Pith. sign in

REVIEW 5 major objections 5 minor 80 references

Semantic-Aware Gaussian Process Calibration with Structured Layerwise Kernels for Deep Neural Networks

T0 review · 5 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 The paper claims that calibrating deep classifier confidence with a layered Gaussian process—one local GP per network layer joined by an additive kernel made of a shared global term and a layer-local term—produces more stable and accurate…

desk verdict Promising layerwise GP calibration idea, but the 'consistently superior' claim is undercut by the paper's own ResNet-18 results and the absence of uncertainty estimates. read the letter →

arxiv 2507.15987 v1 pith:5BCXFQY6 submitted 2025-07-21 cs.LG cs.CV

Kyung-hwan Lee , Kyung-tae Kim This is my paper · ORCID
classification cs.LGcs.CV
keywords confidencecalibrationGaussianprocesslayerwisestructuredkernelintrinsiccoregionalizationmodeldomainshiftsoftmaxresidualuncertaintyquantification
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 tries to establish that confidence calibration for deep classifiers gets better and more stable when the calibration model mirrors the network's own layer structure. The proposed framework, SAL-GP, replaces a single Gaussian process correction with one local GP per network layer, all coupled through an additive structured kernel made of a shared global component and a layer-specific component. The authors report that this multi-layer variant, SAL-GP(ML), lowers expected calibration error more than temperature scaling or any single-layer GP on miscalibrated, shifted test sets, and that its performance barely depends on which layer or pooling method is used. Calibrated confidence is what lets operators trust a model's probability estimates in high-stakes settings, and the usual baselines either give no uncertainty estimate or require an ad hoc choice of which layer to calibrate on.

What carries the argument

The load-bearing object is the additive multi-layer kernel $k((x,\ell),(x',\ell')) = k_{\mathrm{global}}(x,x') + \delta_{\ell\ell'} k_{\mathrm{layer}}(x,x')$, a computationally cheap approximation of the intrinsic coregionalization model in which the coregionalization matrix is replaced by a rank-one global term plus a diagonal layer term. It is paired with the input tuple $(f_\ell, s, \ell)$ and the softmax-residual target $r = c - s$. Because the Kronecker delta kills cross-layer covariance, the layerwise GPs are coupled only through the shared global kernel; the global prediction is defined by evaluating the kernel at the artificial test layer index $\ell^* = -1$, so that $k_{\mathrm{global}}$ alone connects the test point to the training set. This kernel is what lets the model pool information from all layers while keeping inference cheap and layer-interpretable.

What would settle it

Fit the full multi-task kernel $k_{\mathrm{feat}}(x,x')\,B_{\ell\ell'}$ with a learned non-diagonal coregionalization matrix $B$ on the same AConvNet/MSTAR and RNN/PLAsTiCC protocols and compare ECE and predictive variance with SAL-GP(ML); if the non-diagonal model does not beat the diagonal additive approximation, the paper's mechanism is sufficient, and if it does, the additive kernel has discarded the inter-layer dependencies the method claims to use. As a second check, recompute the SAL-GP(ML) global predictive variance with the $k_{\mathrm{layer}}$ self-term removed: if the reported variance barely changes, then the layer-local kernel is not contributing to the global uncertainty the paper presents as joint marginalization.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that an additive, layer-indexed kernel over a GP that regresses softmax residuals can pool calibration signal from every internal layer of a network. Each sample enters as $(f_\ell, s, \ell)$, where $f_\ell$ is the pooled feature vector of layer $\ell$, $s$ is the maximum softmax score, and the regression target is the residual $r = c - s$ between correctness and confidence. The kernel $k((x,\ell),(x',\ell')) = k_{\mathrm{global}}(x,x') + \delta_{\ell\ell'} k_{\mathrm{layer}}(x,x')$ is presented as an approximation of the intrinsic coregionalization model in which cross-layer covariances are set to zero, so layers interact only through the shared global term; the global prediction is made by evaluating the kernel at the unseen test layer $\ell^* = -1$. Empirically the paper finds that this configuration outperforms temperature scaling and all single-layer GPs in the miscalibrated regimes, for instance reducing ECE from 0.18257 to 0.08532 on the AConvNet/MSTAR setup and from 0.26768 to 0.15402 on the RNN/PLAsTiCC setup, while the hierarchical-layer variant SAL-GP(HL) is inconsistent and sometimes worse than simple baselines.

Load-bearing premise

The whole result rests on the assumption that ignoring direct cross-layer statistical dependencies—keeping only a shared global kernel plus layer-local corrections—still captures enough of how uncertainty propagates through a deep network to improve calibration under domain shift.

Editorial extensions

If this is right

  • If the central claim holds, practitioners can post-hoc calibrate a deep classifier without first deciding which layer's features are best, because the multi-layer kernel's global prediction is stable across layers and pooling choices.
  • In severely miscalibrated and domain-shifted settings, structured layerwise GP calibration can cut expected calibration error to roughly half the uncalibrated or temperature-scaled level, while also supplying a per-prediction variance that neither baseline provides.
  • The hierarchical-layer variant SAL-GP(HL) should not be used as a drop-in replacement: the paper's experiments show it underperforming single-layer GP in several configurations, indicating that the additive multi-layer kernel rather than the hierarchy is the active ingredient.
  • For already well-calibrated networks the method's gains are small, so its practical value is concentrated in high-overconfidence, shifted, or under-represented-class regimes.

