Pith. sign in

REVIEW 3 major objections 4 minor 77 references

Enhancing Uncertainty Estimation and Interpretability via Bayesian Non-negative Decision Layer

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A Bayesian non-negative decision layer replaces the softmax layer and improves accuracy, uncertainty, and interpretability.

desk verdict A practically useful Bayesian last-layer with an overstated theoretical guarantee; the empirical results are the real contribution. read the letter →

arxiv 2505.22199 v1 pith:OTPU4DTX submitted 2025-05-28 cs.LG cs.AI

Xinyue Hu , Zhibin Duan , Bo Chen , Mingyuan Zhou This is my paper
classification cs.LGcs.AI
keywords Bayesiannon-negativedecisionlayeruncertaintyestimationmatrixfactorizationdisentangledrepresentationlearningWeibullvariationalinferencesparseinterpretabledeepcategoricallikelihood
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 sets out to fix two longstanding weaknesses of deep classifiers in one move: poor uncertainty estimates and entangled, hard-to-interpret features. It proposes replacing the final softmax layer with a Bayesian Non-negative Decision Layer (BNDL), which treats the last layer as a conditional Bayesian non-negative factor analysis: each image gets a non-negative latent score vector drawn from a gamma distribution, and the class distribution is produced by multiplying that vector by a non-negative factor-loading matrix. Because the latent variables are stochastic, the model can estimate predictive uncertainty in a single forward pass plus lightweight sampling; because the factors are non-negative and sparse, the learned features are supposed to separate into identifiable, semantically meaningful components. The paper reports that BNDL improves or preserves accuracy on CIFAR-10, CIFAR-100, and ImageNet-1k with both ResNet and ViT backbones, produces more reliable uncertainty under the PAvPU metric, and yields more disentangled features, with a partial-identifiability guarantee drawn from non-negative matrix factorization theory.

What carries the argument

The load-bearing object is the factorized generative model of the decision layer, $y_j \mid \theta_j \sim \mathrm{Category}(\theta_j \Phi)$, with gamma priors on the per-sample score $\theta_j$ and global loading $\Phi$, plus the Weibull variational inference network that approximates their posteriors. Substituting the variational means turns the layer into NMF, and the partial identifiability condition of exact NMF (a selective-window row plus $r-1$ zeros per column) is what the authors use to claim disentanglement. The Weibull reparameterization and the analytic KL divergence make stochastic gradient training practical.

What would settle it

Inspect the learned loading matrix $\Phi$ of a trained BNDL on a dataset with $r$ classes: if the number of exactly zero entries in any column is less than $r-1$, or no row is a scalar multiple of a standard basis vector, then the claimed partial-identifiability guarantee is not satisfied for that column. A controlled alternative is to generate synthetic labels from a known sparse non-negative factor model and test whether BNDL's inferred factors recover the generating columns up to permutation and scaling.

Watch

Extended reading notes

Core claim

BNDL's central claim is that a softmax output layer can be replaced by a generative Bayesian non-negative factor model without sacrificing performance, while gaining calibrated uncertainty and interpretability. Formally, the label $y_j$ is drawn from a categorical distribution parameterized by $\theta_j \Phi$, where $\theta_j$ is a sample-specific non-negative factor score with prior $\mathrm{Gamma}(f_\theta(x_j), 1)$ and $\Phi$ is a global non-negative loading matrix with prior $\mathrm{Gamma}(1, 1)$. A Weibull variational network approximates the intractable posteriors, with an analytic gamma-to-Weibull KL divergence, so training is by ordinary SGD on the ELBO. The paper argues that at the expectation of the variational posteriors the layer is exactly a non-negative matrix factorization $Y = \theta \Phi$, and invokes partial identifiability results for NMF to claim the learned factors are unique up to permutation and scaling, which is the formal sense in which the features are disentangled. Experiments across ResNet-18/50 and ViT show accuracy gains (e.g., 95.54 vs 94.98 on CIFAR-10 for ResNet, 77.01 vs 75.33 on ImageNet), higher PAvPU, orders-of-magnitude sparser decision layers, and LIME/GradCAM visualizations that align with object semantics.

Load-bearing premise

The disentanglement guarantee holds only if the trained decision layer satisfies the exact sparsity and selective-window conditions required by the cited NMF identifiability theorem, and the paper checks neither condition directly in the trained models.

