Pith. sign in

REVIEW 3 major objections 5 minor 43 references

ConBatch-BAL: Batch Bayesian Active Learning under Budget Constraints

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding two simple budget-allocation heuristics on top of the Batch-BALD acquisition function makes batch Bayesian active learning practical under varying annotation costs and budget limits, and that on geolocated…

desk verdict Honest, useful empirical study of two simple budget-aware Batch-BALD wrappers; the new datasets are the real contribution, but the missing random-score ablation leaves the central claim under-supported. read the letter →

arxiv 2507.04929 v1 pith:KG6TUVBB submitted 2025-07-07 cs.LG cs.CV

Pablo G. Morato , Charalampos P. Andriotis , Seyran Khademi This is my paper · ORCID
classification cs.LGcs.CV
keywords activelearningbatchbudgetconstraintsmutualinformationMonteCarlodropoutbuildingenergyefficiencyaerialimagerygeospatialdata
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

This paper claims that batch Bayesian active learning can be made practical under per-sample annotation cost variation and budget limits by adding two simple budget-allocation heuristics on top of the Batch-BALD acquisition function: dynamic thresholding, which redistributes the remaining budget across batch steps, and greedy selection, which picks the highest-scoring affordable sample at each step. The authors argue this matters because real annotation campaigns, such as drone-based building inspection, face travel and cost constraints that standard active learning ignores. To test this, they introduce two geolocated aerial-image datasets of Rotterdam buildings and compare the heuristics against a random acquisition baseline under distance, return-trip, and area-cost constraints. On these datasets the constrained heuristics reach target accuracies with 20–43% fewer active learning iterations on build6k and 50–80% fewer on nieman17k than random selection, and under a 2 km budget they even beat the unconstrained random baseline. The paper positions ConBatch-BAL as a general framework because the selection metric and cost model can be replaced.

What carries the argument

The load-bearing components are the two budget-allocation heuristics and the acquisition function they wrap. Dynamic thresholding ConBatch-BAL maintains an adaptive cost threshold $c_{th}=c_{max}/(n_{max}-(i-1))$ that rebalances the remaining budget across remaining batch steps, while greedy ConBatch-BAL selects the highest mutual-information sample among those affordable under the current remaining budget. Both score candidates with the Batch-BALD acquisition function, mutual information between batch predictions and model parameters, computed via joint entropy minus expected conditional entropy. Uncertainty comes from a Monte Carlo dropout Bayesian neural network, treated as an approximate posterior, operating on 384-dimensional DINOv2 features extracted from the aerial images. The cost configurations—distance, distance-return, and area cost—encode how a selected set maps to annotation expenditure, and they turn batch selection into a knapsack-style problem that the heuristics solve sequentially.

What would settle it

Re-run the same benchmark protocol with uncertainty computed by a stronger approximate inference method, such as stochastic-gradient MCMC or deep ensembles, replacing MC dropout, and measure whether ConBatch-BAL still beats random selection on build6k and nieman17k. If the gap shrinks to near zero, the budget heuristics add little; if it persists, MC-dropout noise is not the driver.

Watch

Extended reading notes

Core claim

At its core, the paper establishes that the combinatorial problem of selecting a batch of samples under a budget—where the cost of a sample may depend on previously chosen samples and the uncertainty metric is joint mutual information—can be handled by two straightforward sequential heuristics with no exact optimization. The dynamic thresholding strategy sets an initial per-step cost threshold equal to the total budget divided by the maximum batch size, and after each selection resets the threshold to the remaining budget divided by the remaining steps. The greedy strategy simply restricts attention at each step to samples whose cost does not exceed the remaining budget. Both retain the Batch-BALD mutual information score as the acquisition metric computed on a Monte Carlo dropout BNN fed with frozen DINOv2 embeddings. The paper's central empirical claim is that in the distance cost configuration on the released building datasets, these heuristics reduce the number of active learning iterations needed to hit accuracy targets by 20–43% (build6k) and 50–80% (nieman17k) relative to random acquisition, and that constrained runs with a 2 km budget outperform the unconstrained random baseline on all datasets.

Load-bearing premise

The argument assumes that the mutual-information scores from a Monte Carlo dropout network on frozen DINOv2 features rank annotation value correctly on these noisy aerial building images; if that ranking is unreliable, the reported advantage over random selection could vanish regardless of the budget heuristics.

