arxiv: 2605.01502 · v1 · submitted 2026-05-02 · 💻 cs.CV

Recognition: unknown

RADMI: Latent Information Aggregation as a Proxy for Model Uncertainty

William Stevens , Mohit Prabhushankar , Ghassan AlRegib

Authors on Pith no claims yet

Pith reviewed 2026-05-09 14:20 UTC · model grok-4.3

classification 💻 cs.CV

keywords uncertainty estimationmutual informationsemantic segmentationepistemic uncertaintysingle-pass methoddecoder layersencoder-decoder architecture

0 comments

The pith

Aggregating mutual information between decoder layers serves as a single-pass proxy for model uncertainty in segmentation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes that uncertainty in neural network segmentation predictions can be estimated by tracking how much information is shared between consecutive layers in the decoder. High mutual information signals regions where the network is reconciling conflicting context, such as class boundaries, and this signal correlates strongly with the uncertainty obtained from full ensembles. Because the method runs in one forward pass and requires no architecture changes, it offers a practical alternative to slow ensemble or multi-pass techniques for dense prediction tasks. On a seismic facies benchmark the approach matches ensemble uncertainty better than prior single-pass baselines while producing maps that stay sharply localized to ambiguous areas.

Core claim

RADMI estimates prediction uncertainty by computing mutual information between feature maps from successive decoder layers and linearly aggregating the normalized values across resolutions. The authors observe that elevated inter-layer mutual information arises precisely where the network must integrate conflicting contextual cues, producing a spatial uncertainty map that aligns more closely with deep-ensemble uncertainty than other single-pass methods, with gains of 5.5 percent Pearson and 10.7 percent Spearman correlation on the evaluated segmentation task.

What carries the argument

Resolution-Aggregated Decoder Mutual Information (RADMI), the linear aggregation of normalized mutual information computed between consecutive decoder-layer feature maps.

If this is right

Uncertainty maps become available at the cost of a single forward pass without any architectural modification.
Uncertainty concentrates sharply at class boundaries and other ambiguous regions rather than being diffuse.
Single-pass methods can now reach correlation levels with ensembles that previously required multiple stochastic passes.
Linear aggregation of inter-layer information flow supplies a principled, architecture-agnostic uncertainty signal for encoder-decoder networks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same inter-layer mutual information signal may serve as an uncertainty proxy in other encoder-decoder dense tasks such as depth estimation or medical image analysis.
High-MI regions could be used to guide active learning or targeted data collection without needing an external uncertainty oracle.
The approach implicitly treats the decoder as an information-integration stage whose internal conflicts reveal epistemic uncertainty, an idea that could be tested by ablating decoder depth or skip connections.

Load-bearing premise

Elevated mutual information between decoder layers reliably marks regions where the network is uncertain because it is forced to reconcile conflicting contextual information.

What would settle it

A new segmentation dataset in which RADMI values fail to correlate with ensemble-derived uncertainty or with actual prediction errors at ambiguous boundaries would undermine the proxy claim.

read the original abstract

Epistemic uncertainty estimation is essential for identifying regions where deep learning system outputs may be unreliable. However, existing approaches require computationally expensive ensemble methods or multiple stochastic forward passes, limiting their scalability to dense prediction tasks like segmentation. We propose Resolution-Aggregated Decoder Mutual Information (RADMI), a single-pass method that estimates prediction uncertainty by measuring mutual information (MI) between consecutive decoder layers in segmentation networks. We observe that elevated inter-layer MI correlates with prediction uncertainty, as the network must integrate conflicting contextual information at ambiguous regions such as class boundaries. Evaluating on a seismic facies segmentation benchmark, RADMI achieves the highest correlation with deep ensemble uncertainty among all single-pass methods, outperforming the next-best baselines by 5.5% in Pearson and 10.7% in Spearman correlation coefficients. Compared to baselines that either lack spatial precision or demand significant computational overhead, RADMI yields sharp, boundary-localized uncertainty maps without architectural modifications. Our results suggest that linear aggregation of normalized information flow provides a principled and efficient proxy for prediction uncertainty in encoder-decoder architectures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RADMI offers a cheap single-pass uncertainty proxy by aggregating decoder-layer mutual information, but the link to epistemic uncertainty is observational and all evidence is from one narrow benchmark.