Editorial extensions

If this is right

  • BNDL can replace the softmax layer in standard ResNet and ViT classifiers and yield higher or equal accuracy while adding uncertainty estimation on CIFAR-10, CIFAR-100, ImageNet-1k, and Places-10.
  • Uncertainty from BNDL is usable for flagging likely mistakes: accuracy falls monotonically as estimated uncertainty rises across ten ImageNet-1k subsets.
  • The decision layer can be made orders of magnitude sparser with little accuracy loss, which is what supports the interpretability claim.
  • BNDL's learned features score higher on the SEPIN@k disentanglement metric than a standard ResNet, and LIME and GradCAM visualizations concentrate on the object rather than background.
  • Computational overhead is negligible compared with the backbone, and uncertainty estimation needs one forward pass instead of multiple full-network runs.

Reading between the lines

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

  • The paper's own reported ImageNet decision-layer sparsity (1-sparsity about 0.04) is a density measure, not the $r-1$ exact zeros per column the identifiability theorem asks for, so the formal guarantee may be weaker than claimed for the 1000-class model even if the empirical disentanglement holds.
  • If BNDL's uncertainty is reliable across domains, the same layer could serve as a lightweight out-of-distribution detector: sampling from the posterior and testing whether the predictive p-value drops below threshold would be an OOD signal obtainable without retraining ensembles.
  • The approach could transfer to other output structures, such as multi-label or regression settings, wherever the response is a non-negative combination of latent components; the categorical likelihood would need to be replaced, but the gamma and Weibull machinery is likelihood-agnostic.
  • A clean test of the disentanglement mechanism would be to run BNDL on synthetic data generated from a known ground-truth NMF and check whether inferred factors recover the generating columns up to permutation and scaling; the ImageNet experiments are suggestive but do not provide such a controlled check.
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 / 4 minor

Summary. The paper proposes the Bayesian Non-negative Decision Layer (BNDL), which replaces the final softmax layer of a deep network with a non-negative factor-analysis-style generative layer. Labels are modeled under a categorical likelihood with gamma-distributed factor scores θ and factor loadings Φ, and a Weibull variational inference network approximates the intractable posterior. The authors claim that this layer improves or preserves accuracy, provides reliable uncertainty estimates, and yields disentangled and interpretable features, with theoretical guarantees based on partial identifiability results for non-negative matrix factorization (Gillis and Rajkó, 2023). Experiments are reported on CIFAR-10, CIFAR-100, ImageNet-1k, and Places-10 with ResNet and ViT backbones, using PAvPU for uncertainty, SEPIN for disentanglement, and LIME/Grad-CAM visualizations.

Significance. If the central claims held, BNDL would be a simple and scalable last-layer substitution that jointly improves uncertainty estimation, sparsity, and interpretability without requiring ensembles or multiple full-network passes. The empirical core is a genuine strength: the paper reports comparisons on four datasets, uses two uncertainty/interpretability measures (PAvPU and SEPIN), includes LIME and Grad-CAM visualizations, provides a complexity analysis, and releases code. The empirical results are plausible and potentially useful. However, the theoretical disentanglement guarantee, which is a headline contribution in the abstract, Section 4, and the appendix, is not established for the model as actually trained; the revision needs to either prove the claimed equivalence and verify the required assumptions for the real training objective, or substantially weaken those theoretical claims.