Editorial extensions

If this is right

  • Under a 2 km batch distance constraint, both ConBatch-BAL strategies reach the tested accuracy targets in fewer active learning iterations than the unconstrained random baseline on all three datasets.
  • The advantage over random selection grows at higher accuracy targets and is largest on the cleaner mnist6k benchmark, whereas the noisier building datasets narrow the gap.
  • In the area-cost configuration, greedy ConBatch-BAL generally beats dynamic thresholding because thresholds can lock out costly but informative samples; mnist6k is the exception where thresholding wins.
  • For the distance-return configuration, the same ordering holds but all strategies need more iterations because the return trip tightens the budget.
  • Effectiveness relative to random selection persists across batch sizes of 2, 5, and 10 on build6k.

Reading between the lines

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

  • If the 2 km finding generalizes, city-scale building surveys can budget drone flights by local travel radius rather than per-building fees, since informative buildings appear to cluster within a few kilometers.
  • The framework's separation of acquisition function from cost model suggests the same two heuristics could be dropped onto other acquisition metrics such as BADGE or k-BALD, or onto non-geospatial cost models like per-image expert review time or API pricing.
  • Because the authors only retrain a small BNN on frozen DINOv2 features, the reported iteration savings may mostly reflect the quality of the uncertainty signal; a test with stronger posterior approximations would show whether the heuristics themselves or the MC-dropout ranking drive the gains.
  • The dynamic thresholding failure mode on build6k area-cost—missing informative expensive samples—suggests a natural hybrid: run greedy when informative samples are spatially concentrated, and thresholding when they are dispersed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper proposes two heuristics for batch Bayesian active learning under annotation budget constraints: dynamic thresholding ConBatch-BAL and greedy ConBatch-BAL, both built on the Batch-BALD mutual-information acquisition function with Monte Carlo dropout BNNs. The budget is modeled as a cost constraint on the batch (e.g., total travel distance or summed area-based costs). The authors benchmark the two strategies against random selection on three datasets—two new geolocated aerial-image building datasets (build6k, nieman17k) and a geolocated MNIST variant—under three cost configurations and multiple budgets, reporting that the proposed strategies reach accuracy targets with fewer active learning iterations than random selection, sometimes even outperforming the unconstrained random baseline. They also release the two building datasets, the DINOv2 embeddings, and code for reproducibility.

Significance. If the empirical claims hold, the paper makes two useful contributions: a simple, practical recipe for incorporating budget constraints into Batch-BALD-style acquisition, and two new real-world geolocated building datasets for cost-sensitive active learning research. The experimental scope is broad: three datasets, three cost models, several budget levels, two additional batch sizes, and five seeds per condition, with code and data released. The strengths are the resource contribution and the transparent reproducibility statement. The main open question is whether the reported gains come from the Bayesian mutual-information ranking or simply from the budget-aware spatial selection mechanics; the paper does not yet provide the control experiment needed to separate these, and the statistical evidence is limited to five seeds without significance tests.