Reading between the lines

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

  • Because the additive kernel zeroes out every off-diagonal cross-layer covariance, the improvement over single-layer GP is most plausibly explained by pooling training examples across layers through the shared global kernel rather than by modeling propagation of uncertainty between layers; a flat GP trained on concatenated layer features should reproduce much of the effect.
  • The global prediction at $\ell^* = -1$ uses $k_{\mathrm{layer}}$ only in the prior self-variance, which the paper then subtracts when reporting pure global variance; a direct test would be to compare predictive variance with and without that term, since joint marginalization across layers would imply the layer kernel changes the global posterior rather than just the bookkeeping.
  • A testable extension is to increase network depth well beyond five layers and check whether the benefit scales; the paper's interpretability story treats each layer as a reasoning stage, but its experiments only report aggregate calibration metrics, not evidence that specific layers contribute specific corrections.
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

5 major / 5 minor

Summary. The paper proposes SAL-GP, a Gaussian-process post-hoc calibration framework that pools feature representations from all layers of a deep neural network and models the calibration residual with an additive kernel combining a global (layer-agnostic) kernel and a Kronecker-delta layerwise kernel. Two variants are studied: SAL-GP(HL), a hierarchical layer kernel, and SAL-GP(ML), a reduced ICM-style multi-layer kernel. The manuscript claims that SAL-GP(ML) provides consistent calibration improvement over single-layer GP and temperature scaling, with the global prediction at a test layer index l*=-1 serving as the primary output. Experiments are reported on ConvNet, AConvNet, ResNet-18 (MSTAR) and an RNN (PLAsTiCC), with ECE, MCE, NLL, Brier score, and predictive variance as metrics.

Significance. If the central claim were established, SAL-GP(ML) would be a practically useful calibration method, especially under domain shift, because it offers a tractable way to use all intermediate layers without the intractability of deep GPs. The paper has strengths: it evaluates across several architectures and data modalities, provides ablations over pooling strategies and layer subsets, and the additive kernel in Eq. (62) is a reasonable, computationally efficient approximation of the ICM kernel. However, the empirical evidence is mixed, no uncertainty estimates are reported for the calibration metrics, and the theoretical derivation of the global predictive variance and the 'joint marginalization' claim contain gaps. As presented, the paper's headline claim of consistent superiority is not supported.

major comments (5)
  1. [V-A3, Table III] The paper's claim that SAL-GP(ML) 'consistently enhances calibration stability' (Section VI) is directly contradicted by Table III: on ResNet-18, the best single-layer GP (Layer 4, max pooling) has ECE=0.01972, while SAL-GP(ML) G has ECE=0.02983 (max) and 0.03243 (average), a relative degradation of 51-64%. NLL and Brier also worsen. Table I (max pooling) and Table II (max pooling) show similar contradictions. The claim of consistency is therefore not supported by the reported results.
  2. [III-F, Eqs. (53)-(57)] Equation (57) defines the 'pure global predictive variance' by subtracting k_layer(z*,z*) from the actual predictive variance at the test layer index l*=-1. For the additive kernel of Eq. (51), the predictive variance at l*=-1 is k_global(z*,z*)+k_layer(z*,z*) - k_global^T(K+sigma^2 I)^{-1} k_global, not the expression in Eq. (57) with k_layer(z*,z*) removed. Thus Eq. (57) is not the variance of any Gaussian predictive distribution for the model, and the reported uncertainty estimates for the global prediction are not correct posterior variances. The derivation needs to be either corrected or explicitly labeled as a heuristic approximation with a stated justification.
  3. [V-A1, Eq. (74)] Equation (74) asserts a global posterior p(f_global | X,y) obtained by integrating over the layerwise latent functions, but no derivation is given and this formula is not used to derive the predictive equations in Section III-F or III-G. Moreover, the additive kernel in Eq. (62) sets cross-layer covariances to zero (except through the shared global kernel), so the model does not realize the 'joint marginalization across all layers' claimed in the abstract and in Section III-G. Either a derivation connecting Eq. (74) to the kernel in Eq. (62) must be provided, or the joint-marginalization claim should be removed.
  4. [Tables I-IV, general empirical protocol] No error bars, confidence intervals, or repeated-seed statistics are reported anywhere in the paper. The improvements over the best single-layer GP in Table I (ECE 0.00832 to 0.00797) and Table II (0.08800 to 0.08532) are on the order of a few percent and could easily be within the sampling noise of a single split. Without such statistics, the claim of 'consistent' improvement is not statistically supported. The paper should either provide multiple runs or clearly label the improvements as not statistically differentiated.
  5. [V-B, Table IV] In Table IV, the global prediction SAL-GP(ML) G(1-5) is reported with a predictive variance of 0.00001, which is orders of magnitude smaller than the layerwise SAL-GP(ML) variances (about 0.12) and implausibly small for a GP trained on data with duplicate layer copies. This suggests a different computation for the global prediction that is not explained, and it undermines the paper's uncertainty-quantification claims. The MCE of 0.98859 further indicates that worst-case calibration is nearly uncalibrated despite the lower ECE, which should be discussed explicitly.
minor comments (5)
  1. [III-F, III-G, IV] Typographical errors include 'kenrel' (Eq. 52 area), 'Hierachical' (Section II-D heading), 'Databse' (Section V-A heading), and 'SAL-GP (ML)) L4' in Table III. The paper should be carefully proofread.
  2. [II-D, V-A] The text refers to 'Appendix 2' and 'Section S1' for derivations and further analysis, but no appendix or supplementary document is included in the manuscript. These references should be either included or removed.
  3. [V-A2] The text states 'Figs. 6 and 7' when discussing AConvNet reliability diagrams, but Figs. 6-8 are PLAsTiCC light curves and distributions. The correct references are Figs. 11 and 12.
  4. [IV-A1] The 'Standard CNN' described as Conv(16)->MaxPool->... does not match reference [76] (LeNet-5). Either the architecture description or the reference citation should be corrected.
  5. [V-A1, Eq. (72)] The temperature-scaling loss in Eq. (72) is written in a way that mixes sample-level and batch-level notation; it would be clearer to define L(T) as the sum of per-sample cross-entropy losses. The gradient analysis in Eq. (73) is correct but should state the assumption that the validation set is perfectly classified.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: GP hyperparameters are fitted by marginal likelihood on held-out residuals, the ℓ*=-1 global prediction is a modeling choice rather than a fitted target, and the only self-citation is a non-load-bearing pointer to prior OOD work.