major comments (3)
  1. [Sec. 4, Eq. (14)] The reduction of the ELBO in Eq. (13) to the least-squares non-negative matrix factorization objective in Eq. (14) is asserted without proof. The ELBO contains a categorical cross-entropy expected log-likelihood plus two KL divergence terms; squared error is not a special case of categorical cross-entropy, and the point-mass limit discussed in Sec. 3.3 is not the regime used in training and does not remove the softmax normalization. Consequently, Proposition 1/2, even if correct as a statement about exact NMF, applies to a different optimization problem and cannot by itself support the paper's claim that BNDL provides theoretical guarantees for the learned model.
  2. [Sec. 4 and Appendix A.2] The two assumptions of Proposition 2 are never verified for the trained BNDL. The Selective Window condition requires a row of Φ equal to a scaled unit vector, and the Sparsity Constraint requires the k-th column of Φ to have at least r−1 zeros; with r = rank(Y) = C for one-hot classification labels, this means each column of Φ has at most one nonzero entry. The reported aggregate density values (e.g., 1-sparsity about 0.04 on ImageNet, or the 1-Sparsity values in Table 4) do not imply per-column zeros, and no experiment checks either condition. Therefore the statement in Appendix A.2 that 'BNDL satisfies the sparsity constraint' is unsupported, and the disentanglement guarantee is vacuous as stated.
  3. [Sec. 3.3, around Eqs. (8) and (11)] The claim that 'Equations 8 and 11 ensure E[θj] = λj and E[Φ] = λΦ' is not correct for the Weibull parameterization in Eq. (6): the mean of Weibull(k, λ) is λ Γ(1 + 1/k), not λ. With λ defined as ReLU(·)/exp(1 + 1/k), the mean equals ReLU(·) Γ(1 + 1/k)/exp(1 + 1/k), which equals the encoder output only asymptotically as k → ∞. Since the 'Connection with Non-negative Matrix Factorization' paragraph and the subsequent reformulation in Eq. (14) rely on this equality, the NMF interpretation is not exact for finite k and needs correction or qualification.
minor comments (4)
  1. [Throughout] There are several typographical errors, including 'Excat NMF' (Appendix A.2), 'Weilbull' (Section 3.3), 'Reproduciltly' (Reproducibility Statement), and 'the last two term is' (Section 3.4); these should be corrected.
  2. [Sec. 4 vs Appendix A.2] The same theorem is numbered Proposition 1 in Section 4 and Proposition 2 in Appendix A.2; the numbering should be harmonized.
  3. [Sec. 3.2] The notation θj ∈ R^K_+ is described as a 'factor score matrix', but it is a K-dimensional vector per sample; the text should say 'factor score vector'.
  4. [Fig. 3 and Table 4] The x-axis label '1 - Sparsity' in Figure 3 is at odds with the text's description 'x-axis shows the proportion of non-sparse weights' and with the '1-Sparsity' heading in Table 4; please define and use one consistent sparsity convention.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the empirical claims are measured against external benchmarks, and the theoretical identifiability argument rests on an external NMF theorem rather than on the paper's own definitions.

full rationale

The paper's central claims are that BNDL improves accuracy, uncertainty calibration (PAvPU), and interpretability. These are evaluated on held-out benchmark data (CIFAR-10, CIFAR-100, ImageNet-1k, Places-10) and compared with published baselines; none of the reported metrics are constructed from the method's own assumptions or fitted to the target outcome. The theoretical disentanglement claim in Section 4 imports an identifiability result from Gillis and Rajkó (2023), an external source with no author overlap, so it is not a self-citation chain. The Weibull variational inference machinery cites Zhang et al. (2018), which includes two of the present authors, but the reparameterization and closed-form KL divergence are stated explicitly in Eqs. (6)-(7), and the citation is not the sole support for the inference method; it is standard material and not load-bearing. The main logical weakness is that Eq. (14) replaces the categorical ELBO of Eq. (13) with a least-squares NMF objective and the per-column sparsity/selective-window hypotheses of Proposition 1 are not verified; however, that is an unproven bridge or correctness gap, not a circular reduction in which a prediction is equivalent to its input by construction. No fitted parameter is renamed as a prediction, no known result is repackaged as new, and no uniqueness conclusion is imported from the authors' own prior work. Accordingly, the appropriate circularity finding is none.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central model introduces latent variables theta and Phi, which are standard factor-analysis components rather than new physical entities; no new external entities are postulated.

free parameters (2)
  • Latent dimension K
    The number of non-negative factors for theta and Phi is not stated in the paper; it controls capacity and sparsity.
  • Sparsity threshold alpha
    Used in ReLU(w - alpha) to control decision-layer sparsity in Sec. 5.1.2; value is a hyperparameter that sets the sparsity-accuracy trade-off.
assumptions (3)
  • standard math Partial identifiability theorem for exact NMF under selective window and sparsity constraints (Gillis and Rajko, 2023).
    The paper relies on this external theorem to claim that BNDL learns disentangled features; the theorem itself is proved elsewhere, but the paper must verify its preconditions.
  • domain assumption The learned factor loading matrix Phi satisfies the selective window and sparsity constraints.
    Section 4 asserts BNDL satisfies these conditions due to gamma priors and ReLU sparsity, but it does not verify them for the trained models; the measured sparsity may be insufficient for the theorem's requirement.
  • ad hoc to paper The categorical ELBO objective is equivalent to the squared-error NMF objective ||Y - theta Phi||_F^2.
    Equation (14) replaces the categorical likelihood and KL terms with a least-squares NMF objective without derivation; this equivalence is necessary for the identifiability argument but is not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Uncertainty Estimation and Interpretability via Bayesian Non-negative Decision Layer." pith.science (2026). https://pith.science/paper/OTPU4DTX