read the letter

The main takeaway is that this paper introduces RADMI, which computes mutual information between consecutive decoder layers in a segmentation network, normalizes and aggregates it across resolutions, and uses the result as a single-forward-pass estimate of prediction uncertainty. On a seismic facies dataset it reports higher Pearson and Spearman correlation with deep-ensemble uncertainty than several other single-pass baselines, and the maps look spatially sharp at boundaries without any architecture changes or extra passes. That is the concrete advance: a lightweight way to repurpose internal feature statistics for uncertainty in encoder-decoder models. It sits in the single-pass uncertainty literature but picks a different signal than variance or dropout-based methods. The efficiency and lack of overhead are real practical pluses for dense prediction where ensembles are too slow. The soft spots are straightforward. The justification for why elevated inter-layer MI should track epistemic uncertainty is presented as an observation about conflicting context at boundaries rather than a derivation or bound; nothing rules out other explanations such as feature redundancy or dimensionality effects. All quantitative support is a single correlation comparison on one specialized benchmark, with no ablations isolating the MI term, no error bars, no results on standard segmentation datasets, and no independent grounding against predictive entropy or other uncertainty measures. That leaves the central performance claim vulnerable to dataset-specific statistics. The work is aimed at engineers or researchers who need fast uncertainty maps for segmentation in applied domains and are willing to accept correlational evidence for now. A reader already working on single-pass methods might pick up the implementation idea and test it themselves. I would send it to peer review because the idea is distinct enough and the reported edge is positive, even though the paper would need additional datasets and a clearer account of the MI-uncertainty relationship to be convincing on its own terms.

Referee Report

4 major / 2 minor

Summary. The paper introduces RADMI (Resolution-Aggregated Decoder Mutual Information), a single-pass method for estimating epistemic uncertainty in encoder-decoder segmentation networks. It computes mutual information between consecutive decoder layers, normalizes and linearly aggregates these values across resolutions, and claims this yields uncertainty maps that correlate more strongly with deep-ensemble uncertainty than existing single-pass baselines. The central observation is that elevated inter-layer MI occurs at ambiguous regions (e.g., class boundaries) where the decoder integrates conflicting context. On a seismic facies segmentation benchmark, RADMI reportedly outperforms the next-best single-pass method by 5.5% Pearson and 10.7% Spearman correlation with ensemble uncertainty, while producing spatially sharp maps without architectural changes or multiple forward passes.

Significance. If the empirical link between aggregated decoder-layer MI and epistemic uncertainty generalizes, RADMI would provide a computationally cheap, single-pass alternative to ensembles or MC-dropout for dense prediction tasks. This could improve scalability for applications requiring uncertainty-aware segmentation. However, the current evidence consists solely of correlation coefficients on one specialized dataset with no theoretical derivation, error bars, or external validation, so the practical significance remains provisional.

major comments (4)

[Abstract and §3] Abstract and §3 (method): The assertion that RADMI is a 'principled' proxy because 'elevated inter-layer MI correlates with prediction uncertainty' is presented as an observation rather than derived. No equation or bound is given relating the MI quantity to predictive entropy, ensemble variance, or any other established uncertainty measure; the construction therefore provides no guarantee that the aggregation isolates epistemic uncertainty rather than dataset-specific statistics such as boundary frequency or feature dimensionality.
[Evaluation] Evaluation section: All quantitative support is a single Pearson/Spearman comparison against deep-ensemble uncertainty on one seismic facies dataset. No error bars, multiple random seeds, or ablation isolating the MI aggregation from normalization choices are reported. This makes the 5.5% / 10.7% improvement claim difficult to assess for robustness or statistical significance.
[Abstract and §4] Abstract and §4: The manuscript reports no results on standard segmentation benchmarks (Cityscapes, Pascal VOC, medical imaging datasets). Without these, the claim that RADMI 'outperforms all single-pass methods' and yields 'sharp, boundary-localized uncertainty maps' cannot be evaluated for generalizability beyond the seismic domain.
[§3] §3 (RADMI definition): The method defines uncertainty directly via the same inter-layer MI quantity whose correlation with ensemble uncertainty is then measured. No independent grounding (e.g., calibration plots, OOD detection performance, or comparison to predictive entropy on held-out data) is described, creating a circularity risk that the reported correlations may partly reflect shared sensitivity to the same image statistics rather than a true uncertainty proxy.

