Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Last-layer committee machines for uncertainty estimations of benthic imagery

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A last-layer committee machine estimates per-image uncertainty for benthic imagery with over 95% fewer parameters, matching the accuracy of Monte Carlo dropout and Bayesian approaches.

desk verdict A useful applied comparison showing a 100-head last-layer committee matches MCD accuracy on benthic images at a fraction of the inference cost, but the epistemic-uncertainty claim is not supported by the diversity evidence in the paper. read the letter →

arxiv 2504.16952 v2 pith:U2I6VFL5 submitted 2025-04-22 q-bio.QM

H. Martin Gillis , Isaac Xu , Benjamin Misiuk , Craig J. Brown , Thomas Trappenberg This is my paper · ORCID
classification q-bio.QM
keywords benthichabitatmappinglast-layerensemblecommitteemachineuncertaintyestimationMonteCarlodropoutBayesianneuralnetworkhuman-in-the-loopsingularvaluedecomposition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that attaching 100 thin linear classifiers to a shared pre-trained feature extractor can give per-image uncertainty estimates for seafloor imagery at a fraction of the cost of standard Bayesian approaches. Each committee member is trained with its own cross-entropy loss, then the member softmax outputs are averaged in one forward pass. On the German Bank 2010 and Substrate depth 2 benthic datasets, the reported accuracies are within about one percentage point of Monte Carlo dropout (0.786 vs 0.794 and 0.879 vs 0.885), while the number of network parameters evaluated for uncertainty drops by more than 95%. If correct, this gives marine scientists a practical way to generate prioritized lists of ambiguous, mislabeled, or out-of-distribution images for human review.

What carries the argument

The load-bearing object is the last-layer committee machine: $M$ linear layers $g_m$ appended to a shared feature extractor $f(x; w_f)$, each with randomly initialized weights and its own cross-entropy loss $\ell_m(y, g_m(f; w_{g_m}))$, with gradients backpropagated independently for each member. At inference, the per-sample uncertainty estimate is the average of the $M$ softmax outputs, giving an approximation to the posterior predictive distribution in one forward pass. Diversity is quantified by the coefficient of variation of the Frobenius norms of singular values across members, which remains non-zero but small (about $2.7\times10^{-3}$ to $3.6\times10^{-3}$) after training, and this residual diversity is what the authors lean on for the epistemic-uncertainty claim.

What would settle it

Compute the correlation between LLCM-100 per-sample standard deviation and prediction error on held-out benthic images, including artificially mislabeled and geographically out-of-distribution images; if the correlation is near zero or the spread collapses to the entropy of the mean softmax, the epistemic-uncertainty claim would not be supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that a last-layer committee machine can replace computationally expensive uncertainty estimators such as Bayesian neural networks and Monte Carlo dropout without meaningful loss in classification accuracy or calibration. The method keeps one shared feature extractor and adds $M=100$ single linear layers as committee members; each member is randomly initialized and trained with its own cross-entropy objective, and inference averages the $M$ softmax distributions. The authors report near-identical accuracy to Monte Carlo dropout on both benthic datasets, with a single forward pass and a reduction in parameters evaluated by about 20 times on the benchmark network and more than 95% overall. They also provide diversity analyses based on singular value decompositions of member weights, arguing that random initialization alone leaves the members sufficiently distinct to support epistemic uncertainty.

Load-bearing premise

The claim stands on the assumption that the 100 committee members, after training on the same features, remain different enough that the spread of their predictions is a meaningful measure of model uncertainty rather than numerical noise.

