decomposing uncertainty for large language models through input clarification ensembling." arXiv preprint arXiv:2311.08718 (2023).

Questions for authors
  • What is the new contribution offered by this paper over the existing works listed above?
  • Is there any unique takeaway from this analysis on EGG data? This might be related to my previous concern about the generalization and new contribution.
Reviewer_zbsxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper audits whether predictive uncertainty can identify which unavailable measurements matter using controlled channel withholding in EEG. The paper sets up an experiment by fixing the underlying EEG history, future prediction target, prediction horizon, and number of observed channels while varying channel identity. The paper finds that uncertainty rankings and measurement-consequence rankings are individually reproducible across subject-disjoint splits; however, they exhibit weak or negative alignment with one another.

Critical strengths

1.The paper addresses an important distinction between identifying difficult predictions and identifying which missing measurements matter. 2.The experimental design is sound and focuses on differences associated with the identities of the retained EEG channels. Also provides comprehensive analysis using uncertainty estimators.

Critical weaknesses

1.The paper uses both future-prediction error and teacher-state reconstruction error to assess the consequences of missing EEG channels, prioritizing teacher-state error because it is more reproducible. However, reproducibility does not establish that a metric captures the intended notion of measurement consequence, and future-prediction error is more directly related to the predictive task. How consistent are the mask rankings produced by these two metrics? 2.The paper reports the uncertainty–consequence correlations without confidence intervals. Thus, it’s not clear how robust the results are, also it is difficult to judge whether the weak or negative alignment is distinguishable from positive alignment. It is also unclear how much prediction error differs across EEG channel masks. So, the importance of the observed misalignment remains uncertain. 3.The uncertainty estimator evaluations are limited and reported on the heteroscedastic and deterministic models. Stronger alternatives such as deep ensembles would be helpful because it’s not clear whether the dissociation shows predictive uncertainty or a limitation of these particular estimators.

Questions for authors

See weaknesses

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

The paper trains a small transformer model to predict future EEG spectral states from a 10s signal window and asks whether the ranking of channel subsets by predictive variance (via Gaussian NLL, with learned mean and sigma heads) matches their ranking by error, which bears importance for knowing which channels to use in a sparse setup. It trains the predictor on the dataset with random channel masks, and checks on a held-out cohort (with random splits in two) if the sigma ranking of fixed-size masks correlates with the measured error ranking for the same masks, as it would for a well calibrated model. They find a negative result, that both the sigma and the error mask rankings are consistent across splits, but don’t correlate with each other. They thus claim that the predictive uncertainty indicates hard predictions but does not identify what missing channels would alleviate it.

Critical strengths

Whether a predictor’s uncertainty can be used to choose electrodes for EEG montages is a relevant practical question, and a real dataset is used.

Negative results are reported transparently.

Critical weaknesses

The main result, that the mask sigma ranking doesn’t match the error ranking, appears to derive from the setup. There are many masks (random subsets) so the model see few samples for a given masks and might not be learning the calibrated sigma for specific masks. The calibration was only reported over the aggregate over masks, not over individual masks, which determines the ranking. The size of the mask effect at fixed mask size is also not reported. Most subset pairs cannot be ordered even under bootstrap, which suggests a small effect, in which case the rank correlations would be mostly noise. The relevance of the results is very limited. It is unclear if the main problem of inferring what channels are informative from the output sigma is something people would expect to do, which limits even more the significance of a negative result. A standard way to decide which channel to add is expected information gain, by directly estimating how much a channel reduces the error. The paper doesn’t justify it’s approach of trying to predict only from sigma, nor does it compare with the marginal information gain of each channel in the mask. The error ranking may largely reflect the target construction, as the target is a PCA of regional spectral features over all 19 channels, so the consequence of a mask could simply track how much target variance the channels’ regions carry. Also, the variance, like the error, can only be computed for channels that were recorded, so it cannot score a channel you don’t have, and the method would not applicable in a practical setting to inform on channel you haven’t measured.

There are multiple missing important baselines and ablations that would inform what is behind the results. The variance head performs worse than the plain MSE and than a constant variance on NLL, the known problem with fitting mean and variance jointly, and it is not tested whether this causes the result. It is not determined where the variance’s information comes from, e.g. no ablation removing the subset input, no subset-only or window-only variance model. There is no look into what electrodes are relevant for sigma, e.g. artifact-heavy channels, and the masks are random draws from tens of thousands of possible subsets. A few standard setups, the ones a wearable or ICU setup would actually use, would have been both more relevant and easier to interpret. Sections 6 and 7, on the kNN and the oracle, do not add on the main results. The presentation is poor throughout, e.g. a novel formulation for what is essentially a stratified calibration check and with constant disclaimers after every sentence.

Questions for authors

How much does the error vary depending on which channels were kept? This would indicate how much signal there is for the ranking.

Does the inverted ordering remain if the variance head is trained without the channel subset as an input? That would separate a learned subset prior from what’s in the signal.

Reviewer_VjsKOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper examines whether predictive uncertainty can identify which missing measurements matter, using controlled channel withholding in EEG. The authors of the paper find that uncertainty and prediction-error rankings across channel subsets are individually reproducible but poorly aligned. They also introduce Local Predictive-State Dispersion (LPSD), which correlates with uncertainty, and use a synthetic oracle to examine the limits of interpreting this probe as conditional ambiguity.

Critical strengths
  • Recognising that a prediction is difficult does not necessarily tell us which additional measurements would help. Testing this distinction directly is useful, particularly when uncertainty is used to motivate sensor selection or data acquisition. The paper felt well-motivated.
  • Checking the reproducibility of both uncertainty and consequence rankings addresses the possibility that their disagreement arises solely from an unstable target ranking.
Critical weaknesses

Weaknesses

  • An uncertainty score can correlate with error across recordings while systematically misestimating risk for particular masks. For a fixed mean predictor, the population-optimal Gaussian variance estimates its conditional squared error; averaging within a mask therefore recovers that predictor’s expected error, up to scaling. Poor population mask alignment could consequently reflect estimation, optimisation, or distribution-shift problems. The paper acknowledges these possibilities but does not investigate them enough which gives me some concerns.
  • Section 5.1 states that the uncertainty constructions share reference consequences that are not necessarily their own predictive risks. This makes the comparison harder to interpret: disagreement may reflect differences between predictors or target spaces. I would like to see each score evaluated against its own predictor’s error alongside the shared-reference comparisons. The more closely matched belief/teacher-state comparison is useful, but its estimator specification is incomplete.
  • The appendices identify omissions in the exact masks, future-descriptor construction, belief distribution and score reduction, LPSD mask-matching policy, and fixed-mask diagnostic aggregation. These affect the quantities being measured, rather than just implementation convenience. In particular, the reported fixed-mask correlation is difficult to assess without knowing whether correlations were computed within masks before aggregation. Appendix M’s distinction between reproducing summary arithmetic and reproducing the underlying experiments is a substantial limitation.

Minor Points

  • The paper structure is super difficult to read. There is a new subsection every couple of paragraphs.
  • Throughout the reading of the paper, I wasn't convinced that ICLR was the correct venue for this paper.
Questions for authors
  • Does the dissociation persist when each uncertainty score is compared with its own predictor’s error?
  • Were the evaluated mask counts and patterns supported during training, and what exactly was the training mask distribution?
  • Can the missing estimator definitions and fixed-mask aggregation procedure be supplied, together with quantitative within-history results?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.