minor comments (2)

[§3] Clarify the exact normalization and aggregation formula for RADMI (e.g., how MI values are scaled across layers and resolutions) so that the method can be reproduced from the text alone.
[Abstract and §3] The abstract states 'linear aggregation of normalized information flow' but does not specify whether the normalization is per-layer, per-resolution, or global; this detail should be stated explicitly in the method section.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the constructive comments that help clarify the scope and limitations of our work. We respond point-by-point to the major comments and indicate the revisions we will make.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (method): The assertion that RADMI is a 'principled' proxy because 'elevated inter-layer MI correlates with prediction uncertainty' is presented as an observation rather than derived. No equation or bound is given relating the MI quantity to predictive entropy, ensemble variance, or any other established uncertainty measure; the construction therefore provides no guarantee that the aggregation isolates epistemic uncertainty rather than dataset-specific statistics such as boundary frequency or feature dimensionality.

Authors: We agree that the link is observational rather than theoretically derived from a bound or equation relating inter-layer MI to established uncertainty quantities such as predictive entropy or ensemble variance. The term 'principled' in the abstract was intended to reflect the information-theoretic motivation, but we will revise the abstract and Section 3 to remove this wording, explicitly describe RADMI as an empirical proxy, and note that it does not provide a formal guarantee against confounding with dataset-specific factors. revision: yes
Referee: [Evaluation] Evaluation section: All quantitative support is a single Pearson/Spearman comparison against deep-ensemble uncertainty on one seismic facies dataset. No error bars, multiple random seeds, or ablation isolating the MI aggregation from normalization choices are reported. This makes the 5.5% / 10.7% improvement claim difficult to assess for robustness or statistical significance.

Authors: We acknowledge that the current evaluation lacks error bars, multiple seeds, and ablations. We will recompute the correlations over multiple random seeds to report error bars and add an ablation study that isolates the contribution of the resolution-aggregated MI from the normalization and linear aggregation steps. These changes will be included in the revised evaluation section. revision: yes
Referee: [Abstract and §4] Abstract and §4: The manuscript reports no results on standard segmentation benchmarks (Cityscapes, Pascal VOC, medical imaging datasets). Without these, the claim that RADMI 'outperforms all single-pass methods' and yields 'sharp, boundary-localized uncertainty maps' cannot be evaluated for generalizability beyond the seismic domain.

Authors: We agree that results on standard benchmarks would be needed to support broad claims of outperformance. Our experiments focus on the seismic facies task. In revision we will qualify the abstract and Section 4 claims to refer specifically to the seismic benchmark, remove the unqualified statement that RADMI 'outperforms all single-pass methods,' and add a limitations paragraph discussing potential applicability and challenges for other domains. revision: partial
Referee: [§3] §3 (RADMI definition): The method defines uncertainty directly via the same inter-layer MI quantity whose correlation with ensemble uncertainty is then measured. No independent grounding (e.g., calibration plots, OOD detection performance, or comparison to predictive entropy on held-out data) is described, creating a circularity risk that the reported correlations may partly reflect shared sensitivity to the same image statistics rather than a true uncertainty proxy.

Authors: The inter-layer MI is computed from a single forward pass on the trained network, whereas the reference is uncertainty from an independent deep ensemble; the correlation therefore serves as external validation rather than a circular definition. Nevertheless, to mitigate the perceived circularity risk we will expand Section 3 with an explicit distinction between the internal MI measure and the external ensemble reference, and add qualitative comparisons of RADMI maps against predictive entropy on the same data. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained via external validation

full rationale

The paper defines RADMI explicitly as linear aggregation of normalized inter-layer mutual information in decoder layers, then validates the proxy claim by computing Pearson and Spearman correlations against an independent external benchmark (deep ensemble uncertainty) on a seismic facies dataset. This correlation measurement is not forced by construction or self-definition; the ensemble provides a separate reference distribution. No equations reduce the target uncertainty quantity to the MI aggregation itself, no self-citations bear the central premise, and no uniqueness theorems or ansatzes are imported from prior author work. The derivation chain remains open to empirical falsification outside the fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unproven assumption that inter-layer MI directly tracks epistemic uncertainty and on the empirical observation that linear aggregation of normalized MI produces useful maps; no free parameters are explicitly named in the abstract, but the aggregation and normalization steps implicitly introduce choices.