Editorial extensions

If this is right

  • Single-pass uncertainty: an LLCM-100 produces per-sample standard deviations over committee members in one forward pass, so uncertainty-aware review can scale to the millions of images in large benthic compilations.
  • Parameter reduction: with a shared feature extractor and 100 single-layer heads, the number of network parameters evaluated for uncertainty drops by more than 95% relative to sampling a full network 100 times for Monte Carlo dropout.
  • Accuracy parity: LLCM accuracy is within about one point of Monte Carlo dropout on German Bank 2010 (0.786 vs 0.794) and Substrate depth 2 (0.879 vs 0.885), with calibration restored by temperature scaling.
  • Simplicity: random initialization of the committee members is enough to create diversity; no specialized loss, logit averaging, or diversity-promoting regularizer is required.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: Because the committee members end up nearly identical (singular-value coefficient of variation around 0.003), the per-sample uncertainty could in practice be only slightly more informative than the softmax entropy of the averaged output; this should be checked explicitly before interpreting the spread as epistemic uncertainty.
  • Inference: The efficiency-threshold plots the paper introduces could be used as a model-agnostic protocol to audit any uncertainty estimator on any dataset, not just benthic imagery.
  • Inference: A sharper test of the method would be to measure whether LLCM variance increases on out-of-distribution images from geographic regions absent from training; the paper's experiments focus on in-distribution held-out partitions.
  • Inference: The same single-layer committee could be plugged into an active-learning loop by re-annotating the lowest-confidence samples and adding them to the training pool, which the paper mentions only as a long-term goal.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a last-layer committee machine (LLCM) in which a shared feature extractor feeds M independently initialized linear heads, trained by summed cross-entropy losses, and per-sample uncertainties are obtained in a single forward pass by averaging the head softmax outputs. The authors compare LLCM-100 with Bayesian model averaging (BMA-100), Monte Carlo dropout (MCD-100), and deterministic CNNs on MNIST and on two benthic datasets, German Bank 2010 and Substrate (depth 2), using accuracy, NLL, Brier score, ECE, reliability diagrams, and threshold-based accuracy/efficiency curves. They also report weight-space diversity analyses based on coefficients of variation and SVD singular values, and they claim that random initialization of the committee heads is sufficient for epistemic (model) uncertainty. The headline empirical result is that LLCM-100 accuracy is within roughly one point of MCD-100 on the two benthic datasets while requiring one forward pass instead of 100 stochastic forward passes.

Significance. If the accuracy and uncertainty-ranking results hold, the LLCM is a useful, low-complexity tool for prioritizing ambiguous benthic images for human review. The paper's comparisons are internally consistent, the data and pretrained models are public, calibration is performed on the validation set, and the uncertainty metric plots provide a transparent way to display selective accuracy and efficiency. The work also engages appropriately with recent last-layer Bayesian methods. However, the epistemic-uncertainty contribution is not established by the reported weight-space diversity evidence, the '95% reduction of network parameters' claim is not literally a parameter-count reduction, and the near-identical performance claim lacks seed-level variance. These issues do not invalidate the accuracy/efficiency comparison, but they require revision before the paper can be recommended.

major comments (4)
  1. [Section 4.1, Eq. (5)] The evidence for Contribution 3 (random initialization yields epistemic model uncertainty) is weight-space diversity only, and that evidence is not persuasive. Fig. 1b-d reports CV||SV||F between 2.7e-3 and 3.6e-3 for an LLCM-10 on MNIST, and Fig. 1 shows this metric decreasing during training; Table 3 reports sigma_CVM values ranging from about 9.4 to 8372 depending on hyperparameters, which reflects logit magnitude scaling rather than functional diversity. None of these metrics measures whether the M members' predictive distributions p(y|x; w^g_m) disagree on inputs where the model is uncertain. If residual member differences are negligible, the variance across members in Eq. (5) collapses and the per-sample uncertainty reduces to the entropy of the committee-mean softmax, which is not epistemic uncertainty. Please provide a direct predictive-diversity measurement (e.g., per-sample standard deviation of member softmax probabilities, average pairwise KL or Jensen-Shannon divergence, or fraction of members agreeing with the ensemble mean) computed for the deployed LLCM-100 models on German Bank 2010 and Substrate, and show that these quantities are non-negligible relative to the ensemble-mean entropy. Without this, the epistemic-uncertainty claim is unsupported.
  2. [Abstract; Section 4.2] The abstract's '>95% reduction of network parameters' is not a parameter-count reduction. MCD evaluates the same set of parameters with different dropout masks; counting 100 sampled forward passes as 40M parameters (Section 4.2) is a nonstandard accounting. The ResNet-LLCM adds 100 linear classifiers to the shared ResNet-50 backbone, so its actual parameter count is larger than that of the corresponding single-head CNN, not 95% smaller. The legitimate efficiency claim is that LLCM obtains uncertainties with one backbone forward pass plus M head evaluations, rather than 100 stochastic forward passes. Please restate the efficiency claim in terms of inference-time compute or parameter-evaluations throughout the paper.
  3. [Tables 5 and 6] The claim that LLCM-100 performs 'near-identically' to MCD-100 rests on single runs with no reported variance across random seeds or training repetitions. For example, German Bank 2010 accuracy is 0.786 for LLCM versus 0.794 for MCD on 500 test images, a difference of about 4 images, and Substrate accuracy is 0.879 versus 0.885. Without seed-level standard deviations, confidence intervals, or a paired comparison, the equal-performance claim is not statistically supported. Please report means and standard deviations over at least several seeds for all ensemble methods, or explicitly present the comparison as descriptive rather than as evidence of equivalence.
  4. [Figures 5 and 7] The uncertainty metric plots do not include a baseline that is cheaper than LLCM, such as the max-softmax or softmax entropy of the deterministic CNN. Since the paper's practical contribution is a prioritized list of uncertain samples for human review, the reader needs to know how much of the selective-accuracy benefit comes from the 100-headed architecture rather than from simple confidence scores of a single trained network. Adding this baseline to Figs. 5 and 7 would directly address the concern that, if committee members are nearly identical, LLCM uncertainty scores reduce to the entropy of a shared softmax.
minor comments (5)
  1. [Eq. (8)] The symbol M is used both for the number of committee members and for the number of calibration bins in Eq. (8); please rename the bin index or the bin count to avoid confusion.
  2. [Section 4.2] The phrase 'ca.400K network parameters which is sampled 100x (i.e., 40M network parameters)' should be replaced by a statement about parameter-evaluations or multiply-accumulate operations, since multiplying a parameter count by the number of stochastic forward passes is not a standard parameter count.
  3. [Section 3.3; Tables 5 and 6] The BMA-100 results for the benthic datasets use a last-layer Bayesian network, not a fully Bayesian neural network; please clarify this in the text and in the table labels to avoid terminological confusion.
  4. [Figure captions] The captions for Figs. 2 and 4 say 'using class weights/logit normalization/label smoothing' without indicating which panel corresponds to which combination; please specify this in each panel caption or in the figure legend.
  5. [Throughout] There are several typos and nonstandard phrasings, including 'resulted with' for 'resulted in', 'int he next section' for 'in the next section', 'dateset' for 'dataset', 'preformed' for 'performed', and 'to identity' for 'to identify'; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the LLCM accuracy and uncertainty comparisons are evaluated against held-out labels, and the parameter-reduction claim is arithmetic.