full rationale

The derivation chain is self-contained. The GP is trained on the residual target r_i = c_i - s_i via marginal likelihood, and the calibrated output s'_* = s_* + r_* is standard residual regression; the correctness labels enter as training targets, while all ECE/MCE/NLL numbers are computed on held-out test points, so no reported calibration metric is a renamed fitted parameter. The global prediction at the artificial test-layer index ℓ* = -1 (Eqs. 53-56) is an explicit modeling choice: setting the Kronecker delta to zero gives the ordinary GP predictive mean under the global kernel, which does not make the ECE result true by definition. The additive kernel of Eq. 62 is obtained as a valid algebraic special case of ICM/LCM (diagonal or rank-one-plus-diagonal coregionalization matrix), not by circular appeal. The only self-citation ([79]) appears in a future-work/OOD-filtering sentence and is not load-bearing. Concerns that Eq. 62 uses the same notation as Eq. 51, or that Table III contradicts the 'consistently superior' claim, are internal-consistency and empirical-correctness issues rather than circular reductions, so they do not raise the circularity score.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard GP formulas and a set of untested domain assumptions about layerwise feature semantics; the global prediction at l*=-1 and the additive kernel are specific modeling choices, and kernel hyperparameters are fitted but not reported.

free parameters (4)
  • GP observation noise variance beta = not reported
    Optimized by marginal likelihood on calibration residuals; not reported in text.
  • Global kernel scale and lengthscale = not reported
    Hyperparameters of the Matern (nu=2.5) global kernel, optimized by marginal likelihood; initial values and bounds not specified.
  • Layer kernel scale and lengthscale = not reported
    Hyperparameters of the layerwise kernel, optimized by marginal likelihood; values not provided.
  • Matern smoothness nu = 2.5
    Chosen by hand in Section IV-G for all GPs; not tuned per dataset.
assumptions (6)
  • standard math Standard GP regression formulas (posterior mean and variance, marginal likelihood) are valid.
    Section III-A relies on textbook GP machinery.
  • domain assumption The ICM/LCM kernel framework and its additive approximation are applicable to layerwise calibration.
    Section III-G imports ICM/LCM from multi-task GP literature to justify the additive kernel; no derivation that the approximation preserves calibration coherence.
  • domain assumption Each neural network layer's pooled feature vector, together with max softmax score and layer index, is a sufficient input representation for calibration residuals.
    Section III-F and IV-E define inputs [f_l, s, l] and assume they capture layer semantics.
  • domain assumption The residual r = c - s is a valid calibration target.
    Taken from Qiu et al. [29], [30]; the paper does not test whether this target is well-specified under domain shift.
  • ad hoc to paper Using a test layer index l*=-1 that is absent from training yields a meaningful 'global' prediction.
    Section III-F Eq. (53) defines global prediction by zeroing the delta term; this is a design choice specific to this paper, not a consequence of the model.
  • ad hoc to paper The additive kernel's zero cross-layer covariance still captures the claimed joint marginalization across layers.
    Eqs. (62)-(63) discard off-diagonal B terms, so layers are only coupled through the shared global kernel; the paper asserts joint behavior without proving it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic-Aware Gaussian Process Calibration with Structured Layerwise Kernels for Deep Neural Networks." pith.science (2026). https://pith.science/paper/5BCXFQY6

@misc{pith2026250715987,
  author       = {Pith},
  title        = {Pith review of: Semantic-Aware Gaussian Process Calibration with Structured Layerwise Kernels for Deep Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5BCXFQY6}},
  note         = {Machine review of arXiv:2507.15987}
}
read the original abstract