major comments (3)
  1. [Section 4.1, Algorithm 1] In Algorithm 1, the early-exit condition is "∀x ∈ Dpool \ Ai−1, c(x) > cmax", but the inner selection loop only considers candidates with c(x) ≤ cth. If there exists a candidate with cth < c(x) ≤ cmax and no candidate with c(x) ≤ cth, the foreach set is empty and the subsequent argmax is taken over an empty set, so the algorithm is undefined. The stopping condition should be stated with respect to cth, or the algorithm should include an explicit branch for the case where no candidate is affordable under the current threshold (e.g., relaxing the threshold or terminating). This is a correctness issue in a central contribution and should be fixed in the pseudocode and discussed in the text.
  2. [Section 6; Appendix C.4; Section 7.1] The paper attributes the observed reductions in active learning iterations to the Bayesian mutual-information ranking, but it compares only against random selection and unconstrained Batch-BALD. Under strict budgets, the random baseline acquires very few points per batch (Appendix C.4, Fig. 15), so any strategy that selects nearby affordable points—even with uninformative scores—could beat random. The limitations section (7.1) itself concedes that the MC-dropout uncertainty estimates "might not be as informative." To make the central claim credible, the authors should add a control baseline that uses the same budget/spatial selection mechanics (greedy or thresholding) with random acquisition scores; if such a baseline matches the ConBatch-BAL results, the advantage is not due to the uncertainty ranking.
  3. [Section 6; Figures 3-4; Appendix C.2] The headline reductions (e.g., 20-43% and 50-80% fewer iterations) are computed from means over only 5 seeds, with no significance tests or confidence intervals reported. The accuracy targets (e.g., 0.76/0.78/0.80 for build6k; 0.93/0.95/0.97 for mnist6k) are chosen without a stated rule, and the number of iterations-to-target is sensitive to the target's position relative to the learning-curve plateau. The authors should report per-condition variance, perform paired significance tests or bootstrap intervals, and either justify the targets or show the comparison over a range of targets.
minor comments (5)
  1. [Abstract / Section 8] The abstract and conclusion state that the strategies can "even outperform the unconstrained baseline solutions," but the results in Figure 7a show dynamic thresholding underperforming random selection under the same budget in the area-cost configuration on build6k; please qualify this claim to the configurations and strategies for which it holds.
  2. [Section 5; Figure 2b] The cost models are described verbally, but the incremental cost c(x) used in Algorithms 1 and 2 is not formally defined for the distance and distance-return configurations (e.g., distance to the previously selected point); a precise definition would improve reproducibility.
  3. [Appendix B.2] Hyperparameters are tuned on build6k under an infinite budget and then fixed across all datasets and configurations; please comment on the sensitivity of the conclusions to this choice.
  4. [Section 4; Section 4.1] There are minor typographical issues: "ConBatch- BAL" appears with an extra space in Section 4, and the sentence "The steps for implementing this active learning strategy 2 are outlined in Algorithm 2" is awkwardly worded.
  5. [Section 4.1, footnote] The complexity expression "O(|Dpool|·T·K nmax)" is ambiguous; please clarify whether K is raised to the power nmax or multiplied.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ConBatch-BAL heuristics are not derived from the evaluation metric, and the acquisition scoring is imported from external work (Batch-BALD, MC-dropout); the empirical claims rest on new datasets and benchmark comparisons, not on fitted or self-referential constructs.

full rationale

The paper's derivation chain is self-contained and non-circular. The acquisition function is the Batch-BALD mutual information (Eq. 3-4), attributed to the external prior work of Kirsch et al. [20]; the uncertainty estimates come from Monte Carlo dropout, attributed to Gal and Ghahramani [13]. Algorithms 1 and 2 are explicit heuristic procedures for imposing budget constraints on this external acquisition score; no parameter is fitted to the reported evaluation metric (number of active learning iterations to reach a fixed accuracy target). The ConBatch-BAL strategies maximize mutual information, not the evaluation target, so the reported reductions in iterations are not forced by construction. The hyperparameters are tuned once on build6k under an infinite budget and then held fixed across all datasets and constraints, which is a standard experimental configuration choice rather than a fitted-input-as-prediction pattern. There is no load-bearing self-citation: none of the reference list is authored by the present paper's authors, and no uniqueness theorem or prior result by the same group is invoked to forbid alternative designs. The main weakness identified in the skeptic summary, namely that the budget/spatial mechanics may be responsible for the gains rather than the mutual-information scores, is a missing-ablation concern about attribution, not circularity: the random baseline is an independent stochastic acquisition process, and the paper explicitly concedes in Section 7.1 that MC-dropout uncertainty estimates 'might not be as informative.' That concession weakens the strength of the central interpretive claim but does not make the derivation equivalent to its inputs. Therefore, no circular step can be exhibited under the required standard, and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The strategies themselves contain no fitted constants. The evaluation, however, rests on hand-chosen accuracy targets and model hyperparameters, plus four domain assumptions about uncertainty quality, embedding quality, and the cost model. All of these are stated in the paper, but they are input assumptions rather than derived results.