axioms (1)

domain assumption Elevated mutual information between consecutive decoder layers indicates regions where the network integrates conflicting contextual information.
Stated in the abstract as the mechanistic justification linking MI to uncertainty.

pith-pipeline@v0.9.0 · 5486 in / 1242 out tokens · 24570 ms · 2026-05-09T14:20:15.876251+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 1 canonical work pages · 1 internal anchor

[1]

INTRODUCTION Deep learning has found success in automating seismic interpreta- tion tasks, including facies classification [1], fault delineation [2], and salt body segmentation [3]. By learning hierarchical feature representations directly from migrated seismic volumes, neural net- works can produce interpretations that approach expert-level qual- ity on...
[2]

RELATED WORKS 2.1. Uncertainty Quantification in Semantic Segmentation Uncertainty quantification for semantic segmentation presents unique challenges beyond image classification, as uncertainty must be estimated at the pixel level rather than for the semantics of the entire image. The most common single-pass approaches derive un- certainty from the softm...
[3]

We propose a single-pass uncertainty estimation method based on measuring mutual informa- tion between consecutive decoder layers in segmentation networks

METHODOLOGY Figure 2 illustrates theRADMIworkflow. We propose a single-pass uncertainty estimation method based on measuring mutual informa- tion between consecutive decoder layers in segmentation networks. The core methodology measures the mutual information I(Z l;Z l+1)(1) which quantifies the statistical dependence between adjacent de- coder stagesland...
[4]

EXPERIMENTS AND RESULTS 4.1. Experimental Setup We evaluate our proposedRADMImethod on the F3 seismic facies segmentation benchmark [1], a widely-used dataset for seismic inter- pretation containing 6 lithological classes: Upper North Sea, Middle North Sea, Lower North Sea, Rijnland/Chalk, Scruff, and Zechstein. The F3 block, located in the Dutch sector o...
[5]

CONCLUSION We have presentedRADMI, a single-pass uncertainty estimation method based on mutual information between consecutive decoder layers. By directly measuring inter-layer information flow,RADMI outperforms the next-best baselines by 5.5% in Pearson and 10.7% in Spearman correlation with deep ensemble uncertainty, while producing sharp, boundary-loca...
[6]

A machine-learning benchmark for facies classification,

Yazeed Alaudah, Patrycja Michałowicz, Motaz Alfarraj, and Ghassan AlRegib, “A machine-learning benchmark for facies classification,”Interpretation, vol. 7, no. 3, pp. SE175–SE187, 07 2019

2019
[7]

A large-scale bench- mark on geological fault delineation models: Domain shift, training dynamics, generalizability, evaluation, and inferential behavior,

Jorge Quesada, Chen Zhou, Prithwijit Chowdhury, Moham- mad Alotaibi, Ahmad Mustafa, Yusufjon Kumakov, Mohit Prabhushankar, and Ghassan AlRegib, “A large-scale bench- mark on geological fault delineation models: Domain shift, training dynamics, generalizability, evaluation, and inferential behavior,”IEEE Access, vol. 13, pp. 215110–215131, 2025

2025
[8]

Subsurface structure analysis using com- putational interpretation and learning: A visual signal process- ing perspective,

Ghassan AlRegib, Mohamed Deriche, Zhiling Long, Haibin Di, Zhen Wang, Yazeed Alaudah, Muhammad Amir Shafiq, and Motaz Alfarraj, “Subsurface structure analysis using com- putational interpretation and learning: A visual signal process- ing perspective,”IEEE Signal Processing Magazine, vol. 35, no. 2, pp. 82–98, 2018

2018
[9]

On calibration of modern neural networks,

Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger, “On calibration of modern neural networks,” inInternational conference on machine learning. PMLR, 2017, pp. 1321–1330

2017
[10]