Calibrating the confidence of neural network classifiers is essential for quantifying the reliability of their predictions during inference. However, conventional Gaussian Process (GP) calibration methods often fail to capture the internal hierarchical structure of deep neural networks, limiting both interpretability and effectiveness for assessing predictive reliability. We propose a Semantic-Aware Layer-wise Gaussian Process (SAL-GP) framework that mirrors the layered architecture of the target neural network. Instead of applying a single global GP correction, SAL-GP employs a multi-layer GP model, where each layer's feature representation is mapped to a local calibration correction. These layerwise GPs are coupled through a structured multi-layer kernel, enabling joint marginalization across all layers. This design allows SAL-GP to capture both local semantic dependencies and global calibration coherence, while consistently propagating predictive uncertainty through the network. The resulting framework enhances interpretability aligned with the network architecture and enables principled evaluation of confidence consistency and uncertainty quantification in deep models.

Figures

Figures reproduced from arXiv: 2507.15987 by the authors.

Figure 1
Figure 1. Comparative performance of various calibration methods on MSTAR DB. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Effect of varying BBQ prior hyperparameters on calibration performance for a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Four possible interpretations for constructing a Hierarchical Gaussian Process based on class and neural network layer. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Layer-wise Semantic Representation Architecture of SAL-GP [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: MSTAR Database: Optical images of military targets versus SAR images [1]. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Multiband photometric light curves for three representative PLAsTiCC objects [77], [78] (shown in columns). Each [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Distributions of photometric flux measurements for [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 9
Figure 9. Figure 9: Calibration residual fit plots (top row) and reliability diagrams (bottom row) for each calibration approach on the [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Calibration residual fit plots (top row) and reliability diagrams (bottom row) for each calibration approach on the [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 15
Figure 15. Figure 15: Test phase confusion matrix for PLAsTiCC data [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]
Figure 11
Figure 11. Figure 11: Calibration residual fit plots (top row) and reliability diagrams (bottom row) for each calibration approach on the all [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Calibration residual fit plots (top row) and reliability diagrams (bottom row) for each calibration approach on the all [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Calibration residual fit plots (top row) and reliability diagrams (bottom row) for each calibration approach on the [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]
Figure 14
Figure 14. Figure 14: Calibration residual fit plots (top row) and reliability diagrams (bottom row) for each calibration approach on the [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 16
Figure 16. Figure 16: Calibration residual fit plots (top row) and reliability diagrams (bottom row) for each calibration approach on the [PITH_FULL_IMAGE:figures/full_fig_p026_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

80 extracted references · 69 canonical work pages

  1. [2]

    ”High- way networks.” arXiv preprint arXiv:1505.00387 (2015)

    Srivastava, Rupesh Kumar, Klaus Greff, and J ¨urgen Schmidhuber. ”High- way networks.” arXiv preprint arXiv:1505.00387 (2015)

    arXiv 2015

  2. [3]

    ”Deep residual learning for image recognition.” Proceedings of the IEEE conference on computer vision and pattern recognition

    He, Kaiming, et al. ”Deep residual learning for image recognition.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2016

    2016

  3. [4]

    ”Deep networks with stochastic depth.” Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Nether- lands, October 11–14, 2016, Proceedings, Part IV 14

    Huang, Gao, et al. ”Deep networks with stochastic depth.” Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Nether- lands, October 11–14, 2016, Proceedings, Part IV 14. Springer Interna- tional Publishing, 2016

    work page 2016

  4. [5]

    ”Densely connected convolutional networks.” Proceed- ings of the IEEE conference on computer vision and pattern recognition

    Huang, Gao, et al. ”Densely connected convolutional networks.” Proceed- ings of the IEEE conference on computer vision and pattern recognition. 2017

    work page 2017

  5. [6]

    ”On the use of artificial neural networks in simulation-based manufacturing control.” Journal of Simulation 8.1 (2014): 76-90

    Bergmann, S ¨oren, S ¨oren Stelzer, and Steffen Strassburger. ”On the use of artificial neural networks in simulation-based manufacturing control.” Journal of Simulation 8.1 (2014): 76-90

    work page 2014

  6. [7]

    ”Deep learning.” nature 521.7553 (2015): 436-444

    LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. ”Deep learning.” nature 521.7553 (2015): 436-444

    work page 2015

  7. [8]

    ”Searching for exotic particles in high-energy physics with deep learning.” Nature communica- tions 5.1 (2014): 4308

    Baldi, Pierre, Peter Sadowski, and Daniel Whiteson. ”Searching for exotic particles in high-energy physics with deep learning.” Nature communica- tions 5.1 (2014): 4308

    work page 2014

  8. [9]

    ”Forecasting S&P 500 index using artificial neural networks and design of experiments.” Journal of Industrial Engineering International 9 (2013): 1-9

    Niaki, Seyed Taghi Akhavan, and Saeid Hoseinzade. ”Forecasting S&P 500 index using artificial neural networks and design of experiments.” Journal of Industrial Engineering International 9 (2013): 1-9

    work page 2013

Show all 80 references
  1. [10]

    ”Neural networks applied to discriminate botanical origin of honeys.” Food chemistry 175 (2015): 128-136

    Anjos, Of ´elia, et al. ”Neural networks applied to discriminate botanical origin of honeys.” Food chemistry 175 (2015): 128-136. JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, MARCH 2025 26 Uncalibrated Temp. Scaled Single GP (Best Layer) SAL-GP (ML) SAL-GP (HL) Fig. 16: Calib...