full rationale

The paper's central comparisons are empirically grounded rather than definitionally circular. The LLCM posterior (Eq. 5) is the standard ensemble average of per-member softmax distributions, the training loss (Eq. 4) is a sum of independently backpropagated cross-entropy terms, and the claimed >95% parameter reduction is a direct arithmetic comparison of how many parameters are evaluated for MCD (full network sampled 100x) versus LLCM (shared trunk evaluated once with 100 linear heads). Accuracy, NLL, Brier score, ECE, and the uncertainty metric plots are all computed against held-out test labels, and temperature calibration is performed on a validation split, so no fitted value is renamed as a prediction. Self-citations to BenthicNet (Lowe et al., 2025; Misiuk et al., 2024) and the pre-trained ResNet-50 (Xu et al., 2024) supply the dataset and backbone, but they do not by themselves justify the uncertainty claim; the LLCM is independently compared to MCD and BMA on the same data. The epistemic-uncertainty contribution rests on the diversity analysis of committee-member weights and SVD norms, which is a proxy rather than a direct measure of predictive disagreement, and one could argue the evidence is weak, especially since reported CV∥SV∥F values are very small (2.7e-3 to 3.6e-3 for LLCM-10 on MNIST). That is an evidential limitation, not a circular reduction: the conclusion does not follow by construction from the diversity metric, nor is the metric defined in terms of the conclusion. No load-bearing step equates a fitted parameter with a predicted quantity, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the chosen design. Accordingly, the paper is self-contained against external test labels and benchmarks, and no circular step is identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper rests on standard empirical deep-learning assumptions and the transferability of pre-trained BenthicNet features. The main tuned quantities are committee size, dropout rates, and an unspecified loss-scaling factor; temperature scaling is fitted on the validation set per model. No new entities are introduced.