free parameters (2)
  • Accuracy targets per dataset = build6k: 0.76, 0.78, 0.80; mnist6k: 0.93, 0.95, 0.97; nieman17k: 0.65, 0.68, 0.71
    These targets are chosen without a stated selection rule and define the central comparisons. The reported 20-43% (build6k) and 50-80% (nieman17k) reductions are differences in the number of iterations needed to hit these exact thresholds; different thresholds could change or erase the headline numbers. See Figures 3, 4, 10 and Section 6.
  • BNN hyperparameters = 2 hidden layers, width 256, dropout 0.1, 100 forward passes, 5 batch steps
    Set by hand and fine-tuned on build6k under an infinite budget (Appendix B.2), then reused for all constrained experiments and datasets. The comparison's validity depends on these settings being appropriate for the test scenarios.
assumptions (4)
  • domain assumption Batch-BALD mutual information is an effective acquisition score for ranking annotation value.
    Invoked in Sections 3.2 and 4 as the acquisition function for both strategies; if the uncertainty estimates are noisy (as acknowledged in Section 7.1), the advantage over random selection could be weakened.
  • domain assumption Monte Carlo dropout provides a valid approximate Bayesian posterior.
    Section 3.3 adopts MC-dropout on the basis of [13]; the paper's limitations section notes these estimates may not be very informative.
  • domain assumption Frozen DINOv2 embeddings retain sufficient information for energy-efficiency and typology classification from aerial images.
    Section 6 and Appendix B.1 use a pretrained DINOv2 without fine-tuning for all datasets; if the embeddings are not discriminative for these labels, the whole active learning signal degrades.
  • domain assumption Annotation cost is well modeled by travel distance, distance with return, or area-based cost.
    Section 5 defines these three cost configurations; the paper notes the cost model is case-specific, so the observed gains may not transfer to other cost structures such as monetary fees or expert hours.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ConBatch-BAL: Batch Bayesian Active Learning under Budget Constraints." pith.science (2026). https://pith.science/paper/KG6TUVBB

@misc{pith2026250704929,
  author       = {Pith},
  title        = {Pith review of: ConBatch-BAL: Batch Bayesian Active Learning under Budget Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KG6TUVBB}},
  note         = {Machine review of arXiv:2507.04929}
}
read the original abstract

Varying annotation costs among data points and budget constraints can hinder the adoption of active learning strategies in real-world applications. This work introduces two Bayesian active learning strategies for batch acquisition under constraints (ConBatch-BAL), one based on dynamic thresholding and one following greedy acquisition. Both select samples using uncertainty metrics computed via Bayesian neural networks. The dynamic thresholding strategy redistributes the budget across the batch, while the greedy one selects the top-ranked sample at each step, limited by the remaining budget. Focusing on scenarios with costly data annotation and geospatial constraints, we also release two new real-world datasets containing geolocated aerial images of buildings, annotated with energy efficiency or typology classes. The ConBatch-BAL strategies are benchmarked against a random acquisition baseline on these datasets under various budget and cost scenarios. The results show that the developed ConBatch-BAL strategies can reduce active learning iterations and data acquisition costs in real-world settings, and even outperform the unconstrained baseline solutions.

Figures

Figures reproduced from arXiv: 2507.04929 by the authors.