    2015

  2. [11]

    ”Applications of arti- ficial neural networks in health care organizational decision-making: A scoping review.” PloS one 14.2 (2019): e0212356

    Shahid, Nida, Tim Rappon, and Whitney Berta. ”Applications of arti- ficial neural networks in health care organizational decision-making: A scoping review.” PloS one 14.2 (2019): e0212356

    2019

  3. [12]

    Fienberg

    DeGroot, Morris H., and Stephen E. Fienberg. ”The comparison and evaluation of forecasters.” Journal of the Royal Statistical Society: Series D (The Statistician) 32.1-2 (1983): 12-22

    1983

  4. [13]

    ”Predicting good prob- abilities with supervised learning.” Proceedings of the 22nd international conference on Machine learning

    Niculescu-Mizil, Alexandru, and Rich Caruana. ”Predicting good prob- abilities with supervised learning.” Proceedings of the 22nd international conference on Machine learning. 2005

    2005

  5. [14]

    ”On calibration of modern neural networks.” Inter- national conference on machine learning

    Guo, Chuan, et al. ”On calibration of modern neural networks.” Inter- national conference on machine learning. PMLR, 2017

    2017

  6. [15]

    ”End to end learning for self-driving cars.” arXiv preprint arXiv:1604.07316 (2016)

    Bojarski, Mariusz, et al. ”End to end learning for self-driving cars.” arXiv preprint arXiv:1604.07316 (2016)

    2016 arXiv

  7. [16]

    ”Calibrating predictive model estimates to support personalized medicine.” Journal of the American Medical Informatics Association 19.2 (2012): 263-274

    Jiang, Xiaoqian, et al. ”Calibrating predictive model estimates to support personalized medicine.” Journal of the American Medical Informatics Association 19.2 (2012): 263-274

    2012

  8. [17]

    ”Predicting with confidence and tolerance.” Nature methods 15.11 (2018): 843-845

    Altman, Naomi, and Martin Krzywinski. ”Predicting with confidence and tolerance.” Nature methods 15.11 (2018): 843-845

    2018

  9. [18]

    ”Errors in predictor variables.” (2024): 4-6

    Altman, Naomi, and Martin Krzywinski. ”Errors in predictor variables.” (2024): 4-6

    2024

  10. [19]

    ”Probabilistic machine learning and artificial in- telligence.” Nature 521.7553 (2015): 452-459

    Ghahramani, Zoubin. ”Probabilistic machine learning and artificial in- telligence.” Nature 521.7553 (2015): 452-459

    2015

  11. [20]

    ”Simple and scalable predictive uncertainty estimation using deep en- sembles.” Advances in neural information processing systems 30 (2017)

    Lakshminarayanan, Balaji, Alexander Pritzel, and Charles Blundell. ”Simple and scalable predictive uncertainty estimation using deep en- sembles.” Advances in neural information processing systems 30 (2017)

    2017

  12. [21]

    ”Dropout as a bayesian approxi- mation: Representing model uncertainty in deep learning.” international conference on machine learning

    Gal, Yarin, and Zoubin Ghahramani. ”Dropout as a bayesian approxi- mation: Representing model uncertainty in deep learning.” international conference on machine learning. PMLR, 2016

    2016

  13. [22]

    ”Neural processes.” arXiv preprint arXiv:1807.01622 (2018)

    Garnelo, Marta, et al. ”Neural processes.” arXiv preprint arXiv:1807.01622 (2018)

    2018 arXiv

  14. [23]

    ”Classification with Bayesian neural networks.” Ma- chine Learning Challenges Workshop

    Neal, Radford M. ”Classification with Bayesian neural networks.” Ma- chine Learning Challenges Workshop. Berlin, Heidelberg: Springer Berlin Heidelberg, 2005

    2005

  15. [24]

    ”A practical Bayesian framework for backpropaga- tion networks.” Neural computation 4.3 (1992): 448-472

    MacKay, David JC. ”A practical Bayesian framework for backpropaga- tion networks.” Neural computation 4.3 (1992): 448-472

    1992

  16. [25]

    ”Weight uncertainty in neural network.” Inter- national conference on machine learning

    Blundell, Charles, et al. ”Weight uncertainty in neural network.” Inter- national conference on machine learning. PMLR, 2015

    2015

  17. [26]

    ”What uncertainties do we need in bayesian deep learning for computer vision?.” Advances in neural in- formation processing systems 30 (2017)

    Kendall, Alex, and Yarin Gal. ”What uncertainties do we need in bayesian deep learning for computer vision?.” Advances in neural in- formation processing systems 30 (2017)

    2017

  18. [27]

    ”Non- parametric calibration for classification.” International Conference on Artificial Intelligence and Statistics

    Wenger, Jonathan, Hedvig Kjellstr ¨om, and Rudolph Triebel. ”Non- parametric calibration for classification.” International Conference on Artificial Intelligence and Statistics. PMLR, 2020