A unified framework for evaluating the robustness of machine-learning interpretability for prospect risking,

Prithwijit Chowdhury, Ahmad Mustafa, Mohit Prabhushankar, and Ghassan AlRegib, “A unified framework for evaluating the robustness of machine-learning interpretability for prospect risking,”Geophysics, vol. 90, no. 3, pp. IM103–IM118, 2025

2025
[11]

What uncertainties do we need in bayesian deep learning for computer vision?,

Alex Kendall and Yarin Gal, “What uncertainties do we need in bayesian deep learning for computer vision?,”Advances in neural information processing systems, vol. 30, 2017

2017
[12]

Effective data selection for seismic interpretation through dis- agreement,

Ryan Benkert, Mohit Prabhushankar, and Ghassan AlRegib, “Effective data selection for seismic interpretation through dis- agreement,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–12, 2024. [8]Enrich the interpretation of seismic image segmentation by estimating epistemic uncertainty, vol. SEG Technical Pro- gram Expanded Abstracts 202...

2024
[13]

Simple and scalable predictive uncertainty estima- tion using deep ensembles,

Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell, “Simple and scalable predictive uncertainty estima- tion using deep ensembles,”Advances in neural information processing systems, vol. 30, 2017

2017
[14]

Dropout as a bayesian approximation: Representing model uncertainty in deep learn- ing,

Yarin Gal and Zoubin Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learn- ing,” inProceedings of The 33rd International Conference on Machine Learning, Maria Florina Balcan and Kilian Q. Wein- berger, Eds., New York, New York, USA, 20–22 Jun 2016, vol. 48 ofProceedings of Machine Learning Research, pp. 1050–1059, PMLR

2016
[15]

Reliable uncertainty estimation for seismic interpretation with prediction switches,

Ryan Benkert, Mohit Prabhushankar, and Ghassan AlRegib, “Reliable uncertainty estimation for seismic interpretation with prediction switches,” inSEG/AAPG International Meeting for Applied Geoscience & Energy, 08 2022, vol. SEG/AAPG In- ternational Meeting for Applied Geoscience & Energy ofSEG International Exposition and Annual Meeting

2022
[16]

U-net: Convolutional networks for biomedical image segmentation,

Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” 2015

2015
[17]

Man-recon: Manifold learning for reconstruction with deep autoencoder for smart seismic interpretation,

Ahmad Mustafa and Ghassan AlRegib, “Man-recon: Manifold learning for reconstruction with deep autoencoder for smart seismic interpretation,” in2021 IEEE International Confer- ence on Image Processing (ICIP). IEEE, 2021, pp. 2953–2957

2021
[18]

A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks

Dan Hendrycks and Kevin Gimpel, “A baseline for detecting misclassified and out-of-distribution examples in neural net- works,”arXiv preprint arXiv:1610.02136, 2016

work page internal anchor Pith review arXiv 2016
[19]

Counterfactual gradients-based quantification of prediction trust in neural net- works,

Mohit Prabhushankar and Ghassan AlRegib, “Counterfactual gradients-based quantification of prediction trust in neural net- works,” in2024 IEEE 7th International Conference on Mul- timedia Information Processing and Retrieval (MIPR). IEEE, 2024, pp. 529–535

2024
[20]

V oice: Variance of induced contrastive explanations to quantify uncertainty in neural network interpretability,

Mohit Prabhushankar and Ghassan AlRegib, “V oice: Variance of induced contrastive explanations to quantify uncertainty in neural network interpretability,”IEEE Journal of Selected Top- ics in Signal Processing, vol. 19, no. 1, pp. 19–31, 2024

2024
[21]

The information bottleneck method,

Naftali Tishby, Fernando C. Pereira, and William Bialek, “The information bottleneck method,” 2000

2000
[22]

Opening the black box of deep neural networks via information,

Ravid Shwartz-Ziv and Naftali Tishby, “Opening the black box of deep neural networks via information,” 2017

2017
[23]

Deep learning and the information bottleneck principle,

Naftali Tishby and Noga Zaslavsky, “Deep learning and the information bottleneck principle,” 2015

2015
[24]

Thomas M Cover and Joy A Thomas,Elements of Information Theory, Wiley-Interscience, 2nd edition, 2006

2006