free parameters (4)
  • Committee size M = 100
    M=100 for all experiments (Table 2); no ablation over M is reported, so sensitivity to this choice is unknown.
  • Loss scaling factor = Unspecified
    Section 3.3 says the total committee loss is scaled by M and calls the factor an additional hyperparameter, but does not state the value or formula.
  • Dropout rates = 0.1 (MNIST), 0.01 (ResNet)
    Chosen per dataset family (Table 2); MCD uses the same training rate at inference.
  • Temperature scaling = Optimized per model in [0.01, 10.0]
    Bayesian optimization on the validation set; standard post-hoc calibration, but the LLCM uses a separate temperature per committee member, adding 100 tuned values.
assumptions (4)
  • domain assumption Softmax averaging over committee members approximates the posterior predictive distribution (Eq. 5).
    The paper equates 1/M times the sum of softmax outputs with p(y|x,D) as in Bayesian model averaging, a standard deep-ensemble approximation that is not verified under distribution shift.
  • domain assumption The pre-trained BenthicNet ResNet-50 representation transfers sufficiently to German Bank 2010 and Substrate (depth 2).
    The whole network is fine-tuned at low learning rate, so the claim is not about a frozen trunk; but the LLCM's value depends on the shared features being meaningful for both sub-datasets.
  • standard math Adam with OneCycleLR reliably optimizes the summed committee loss (Eq. 4).
    Standard empirical optimization assumption; the loss is a sum of identical-structure cross-entropy terms, which is treated as jointly optimizable.
  • domain assumption Test labels are treated as ground truth for accuracy and uncertainty evaluation.
    Section 2 acknowledges annotation ambiguity and inter-user disagreement, yet all metrics use the CATAMI labels as correct, which could bias the uncertainty evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Last-layer committee machines for uncertainty estimations of benthic imagery." pith.science (2026). https://pith.science/paper/U2I6VFL5

@misc{pith2026250416952,
  author       = {Pith},
  title        = {Pith review of: Last-layer committee machines for uncertainty estimations of benthic imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U2I6VFL5}},
  note         = {Machine review of arXiv:2504.16952}
}
read the original abstract

Automating the annotation of benthic imagery (i.e., images of the seafloor and its associated organisms, habitats, and geological features) is critical for monitoring rapidly changing ocean ecosystems. Deep learning approaches have succeeded in this purpose; however, consistent annotation remains challenging due to ambiguous seafloor images, potential inter-user annotation disagreements, and out-of-distribution samples. Marine scientists implementing deep learning models often obtain predictions based on one-hot representations trained using a cross-entropy loss objective with softmax normalization, resulting with a single set of model parameters. While efficient, this approach may lead to overconfident predictions for context-challenging datasets, raising reliability concerns that present risks for downstream tasks such as benthic habitat mapping and marine spatial planning. In this study, we investigated classification uncertainty as a tool to improve the labeling of benthic habitat imagery. We developed a framework for two challenging sub-datasets of the recently publicly available BenthicNet dataset using Bayesian neural networks, Monte Carlo dropout inference sampling, and a proposed single last-layer committee machine. This approach resulted with a > 95% reduction of network parameters to obtain per-sample uncertainties while obtaining near-identical performance compared to computationally more expensive strategies such as Bayesian neural networks, Monte Carlo dropout, and deep ensembles. The method proposed in this research provides a strategy for obtaining prioritized lists of uncertain samples for human-in-the-loop interventions to identify ambiguous, mislabeled, out-of-distribution, and/or difficult images for enhancing existing annotation tools for benthic mapping and other applications.

Figures

Figures reproduced from arXiv: 2504.16952 by the authors.

Figure 1
Figure 1. Network diversity (CVSVi M and CV∥SV∥F ) of a LLCM-10 and the MNIST dataset. Panel (a) shows the results of having identically initialized members, whereas panels (b–d) shows the results of random initialization us￾ing class weights / logit normalization / label smoothing (amount of smoothing, 0.1). These hyperparameters are either applied (denoted by +) or omitted (de￾noted by −). CV across all committee members m … view at source ↗
Figure 2
Figure 2. Reliability diagrams for the MNIST dataset using a LLCM-100 network (p = 0.1). Non-calibrated (left panels) and calibrated (right panels) using class weights / logit normalization / label smoothing (amount of smoothing, 0.1). These hyperparameters are either applied (denoted by +) or omitted (denoted by −). the baseline CNN. However, the advantage of the LLCM com￾pared to BMA or MCD is the reduced compute requiremen… view at source ↗
Figure 3
Figure 3. Uncertain metric plots for the BMA-100, MCD-100 (−/ − /−), and LLCM-100 (−/ − /−) networks and the MNIST dataset using class weights / logit normalization / label smoothing (amount of smoothing, 0.1). These hyper￾parameters are either applied (denoted by +) or omitted (denoted by −). Each trace is created by applying increasing thresholds (step size 0.001) of confi￾dence values and re-calculating metrics. In summary… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Reliability diagrams for the German Bank 2010 dataset using a LLCM￾100 network (p = 0.01). Non-calibrated (left panels) and calibrated (right panels) using class weights / logit normalization / label smoothing (amount of smoothing, 0.1). These hyperparameters are eithe…
Figure 5
Figure 5. Figure 5: Uncertain metric plots for the BMA-100, MCD-100 (−/ − /−), and LLCM-100 (−/ − /−) networks and the German Bank 2010 dataset using class weights / logit normalization / label smoothing (amount of smoothing, 0.1). These hyperparameters are either applied (denoted by +) o…
Figure 6
Figure 6. Figure 6: Reliability diagrams for the Substrate (depth 2) dataset using a LLCM￾100 network (p = 0.01). Non-calibrated (left panels) and calibrated (right panels) using class weights / logit normalization / label smoothing (amount of smoothing, 0.1). These hyperparameters are ei…
Figure 7
Figure 7. Figure 7: Uncertain metric plots for the BMA-100, MCD-100 (−/ − /−), and LLCM-100 (−/ − /−) networks and the Substrate (depth 2) dataset using class weights / logit normalization / label smoothing (amount of smoothing, 0.1). These hyperparameters are either applied (denoted by +…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Improving Detection of Rare Nodes in Hierarchical Multi-Label Learning