    2020

  19. [29]

    ”Quantifying point- prediction uncertainty in neural networks via residual estimation with an i/o kernel.” arXiv preprint arXiv:1906.00588 (2019)

    Qiu, Xin, Elliot Meyerson, and Risto Miikkulainen. ”Quantifying point- prediction uncertainty in neural networks via residual estimation with an i/o kernel.” arXiv preprint arXiv:1906.00588 (2019)

    2019 arXiv

  20. [30]

    ”Detecting misclassification errors in neural networks with a gaussian process model.” Proceedings of the AAAI Conference on Artificial Intelligence

    Qiu, Xin, and Risto Miikkulainen. ”Detecting misclassification errors in neural networks with a gaussian process model.” Proceedings of the AAAI Conference on Artificial Intelligence. V ol. 36. No. 7. 2022

    2022

  21. [31]

    ”Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers.” Icml

    Zadrozny, Bianca, and Charles Elkan. ”Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers.” Icml. V ol

  22. [32]

    Zadrozny, Bianca, and Charles Elkan. ”Transforming classifier scores into accurate multiclass probability estimates.” Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. 2002

    2002

  23. [33]

    ”Ob- taining well calibrated probabilities using bayesian binning.” Proceedings of the AAAI conference on artificial intelligence

    Naeini, Mahdi Pakdaman, Gregory Cooper, and Milos Hauskrecht. ”Ob- taining well calibrated probabilities using bayesian binning.” Proceedings of the AAAI conference on artificial intelligence. V ol. 29. No. 1. 2015

    2015

  24. [34]

    ”Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods.” Advances in large margin classifiers 10.3 (1999): 61-74

    Platt, John. ”Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods.” Advances in large margin classifiers 10.3 (1999): 61-74

    1999

  25. [35]

    ”Distilling the knowl- edge in a neural network.” arXiv preprint arXiv:1503.02531 (2015)

    Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean. ”Distilling the knowl- edge in a neural network.” arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  26. [36]

    Information theory and statistical mechanics

    Jaynes, Edwin T. Information theory and statistical mechanics. Physical review, 106(4):620, 1957

    1957

  27. [37]

    ”The elements of statistical learning: Data mining, inference, and prediction.” (2009)

    Friedman, Jerome. ”The elements of statistical learning: Data mining, inference, and prediction.” (2009)

    2009

  28. [38]

    McAuliffe

    Blei, David M., Alp Kucukelbir, and Jon D. McAuliffe. ”Variational inference: A review for statisticians.” Journal of the American statistical Association 112.518 (2017): 859-877

    2017

  29. [39]

    ”Ensemble deep learning: A review.” Engi- neering Applications of Artificial Intelligence 115 (2022): 105151

    Ganaie, Mudasir A., et al. ”Ensemble deep learning: A review.” Engi- neering Applications of Artificial Intelligence 115 (2022): 105151

    2022

  30. [40]

    ”Towards Improved Variational Inference for Deep Bayesian Models.” arXiv preprint arXiv:2401.12418 (2024)

    Ober, Sebastian W. ”Towards Improved Variational Inference for Deep Bayesian Models.” arXiv preprint arXiv:2401.12418 (2024)

    2024 arXiv

  31. [41]

    ”Deep neural networks as gaussian processes.” arXiv preprint arXiv:1711.00165 (2017)

    Lee, Jaehoon, et al. ”Deep neural networks as gaussian processes.” arXiv preprint arXiv:1711.00165 (2017)

    2017 arXiv

  32. [42]

    ”Deep kernel learning.” Artificial intel- ligence and statistics

    Wilson, Andrew Gordon, et al. ”Deep kernel learning.” Artificial intel- ligence and statistics. PMLR, 2016

    2016

  33. [43]

    Improving Output Uncertainty Estimation and Gen eralization in Deep Learning via Neural Network Gaussian Processes

    Tomoharu Iwata and Zoubin Ghahramani. Improving Output Uncertainty Estimation and Gen eralization in Deep Learning via Neural Network Gaussian Processes. arXiv e-prints, art. arXiv:1707.05922, Jul 2017

    2017 arXiv

  34. [44]

    ”Revisiting unreasonable effectiveness of data in deep learning era.” Proceedings of the IEEE international conference on computer vision

    Sun, Chen, et al. ”Revisiting unreasonable effectiveness of data in deep learning era.” Proceedings of the IEEE international conference on computer vision. 2017

    2017

  35. [45]

    MSTAR Overview, moving and stationary target acquisition and recognition MSTAR databse, Sandia national laboratories, https://www.sdms.afrl.af.mil/index.php?collection=mstar

  36. [46]

    ”Deep learning for remote sensing data: A technical tutorial on the state of the art.” IEEE Geoscience and remote sensing magazine 4.2 (2016): 22-40

    Zhang, Liangpei, Lefei Zhang, and Bo Du. ”Deep learning for remote sensing data: A technical tutorial on the state of the art.” IEEE Geoscience and remote sensing magazine 4.2 (2016): 22-40. JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, MARCH 2025 27

    2016

  37. [47]

    ”A comprehensive survey on SAR ATR in deep- learning era.” Remote Sensing 15.5 (2023): 1454