Figure 1
Figure 1. Proposed batch Bayesian active learning strategies, dynamic thresholding ConBatch [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Released building datasets and implemented cost configurations. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Benchmark results for the distance cost configuration across datasets and budget con￾straints. Markers represent the number of active learning iterations required to reach model ac￾curacy targets on the test set for each seed, with the mean indicated by a vertical line. Greedy ConBatch-BAL and dynamic thresholding ConBatch-BAL outperform the random selection base￾line. The second configuration, distance return, impo… view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Benchmark results for the area cost configuration across datasets and budget constraints. Markers represent the number of active learning iterations required to reach model accuracy targets on the test set for each seed, with the mean indicated by a vertical line. Dyna…
Figure 5
Figure 5. Figure 5: Normalized mutual information as a function of geographical distance for candidate [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Mutual information for batches collected by ConBatch-BAL strategies on the [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Results for the area cost configuration on the build6k dataset under a budget constraint of 101 cost units. (Left) Dynamic thresholding underperforms greedy ConBatch-BAL and random selection strategies. (Right) Greedy ConBatch-BAL selects costly but informative buildin…
Figure 8
Figure 8. Figure 8: Datasets curation process. (Top) Building footprints and tabular data are cross-referenced [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Pipeline: from aerial images and embeddings obtained by the foundation model DINOv2 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Benchmark results for the distance return cost configuration across datasets and budget constraints. Markers represent the number of active learning iterations required to reach a target model accuracy on the test set for each seed, with the mean indicated by a vertic…
Figure 11
Figure 11. Figure 11: Active learning curves for all tested strategies across datasets and budget constraints. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Benchmark results for the distance configuration, showcasing the number of acquired samples and acquisition cost across datasets and budget constraints. Markers represent (top) the total number of acquired samples and (bottom) the traveled distance required to reach a…
Figure 13
Figure 13. Figure 13: Benchmark results for the area cost configuration, showcasing the number of acquired samples and acquisition cost across datasets and budget constraints. Markers represent (top) the total number of acquired samples and (bottom) the cost required to reach a target mode…
Figure 14
Figure 14. Figure 14: Benchmark results for the distance return configuration, showcasing the number of acquired samples and acquisition cost across datasets and budget constraints. Markers represent (top) the total number of acquired samples and (bottom) the traveled distance required to …
Figure 15
Figure 15. Figure 15: Batch acquisition cost (left) and acquired points (right) at each active learning iteration [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: Batch acquisition cost (left) and acquired points (right) at each active learning iteration [PITH_FULL_IMAGE:figures/full_fig_p026_16.png]
Figure 17
Figure 17. Figure 17: Batch acquisition cost (left) and acquired points (right) at each active learning iteration [PITH_FULL_IMAGE:figures/full_fig_p027_17.png]
Figure 18
Figure 18. Figure 18: Benchmark results for the build 6k dataset under the distance configuration, with varying budget constraints and batch sizes (2, 5, and 10). Markers represent (top) the number of active learning iterations and (bottom) the total number of acquired samples required to …
Figure 19
Figure 19. Figure 19: Active learning curves for the build 6k dataset under the distance configuration, with varying budget constraints and batch sizes (2, 5, and 10). Model accuracy on the test set is shown over active learning iterations, with the mean over 5 seeds represented by a solid…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [28]

    Batch multi-fidelity active learning with budget constraints

    Shibo Li, Jeff M Phillips, Xin Yu, Robert Kirby, and Shandian Zhe. Batch multi-fidelity active learning with budget constraints. Advances in Neural Information Processing Systems , 35: 995–1007, 2022

    work page 2022

  2. [27]

    Optimised probabilistic active learning (OPAL) for fast, non-myopic, cost-sensitive active classification

    Georg Krempl, Daniel Kottke, and Vincent Lemaire. Optimised probabilistic active learning (OPAL) for fast, non-myopic, cost-sensitive active classification. Machine Learning, 100:449– 476, 2015

    work page 2015

  3. [25]

    Active learning for cost-sensitive classification

    Akshay Krishnamurthy, Alekh Agarwal, Tzu-Kuo Huang, Hal Daum´ e III, and John Langford. Active learning for cost-sensitive classification. Journal of Machine Learning Research, 20(65): 1–50, 2019

    work page 2019

  4. [1]

    A survey on active learning and human-in-the-loop deep learning for medical image analysis.Medical image analysis, 71:102062, 2021

    Samuel Budd, Emma C Robinson, and Bernhard Kainz. A survey on active learning and human-in-the-loop deep learning for medical image analysis.Medical image analysis, 71:102062, 2021

    work page 2021

  5. [2]

    Active learning for improved semi-supervised semantic segmentation in satellite images

    Shasvat Desai and Debasmita Ghose. Active learning for improved semi-supervised semantic segmentation in satellite images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 553–563, 2022

    work page 2022

  6. [3]

    Active learning for structural reliability: Survey, general framework and benchmark

    Maliki Moustapha, Stefano Marelli, and Bruno Sudret. Active learning for structural reliability: Survey, general framework and benchmark. Structural Safety, 96:102174, 2022

    work page 2022

  7. [4]

    Deep Bayesian active learning with image data

    Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. In International Conference on Machine Learning , pages 1183–1192. PMLR, 2017

    work page 2017

  8. [5]

    Bayesian neural networks and density networks

    David JC MacKay. Bayesian neural networks and density networks. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 354(1):73–80, 1995

    work page 1995

Show all 43 references
  1. [6]

    Bayesian learning for neural networks , volume 118

    Radford M Neal. Bayesian learning for neural networks , volume 118. Springer Science & Business Media, 2012

    2012

  2. [7]

    Deep Bayesian active learning-to-rank with relative annotation for estimation of ulcerative colitis severity