    cs.LG 2026-02 conditional novelty 6.0 of 10

    A node-weighted loss combining inverse-frequency weighting and ensemble-uncertainty focal terms improves recall of rare classes in hierarchical multi-label models by up to ~5x.

Reference graph

Works this paper leans on

50 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    , author Pourpanah, F

    author Abdar, M. , author Pourpanah, F. , author Hussain, S. , author Rezazadegan, D. , author Liu, L. , author Ghavamzadeh, M. , author Fieguth, P. , author Cao, X. , author Khosravi, A. , author Acharya, U.R. , author Makarenkov, V. , author Nahavandi, S. , year 2021 . title A review of uncertainty quantification in deep learning: Techniques, applicatio...

    doi:10.1016/j.inffus.2021.05.008 2021

  2. [2]

    , author Hill, N

    author Althaus, F. , author Hill, N. , author Edwards, L. , author Ferrari, R. , year 2014 . title CATAMI classification scheme for scoring marine biota and substrata in underwater imagery: A pictorial guide to the collaborative and annotation tools foranalysis of marine imagery and video ( CATAMI ) classification scheme . type Technical Report . CATAMI.o...

    work page 2014

  3. [3]

    , author Hill, N

    author Althaus, F. , author Hill, N. , author Ferrari, R. , author Edwards, L. , author Przeslawski, R. , author Schönberg, C.H.L. , author Stuart-Smith, R. , author Barrett, N. , author Edgar, G. , author Colquhoun, J. , author Tran, M. , author Jordan, A. , author Rees, T. , author Gowlett-Holmes, K. , year 2015 . title A standardised vocabulary for ide...

    2015

  4. [4]

    , author Pitas, K

    author Arbel, J. , author Pitas, K. , author Vladimirova, M. , author Fortuin, V. , year 2023 . title A primer on B ayesian neural networks: review and debates . howpublished arXiv preprint . :10.48550/arXiv.2309.16314

    arXiv doi:10.48550/arxiv.2309.16314 2023

  5. [5]

    , author Hobley, B

    author Arosio, R. , author Hobley, B. , author Wheeler, A.J. , author Sacchetti, F. , author Conti, L.A. , author Furey, T. , author Lim, A. , year 2023 . title Fully convolutional neural networks applied to large-scale marine morphology mapping . journal Frontiers in Marine Science volume 10 . :10.3389/fmars.2023.1228867

    work page arXiv 2023

  6. [6]

    , author Treibitz, T

    author Beijbom, O. , author Treibitz, T. , author Kline, D.I. , author Eyal, G. , author Khen, A. , author Neal, B. , author Loya, Y. , author Mitchell, B.G. , author Kriegman, D. , year 2016 . title Improving automated annotation of benthic survey images using wide-band fluorescence . journal Scientific Reports volume 6 , pages 1--11 . :10.1038/srep23166

    doi:10.1038/srep23166 2016

  7. [7]

    , author Chen, J.P

    author Bingham, E. , author Chen, J.P. , author Jankowiak, M. , author Obermeyer, F. , author Pradhan, N. , author Karaletsos, T. , author Singh, R. , author Szerlip, P. , author Horsfall, P. , author Goodman, N.D. , year 2019 . title Pyro: deep universal probabilistic programming . journal Journal of Machine Learning Research volume 20 , pages 973--978 ....

    work page 2019

  8. [8]

    , author Sameoto, J.A

    author Brown, C.J. , author Sameoto, J.A. , author Smith, S.J. , year 2012 . title Multiple methods, maps, and management applications: Purpose made seafloor maps in support of ocean management . journal Journal of Sea Research volume 72 , pages 1--13 . :10.1016/j.seares.2012.04.009

    doi:10.1016/j.seares.2012.04.009 2012

Show all 50 references
  1. [9]

    , author Smith, S.J

    author Brown, C.J. , author Smith, S.J. , author Lawton, P. , author Anderson, J.T. , year 2011 . title Benthic habitat mapping: a review of progress towards improved understanding of the spatial ecology of the seafloor using acoustic techniques . journal Estuarine, Coastal an...