    Li, Jianwei, et al. ”A comprehensive survey on SAR ATR in deep- learning era.” Remote Sensing 15.5 (2023): 1454

    2023

  38. [48]

    ”Change detection in synthetic aperture radar images based on deep neural networks.” IEEE transactions on neural networks and learning systems 27.1 (2015): 125-138

    Gong, Maoguo, et al. ”Change detection in synthetic aperture radar images based on deep neural networks.” IEEE transactions on neural networks and learning systems 27.1 (2015): 125-138

    2015

  39. [49]

    ”Target classification using the deep convolutional networks for SAR images.” IEEE transactions on geoscience and remote sensing 54.8 (2016): 4806-4817

    Chen, Sizhe, et al. ”Target classification using the deep convolutional networks for SAR images.” IEEE transactions on geoscience and remote sensing 54.8 (2016): 4806-4817

    2016

  40. [50]

    ”Very deep convolutional net- works for large-scale image recognition.” arXiv preprint arXiv:1409.1556 (2014)

    Simonyan, Karen, and Andrew Zisserman. ”Very deep convolutional net- works for large-scale image recognition.” arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  41. [51]

    ”Rethinking the inception architecture for computer vision.” Proceedings of the IEEE conference on computer vision and pattern recognition

    Szegedy, Christian, et al. ”Rethinking the inception architecture for computer vision.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2016

    2016

  42. [52]

    M ¨uller, Rafael, Simon Kornblith, and Geoffrey E. Hinton. ”When does label smoothing help?.” Advances in neural information processing systems 32 (2019)

    2019

  43. [53]

    Lawrence

    Alvarez, Mauricio A., Lorenzo Rosasco, and Neil D. Lawrence. ”Kernels for vector-valued functions: A review.” Foundations and Trends® in Machine Learning 4.3 (2012): 195-266

    2012

  44. [54]

    ”Multi-task Gaussian process prediction.” Advances in neural information processing systems 20 (2007)

    Bonilla, Edwin V ., Kian Chai, and Christopher Williams. ”Multi-task Gaussian process prediction.” Advances in neural information processing systems 20 (2007)

    2007

  45. [55]

    Lawrence

    Alvarez, Mauricio A., and Neil D. Lawrence. ”Computationally efficient convolved multiple output Gaussian processes.” The Journal of Machine Learning Research 12 (2011): 1459-1500

    2011

  46. [56]

    ”It is all in the noise: Efficient multi-task Gaussian process inference with structured residuals.” Advances in neural information processing systems 26 (2013)

    Rakitsch, Barbara, et al. ”It is all in the noise: Efficient multi-task Gaussian process inference with structured residuals.” Advances in neural information processing systems 26 (2013)

    2013

  47. [57]

    ”Simple and principled uncertainty estimation with deterministic deep learning via distance awareness.” Advances in neural information processing systems 33 (2020): 7498-7512

    Liu, Jeremiah, et al. ”Simple and principled uncertainty estimation with deterministic deep learning via distance awareness.” Advances in neural information processing systems 33 (2020): 7498-7512

    2020

  48. [58]

    ”Gaussian processes for regression.” Advances in neural information processing systems 8 (1995)

    Williams, Christopher, and Carl Rasmussen. ”Gaussian processes for regression.” Advances in neural information processing systems 8 (1995)

    1995

  49. [59]

    Gaussian pro- cesses for machine learning

    Williams, Christopher KI, and Carl Edward Rasmussen. Gaussian pro- cesses for machine learning. V ol. 2. No. 3. Cambridge, MA: MIT press, 2006

    2006

  50. [60]

    Lawrence, Neil D., and Andrew J. Moore. ”Hierarchical Gaussian process latent variable models.” Proceedings of the 24th international conference on Machine learning. 2007

    2007

  51. [61]

    Lawrence, and Magnus Rattray

    Hensman, James, Neil D. Lawrence, and Magnus Rattray. ”Hierarchical Bayesian modelling of gene expression time series across irregularly sampled replicates and clusters.” BMC bioinformatics 14 (2013): 1-12

    2013

  52. [62]

    Lawrence

    Titsias, Michalis, and Neil D. Lawrence. ”Bayesian Gaussian process latent variable model.” Proceedings of the thirteenth international con- ference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 2010

    2010

  53. [63]

    Lawrence

    Damianou, Andreas, and Neil D. Lawrence. ”Deep gaussian processes.” Artificial intelligence and statistics. PMLR, 2013

    2013

  54. [64]

    ”Doubly stochastic variational inference for deep Gaussian processes.” Advances in neural information processing systems 30 (2017)

    Salimbeni, Hugh, and Marc Deisenroth. ”Doubly stochastic variational inference for deep Gaussian processes.” Advances in neural information processing systems 30 (2017)

    2017

  55. [65]

    ”Monte Carlo implementation of Gaussian pro- cess models for Bayesian regression and classification.” arXiv preprint physics/9701026 (1997)

    Neal, Radford M. ”Monte Carlo implementation of Gaussian pro- cess models for Bayesian regression and classification.” arXiv preprint physics/9701026 (1997)

    1997 arXiv

  56. [66]