    Takeaki Kadota, Hideaki Hayashi, Ryoma Bise, Kiyohito Tanaka, and Seiichi Uchida. Deep Bayesian active learning-to-rank with relative annotation for estimation of ulcerative colitis severity. Medical Image Analysis, 97:103262, 2024

    2024

  3. [8]

    Active learning with convolutional neural networks for hyperspectral image classification using a new Bayesian approach

    Juan Mario Haut, Mercedes E Paoletti, Javier Plaza, Jun Li, and Antonio Plaza. Active learning with convolutional neural networks for hyperspectral image classification using a new Bayesian approach. IEEE Transactions on Geoscience and Remote Sensing, 56(11):6440–6461, 2018

    2018

  4. [9]

    Deep Bayesian active learning for natural language processing: Results of a large-scale empirical study

    Aditya Siddhant and Zachary C Lipton. Deep Bayesian active learning for natural language processing: Results of a large-scale empirical study. arXiv preprint arXiv:1808.05697 , 2018

    2018 arXiv

  5. [10]

    Challenges in Markov chain Monte Carlo for Bayesian neural networks

    Theodore Papamarkou, Jacob Hinkle, M Todd Young, and David Womble. Challenges in Markov chain Monte Carlo for Bayesian neural networks. Statistical Science, 37(3):425–442, 2022

    2022

  6. [11]

    A complete recipe for stochastic gradient MCMC

    Yi-An Ma, Tianqi Chen, and Emily Fox. A complete recipe for stochastic gradient MCMC. Advances in Neural Information Processing Systems , 28, 2015

    2015

  7. [12]

    Bayesian model comparison and backprop nets

    David MacKay. Bayesian model comparison and backprop nets. Advances in Neural Informa- tion Processing Systems, 4, 1991

    1991

  8. [13]

    Dropout as a Bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning , pages 1050–1059. PMLR, 2016

    2016

  9. [14]

    Deep Bayesian active learning, A brief survey on recent advances

    Salman Mohamadi and Hamidreza Amindavar. Deep Bayesian active learning, A brief survey on recent advances. arXiv preprint arXiv:2012.08044 , 2020. 14

    2012 arXiv

  10. [15]

    A survey on Bayesian deep learning

    Hao Wang and Dit-Yan Yeung. A survey on Bayesian deep learning. ACM Computing Surveys, 53(5):1–37, 2020

    2020

  11. [16]

    A review of uncertainty quantification in deep learning: Techniques, applications and challenges

    Moloud Abdar, Farhad Pourpanah, Sadiq Hussain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul Fieguth, Xiaochun Cao, Abbas Khosravi, U Rajendra Acharya, et al. A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information f...

    2021

  12. [17]

    Uncertainty in deep learning

    Yarin Gal. Uncertainty in deep learning . PhD thesis, University of Cambridge, 2016

    2016

  13. [18]

    A mathematical theory of communication

    Claude Elwood Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27(3):379–423, 1948

    1948

  14. [19]

    Elementary applied statistics

    Linton G Freeman. Elementary applied statistics . John Wiley & Sons, 1975

    1975

  15. [20]

    Batchbald: Efficient and diverse batch acquisition for deep Bayesian active learning

    Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep Bayesian active learning. Advances in Neural Information Processing Systems, 32, 2019

    2019

  16. [21]

    Speeding up BatchBALD: A k-BALD family of approximations for active learning

    Andreas Kirsch. Speeding up BatchBALD: A k-BALD family of approximations for active learning. arXiv preprint arXiv:2301.09490 , 2023

    2023 arXiv

  17. [22]

    Deep batch active learning by diverse, uncertain gradient lower bounds

    Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agar- wal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019

    1906 arXiv

  18. [23]

    Bayesian batch active learning as sparse subset approximation

    Robert Pinsler, Jonathan Gordon, Eric Nalisnick, and Jos´ e Miguel Hern´ andez-Lobato. Bayesian batch active learning as sparse subset approximation. Advances in Neural Infor- mation Processing Systems, 32, 2019

    2019

  19. [24]

    Stochastic batch acquisition: A simple baseline for deep active learning

    Andreas Kirsch, Sebastian Farquhar, Parmida Atighehchian, Andrew Jesson, Frederic Branchaud-Charron, and Yarin Gal. Stochastic batch acquisition: A simple baseline for deep active learning. arXiv preprint arXiv:2106.12059 , 2021