    2011 doi

  2. [10]

    , author Beyer, L

    author Dosovitskiy, A. , author Beyer, L. , author Kolesnikov, A. , author Weissenborn, D. , author Zhai, X. , author Unterthiner, T. , author Dehghani, M. , author Minderer, M. , author Heigold, G. , author Gelly, S. , author Uszkoreit, J. , author Houlsby, N. , year 2021 . t...

    2021 doi arXiv

  3. [11]

    , author Maulik, R

    author Egele, R. , author Maulik, R. , author Raghavan, K. , author Lusch, B. , author Guyon, I. , author Balaprakash, P. , year 2021 . title AutoDEUQ : automated deep ensemble with uncertainty quantification . howpublished arXiv preprint . :10.48550/arXiv.2110.13511

    2021 doi arXiv

  4. [12]

    , author Ghahramani, Z

    author Gal, Y. , author Ghahramani, Z. , year 2015 . title Dropout as a B ayesian approximation: representing model uncertainty in deep learning , in: booktitle International Conference on Machine Learning (ICML) , pp. pages 1--12 . :10.48550/arXiv.1506.02142

    2015 doi arXiv

  5. [13]

    , author Hron, J

    author Gal, Y. , author Hron, J. , author Kendall, A. , year 2017 . title Concrete dropout . howpublished arXiv preprint . :10.48550/arXiv.1705.07832

    2017 doi arXiv

  6. [14]

    , author Tassi, C.R.N

    author Gawlikowski, J. , author Tassi, C.R.N. , author Ali, M. , author Lee, J. , author Humt, M. , author Feng, J. , author Kruspe, A. , author Triebel, R. , author Jung, P. , author Roscher, R. , author Shahzad, M. , author Yang, W. , author Bamler, R. , author Zhu, X.X. , y...

    2023 doi

  7. [15]

    , author Pleiss, G

    author Guo, C. , author Pleiss, G. , author Sun, Y. , author Weinberger, K.Q. , year 2017 . title On calibration of modern neural networks , in: booktitle International Conference on Machine Learning (ICML) , pp. pages 1--14 . :10.48550/arXiv.1706.04599

    2017 doi arXiv

  8. [16]

    , author Willes, J

    author Harrison, J. , author Willes, J. , author Snoek, J. , year 2024 . title Variational B ayesian last layers , in: booktitle International Conference on Learning Representations (ICLR) , pp. pages 1--33 . :10.48550/arXiv.2404.11599

    2024 doi arXiv

  9. [17]

    , author Zhang, X

    author He, K. , author Zhang, X. , author Ren, S. , author Sun, J. , year 2016 . title Deep residual learning for image recognition , in: booktitle Conference on Computer Vision and Pattern Recognition (CVPR) , pp. pages 770--778 . :10.1109/CVPR.2016.90

    2016 doi

  10. [18]

    , author Wright, D

    author Holm, A.N. , author Wright, D. , author Augenstein, I. , year 2023 . title Revisiting softmax for uncertainty approximation in text classification . journal Information volume 14 . :10.3390/info14070420

    2023 doi

  11. [19]

    , author Li, Y

    author Huang, G. , author Li, Y. , author Pleiss, G. , author Liu, Z. , author Hopcroft, J.E. , author Weinberger, K.Q. , year 2017 . title Snapshot ensembles: train 1, get M for free , in: booktitle International Conference on Learning Representations (ICLR) , pp. pages 1--14...

    2017 doi arXiv

  12. [20]

    , author Brooke, B.P

    author Huang, Z. , author Brooke, B.P. , author Harris, P.T. , year 2011 . title A new approach to mapping marine benthic habitats using physical environmental data . journal Continental Shelf Research volume 31 , pages S4--S16 . :10.1016/j.csr.2010.03.012

    2011 doi

  13. [21]

    , author Johansen, A.S

    author Humblot-Renaux, G. , author Johansen, A.S. , author Schmidt, J.E. , author Irlind, A.F. , author Madsen, N. , author Moeslund, T.B. , author Pedersen, M. , year 2024 . title Underwater uncertainty: a multi-annotator image dataset for benthic habitat classification , in:...