    Deep learning

    Goodfellow, Ian, et al. Deep learning. V ol. 1. No. 2. Cambridge: MIT press, 2016

    2016

  57. [67]

    ”Risk versus uncertainty in deep learning: Bayes, bootstrap and the dangers of dropout.” NIPS workshop on bayesian deep learning

    Osband, Ian. ”Risk versus uncertainty in deep learning: Bayes, bootstrap and the dangers of dropout.” NIPS workshop on bayesian deep learning. V ol. 192. MIT Press, 2016

    2016

  58. [68]

    ”Bayesian deep learning and a probabilistic perspective of generalization.” Advances in neural information processing systems 33 (2020): 4697-4708

    Wilson, Andrew G., and Pavel Izmailov. ”Bayesian deep learning and a probabilistic perspective of generalization.” Advances in neural information processing systems 33 (2020): 4697-4708

    2020

  59. [69]

    ”Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.” Advances in neural information processing systems 32 (2019)

    Ovadia, Yaniv, et al. ”Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.” Advances in neural information processing systems 32 (2019)

    2019

  60. [70]

    ”Benchmarking neural network robustness to common corruptions and perturbations.” arXiv preprint arXiv:1903.12261 (2019)

    Hendrycks, Dan, and Thomas Dietterich. ”Benchmarking neural network robustness to common corruptions and perturbations.” arXiv preprint arXiv:1903.12261 (2019)

    2019 arXiv

  61. [71]

    ”Accurate uncertainties for deep learning using calibrated regression.” International conference on machine learning

    Kuleshov, V olodymyr, Nathan Fenner, and Stefano Ermon. ”Accurate uncertainties for deep learning using calibrated regression.” International conference on machine learning. PMLR, 2018

    2018

  62. [72]

    ”A simple baseline for bayesian uncertainty in deep learning.” Advances in neural information processing systems 32 (2019)

    Maddox, Wesley J., et al. ”A simple baseline for bayesian uncertainty in deep learning.” Advances in neural information processing systems 32 (2019)

    2019

  63. [73]

    ”Finding structure in time.” Cognitive science 14.2 (1990): 179-211

    Elman, Jeffrey L. ”Finding structure in time.” Cognitive science 14.2 (1990): 179-211

    1990

  64. [74]

    ”Long short-term memory.” Neural computation 9.8 (1997): 1735-1780

    Hochreiter, Sepp, and J ¨urgen Schmidhuber. ”Long short-term memory.” Neural computation 9.8 (1997): 1735-1780

    1997

  65. [75]

    ”Learning phrase representations using RNN encoder-decoder for statistical machine translation.” arXiv preprint arXiv:1406.1078 (2014)

    Cho, Kyunghyun, et al. ”Learning phrase representations using RNN encoder-decoder for statistical machine translation.” arXiv preprint arXiv:1406.1078 (2014)

    2014 arXiv

  66. [76]

    ”Gradient-based learning applied to document recognition.” Proceedings of the IEEE 86.11 (2002): 2278-2324

    LeCun, Yann, et al. ”Gradient-based learning applied to document recognition.” Proceedings of the IEEE 86.11 (2002): 2278-2324

    2002

  67. [77]

    ”The photometric lsst astronomical time- series classification challenge (plasticc): Data set.” arXiv preprint arXiv:1810.00001 (2018)

    Allam Jr, Tarek, et al. ”The photometric lsst astronomical time- series classification challenge (plasticc): Data set.” arXiv preprint arXiv:1810.00001 (2018)

    2018 arXiv

  68. [78]

    Kessler, R., et al. ”Models and simulations for the photometric LSST as- tronomical time series classification challenge (PLAsTiCC).” Publications of the Astronomical Society of the Pacific 131.1003 (2019): 094501

    2019

  69. [79]

    ”Adaptive Residual Transfor- mation for Enhanced Feature-Based OOD Detection in SAR Imagery.” arXiv preprint arXiv:2411.00274 (2024)

    Lee, Kyung-hwan, and Kyung-tae Kim. ”Adaptive Residual Transfor- mation for Enhanced Feature-Based OOD Detection in SAR Imagery.” arXiv preprint arXiv:2411.00274 (2024)

    2024 arXiv

  70. [80]

    [On- line]

    ”PLAsTiCC 2018 Simple RNN PyTorch,” Kaggle notebook, 2018. [On- line]. Available: https://www.kaggle.com/code/johnfarrell/plasticc-2018- simple-rnn-pytorch/notebook. [Accessed: July 18, 2025]

    2018

  71. [81]

    ”Trainable calibration measures for neural networks from kernel mean embeddings.” Interna- tional Conference on Machine Learning

    Kumar, Aviral, Sunita Sarawagi, and Ujjwal Jain. ”Trainable calibration measures for neural networks from kernel mean embeddings.” Interna- tional Conference on Machine Learning. PMLR, 2018

    2018

  72. [82]

    ”Beyond temperature scaling: Obtaining well- calibrated multi-class probabilities with dirichlet calibration.” Advances in neural information processing systems 32 (2019)

    Kull, Meelis, et al. ”Beyond temperature scaling: Obtaining well- calibrated multi-class probabilities with dirichlet calibration.” Advances in neural information processing systems 32 (2019)

    2019

Pith tools

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