    2021 arXiv

  20. [26]

    Learning cost-sensitive active classifiers

    Russell Greiner, Adam J Grove, and Dan Roth. Learning cost-sensitive active classifiers. Artificial Intelligence, 139(2):137–174, 2002

    2002

  21. [29]

    Aid: A benchmark data set for performance evaluation of aerial scene classification

    Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing , 55(7):3965–3981, 2017. 15

    2017

  22. [30]

    xView: Objects in context in overhead imagery

    Darius Lam, Richard Kuzma, Kevin McGee, Samuel Dooley, Michael Laielli, Matthew Klaric, Yaroslav Bulatov, and Brendan McCord. xView: Objects in context in overhead imagery. arXiv preprint arXiv:1802.07856 , 2018

    2018 arXiv

  23. [31]

    Dota: A large-scale dataset for object detection in aerial images

    Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pag...

    2018

  24. [32]

    SPAGRI-AI: Smart precision agriculture dataset of aerial images at different heights for crop and weed detection using super-resolution

    Martin Jonak, Jan Mucha, Stepan Jezek, Daniel Kovac, and Kornel Cziria. SPAGRI-AI: Smart precision agriculture dataset of aerial images at different heights for crop and weed detection using super-resolution. Agricultural Systems, 216:103876, 2024

    2024

  25. [33]

    Estimating building energy efficiency from street view imagery, aerial imagery, and land surface temperature data

    Kevin Mayer, Lukas Haas, Tianyuan Huang, Juan Bernab´ e-Moreno, Ram Rajagopal, and Martin Fischer. Estimating building energy efficiency from street view imagery, aerial imagery, and land surface temperature data. Applied Energy, 333:120542, 2023

    2023

  26. [34]

    Review of 50 years of EU energy efficiency policies for buildings

    Marina Economidou, Valeria Todeschi, Paolo Bertoldi, Delia D’Agostino, Paolo Zangheri, and Luca Castellazzi. Review of 50 years of EU energy efficiency policies for buildings. Energy and buildings, 225:110322, 2020

    2020

  27. [35]

    The relationship be- tween operational energy demand and embodied energy in Dutch residential buildings

    A Koezjakov, D Urge-Vorsatz, W Crijns-Graus, and M Van den Broek. The relationship be- tween operational energy demand and embodied energy in Dutch residential buildings. Energy and Buildings , 165:233–245, 2018

    2018

  28. [36]

    PDOK - Publieke Dienstverlening Op de Kaart

    PDOK. PDOK - Publieke Dienstverlening Op de Kaart. https://www.pdok.nl/, 2024. Ac- cessed: 2024-09-27

    2024

  29. [37]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth´ ee Darcet, Th´ eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khali- dov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 , 2023

    2023 arXiv

  30. [38]

    Understanding architecture age and style through deep learning

    Maoran Sun, Fan Zhang, Fabio Duarte, and Carlo Ratti. Understanding architecture age and style through deep learning. Cities, 128:103787, 2022

    2022

  31. [39]

    Bayesian learning via stochastic gradient Langevin dynamics

    Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11) , pages 681–688, 2011

    2011

  32. [40]

    Stochastic gradient hamiltonian monte carlo

    Tianqi Chen, Emily Fox, and Carlos Guestrin. Stochastic gradient hamiltonian monte carlo. In International Conference on Machine Learning , pages 1683–1691, 2014

    2014

  33. [41]

    Cycli- cal stochastic gradient MCMC for Bayesian deep learning

    Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cycli- cal stochastic gradient MCMC for Bayesian deep learning. arXiv preprint arXiv:1902.03932 , 2019

    1902 arXiv

  34. [42]

    3D BAG viewer

    3D BAG. 3D BAG viewer. https://3dbag.nl/en/viewer, 2024. Accessed: 2024-09-27

    2024

  35. [43]

    R VO - Rijksdienst voor Ondernemend Nederland

    R VO. R VO - Rijksdienst voor Ondernemend Nederland. https://www.rvo.nl/, 2024. Ac- cessed: 2024-09-27. 16 A Building datasets curation process We leverage open data sources and web services available in the Netherlands to create the released building datasets. Specifically, w...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.