    2024

  14. [22]

    , author Waegeman, W

    author Hüllermeier, E. , author Waegeman, W. , year 2021 . title Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods , in: booktitle Machine Learning . publisher Springer . volume volume 110 , pp. pages 457--506 . :10.1038/s41597-02...

    2021 doi

  15. [23]

    , author Althaus, F

    author Jackett, C. , author Althaus, F. , author Maguire, K. , author Farazi, M. , author Scoulding, B. , author Untiedt, C. , author Ryan, T. , author Shanks, P. , author Brodie, P. , author Williams, A. , year 2023 . title A benthic substrate classification method for seabed...

    2023

  16. [24]

    , author Buntine, W

    author Jospin, L.V. , author Buntine, W. , author Boussaid, F. , author Laga, H. , author Bennamoun, M. , year 2020 . title Hands-on B ayesian neural networks: A tutorial for deep learning users . :10.1109/MCI.2022.3155327

    2020

  17. [25]

    , author Orenstein, E

    author Katija, K. , author Orenstein, E. , author Schlining, B. , author Lundsten, L. , author Barnard, K. , author Sainz, G. , author Boulais, O. , author Cromwell, M. , author Butler, E. , author Woodward, B. , author Bell, K.L.C. , year 2022 . title Fathomnet: A global imag...

    2022 doi

  18. [26]

    , author Pritzel, A

    author Lakshminarayanan, B. , author Pritzel, A. , author Blundell, C. , year 2017 . title Simple and scalable predictive uncertainty estimation using deep ensembles , in: booktitle Conference on Neural Information Processing Systems (NeurIPS) , pp. pages 1--15 . :10.48550/arX...

    2017 doi arXiv

  19. [27]

    , author Purushwalkam, S

    author Lee, S. , author Purushwalkam, S. , author Cogswell, M. , author Crandall, D. , author Batra, D. , year 2015 . title Why M heads are better than one: training a diverse ensemble of deep networks . howpublished arXiv preprint . :10.48550/arXiv.1511.06314

    2015 doi arXiv

  20. [28]

    , editor Wang, F

    editor Li, X. , editor Wang, F. (Eds.), year 2023 . title Artificial intelligence oceanography . publisher Springer Nature Singapore . :10.1007/978-981-19-6375-9

    2023 doi

  21. [29]

    , author Wang, L

    author Liu, P. , author Wang, L. , author Ranjan, R. , author He, G. , author Zhao, L. , year 2022 . title A survey on active deep learning: From model driven to data driven . journal ACM Computing Surveys volume 54 , pages 1--34 . :10.1145/3510414

    2022 doi

  22. [30]

    , author Kumar, M

    author Louppe, G. , author Kumar, M. , year 2016 . title Bayesian optimization with skopt . https://scikit-optimize.github.io/stable/auto_examples/bayesian-optimization.html

    2016

  23. [31]

    , author Misiuk, B

    author Lowe, S.C. , author Misiuk, B. , author Xu, I. , author Abdulazizov, S. , author Baroi, A.R. , author Bastos, A.C. , author Best, M. , author Ferrini, V. , author Friedman, A. , author Hart, D. , author Hoegh-Guldberg, O. , author Ierodiaconou, D. , author Mackin-McLaug...

    2025

  24. [32]

    , author Brown, C.J

    author Misiuk, B. , author Brown, C.J. , year 2024 . title Benthic habitat mapping: a review of three decades of mapping biological patterns on the seafloor . journal Estuarine, Coastal and Shelf Science volume 296 , pages 1--27 . :10.1016/j.ecss.2023.108599

    2024

  25. [33]

    , author Lowe, S

    author Misiuk, B. , author Lowe, S. , author Xu, I. , year 2024 . title Benthic N et . howpublished Federated Research Data Repository . :10.20383/103.0614

    2024 doi

  26. [34]

    , year 1995

    author Neal, R.M. , year 1995 . title Bayesian learning for neural networks . type Phd thesis . University of Toronto

    1995

  27. [35]

    , author Fertig, E

    author Ovadia, Y. , author Fertig, E. , author Ren, J. , author Nado, Z. , author Sculley, D. , author Nowozin, S. , author Dillon, J.V. , author Lakshminarayanan, B. , author Snoek, J. , year 2019 . title Can you trust your model's uncertainty? evaluating predictive uncertain...

    2019 doi arXiv

  28. [36]

    , author Leibfried, F

    author Pearce, T. , author Leibfried, F. , author Brintrup, A. , author Zaki, M. , author Neely, A. , year 2018 . title Uncertainty in neural networks: Approximately bayesian ensembling , in: booktitle International Conference on Artificial Intelligence and Statistics (AISTATS...

    2018 doi arXiv

  29. [37]

    title MNIST

    author PyTorch , year 2025 a. title MNIST . https://pytorch.org/vision/main/generated/torchvision.datasets.MNIST.html

    2025

  30. [38]

    title OneCycleLR

    author PyTorch , year 2025 b. title OneCycleLR . https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.OneCycleLR.html

    2025

  31. [39]

    , author Dubinsky, Z

    author Raphael, A. , author Dubinsky, Z. , author Iluz, D. , author Benichou, J.I.C. , author Netanyahu, N.S. , year 2020 . title Deep neural network recognition of shallow water corals in the G ulf of E ilat ( A qaba) . journal Scientific Reports volume 10 , pages 1--11 . :10...

    2020 doi

  32. [40]

    , author Zisserman, A

    author Simonyan, K. , author Zisserman, A. , year 2014 . title Very deep convolutional networks for large-scale image recognition , in: booktitle International Conference on Learning Representations (ICLR) , pp. pages 1--14 . :10.48550/arXiv.1409.1556

    2014 doi arXiv

  33. [41]

    , author Topin, N

    author Smith, L.N. , author Topin, N. , year 2017 . title Super-convergence: Very fast training of neural networks using large learning rates . :10.48550/arXiv.1708.07120

    2017 doi arXiv

  34. [42]

    , author Knoll, C

    author Steger, S. , author Knoll, C. , author Klein, B. , author Fröning, H. , author Pernkopf, F. , year 2024 . title Function space diversity for uncertainty prediction via repulsive last-layer ensembles , in: booktitle Structured Probabilistic Inference & Generative Modelin...

    2024 doi arXiv

  35. [43]

    , author Vanhoucke, V

    author Szegedy, C. , author Vanhoucke, V. , author Ioffe, S. , author Shlens, J. , author Wojna, Z. , year 2015 . title Rethinking the I nception architecture for computer vision . howpublished arXiv preprint . :10.48550/arXiv.1512.00567

    2015 doi arXiv

  36. [44]

    , year 2023

    author Valdenegro-Toro, M. , year 2023 . title Sub-ensembles for fast uncertainty estimation in neural networks , in: booktitle Proceedings of the International Conference on Computer Vision (ICCV) Workshops , pp. pages 4119--4127 . :https://openaccess.thecvf.com/content/ICCV2...

    2023

  37. [45]

    , author Xie, R

    author Wei, H. , author Xie, R. , author Cheng, H. , author Feng, L. , author An, B. , author Li, Y. , year 2022 . title Mitigating neural network overconfidence with logit normalization . howpublished arXiv preprint . :10.48550/arXiv.2205.09310

    2022 doi arXiv

  38. [46]

    , author Tran, D

    author Wen, Y. , author Tran, D. , author Ba, J. , year 2020 . title BatchEnsemble : an alternative approach to efficient ensemble and lifelong learning . howpublished arXiv preprint . :10.48550/arXiv.2002.06715

    2020 doi arXiv

  39. [47]

    , author Dai, M

    author Winther, J.G. , author Dai, M. , author Rist, T. , author Hoel, A.H. , author Li, Y. , author Trice, A. , author Morrissey, K. , author Juinio-Meñez, M.A. , author Fernandes, L. , author Unger, S. , author Scarano, F.R. , author Halpin, P. , author Whitehouse, S. , year...

    2020 doi

  40. [48]

    , author Ma, Z

    author Xie, J. , author Ma, Z. , author Lei, J. , author Zhang, G. , author Xue, J.H. , author Tan, Z.H. , author Guo, J. , year 2021 . title Advanced dropout: a model-free methodology for B ayesian dropout optimization . journal IEEE Transactions on Pattern Analysis and Machi...

    2021

  41. [49]

    , author Misiuk, B

    author Xu, I. , author Misiuk, B. , author Lowe, S.C. , author Gillis, H.M. , author Brown, C.J. , author Trappenberg, T. , year 2024 . title Hierarchical multi-label classification with missing information for benthic habitat imagery , in: booktitle International Joint Confer...

    2024

  42. [50]

    , author Rahman, A.U

    author Yasir, M. , author Rahman, A.U. , author Gohar, M. , year 2021 . title Habitat mapping using deep neural networks . journal Multimedia Systems volume 27 , pages 679--690 . :10.1007/s00530-020-00695-0

    2021 doi

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.