@misc{pith2026250522199,
  author       = {Pith},
  title        = {Pith review of: Enhancing Uncertainty Estimation and Interpretability via Bayesian Non-negative Decision Layer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTPU4DTX}},
  note         = {Machine review of arXiv:2505.22199}
}
read the original abstract

Although deep neural networks have demonstrated significant success due to their powerful expressiveness, most models struggle to meet practical requirements for uncertainty estimation. Concurrently, the entangled nature of deep neural networks leads to a multifaceted problem, where various localized explanation techniques reveal that multiple unrelated features influence the decisions, thereby undermining interpretability. To address these challenges, we develop a Bayesian Non-negative Decision Layer (BNDL), which reformulates deep neural networks as a conditional Bayesian non-negative factor analysis. By leveraging stochastic latent variables, the BNDL can model complex dependencies and provide robust uncertainty estimation. Moreover, the sparsity and non-negativity of the latent variables encourage the model to learn disentangled representations and decision layers, thereby improving interpretability. We also offer theoretical guarantees that BNDL can achieve effective disentangled learning. In addition, we developed a corresponding variational inference method utilizing a Weibull variational inference network to approximate the posterior distribution of the latent variables. Our experimental results demonstrate that with enhanced disentanglement capabilities, BNDL not only improves the model's accuracy but also provides reliable uncertainty estimation and improved interpretability.

Figures

Figures reproduced from arXiv: 2505.22199 by the authors.

Figure 1
Figure 1. Illustration of the graphical models. 1(a): the predictive process of output Y for the baseline Deep Neural Network; 1(b): the generative model of DNNs with introducing stochastic latent variable θ ; 1(c): the generation model of the Bayesian non-negative decision layer; and 1(d): corresponding approximate inference for latent variables θ. narayanan et al., 2017) combine neural networks with stochastic outputs to qu… view at source ↗
Figure 2
Figure 2. The leftmost line chart illustrates the average uncertainty and accuracy across subsets of [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Sparsity-accuracy trade-offs for BNDL and Debuggable Network (Wong et al., 2021). [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The LIME visualizations for BNDL and ResNet-50, focusing on the largest [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Uncertainty Vs Test Acc Curve on other datasets and model. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: The LIME visualization results for BNDL ,ResNet-50 Debuggable Networks and BM, [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: We applied GradCAM to visualize ResNet50 and BNDL, and the corresponding heatmaps [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: The LIME visualization results for BNDL and ResNet-50, focusing on the largest [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: The LIME visualization results for BNDL and ResNet-50, focusing on the largest [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: The LIME visualization results for BNDL and ResNet-50, focusing on the largest [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]
Figure 11
Figure 11. Figure 11: The LIME visualization results for BNDL and ResNet-50, focusing on the largest [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: The LIME visualization results for BNDL and ResNet-50, focusing on the largest [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 37 canonical work pages

  1. [1]

    Sanity checks for saliency maps

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018

    2018

  2. [2]

    Deep evidential regression

    Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. Advances in neural information processing systems, 33: 0 14927--14937, 2020

    work page 2020

  3. [3]

    Concrete problems in ai safety

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Man \'e . Concrete problems in ai safety. arXiv preprint arXiv:1606.06565, 2016

    arXiv 2016

  4. [4]

    Weight uncertainty in neural network

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pp.\ 1613--1622. PMLR, 2015

    2015

  5. [5]

    A recurrent latent variable model for sequential data

    Junyoung Chung, Kyle Kastner, Laurent Dinh, Kratarth Goel, Aaron C Courville, and Yoshua Bengio. A recurrent latent variable model for sequential data. Advances in neural information processing systems, 28, 2015

    work page 2015

  6. [6]

    Laplace redux-effortless bayesian deep learning

    Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace redux-effortless bayesian deep learning. Advances in Neural Information Processing Systems, 34: 0 20089--20103, 2021

    2021

  7. [7]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp.\ 248--255. Ieee, 2009

    2009

  8. [8]

    Variational classification

    Shehzaad Dhuliawala, Mrinmaya Sachan, and Carl Allen. Variational classification. Trans. Mach. Learn. Res., 2024, 2023 a . URL https://api.semanticscholar.org/CorpusID:258740937

    work page 2024

Show all 77 references
  1. [9]

    Variational classification

    Shehzaad Dhuliawala, Mrinmaya Sachan, and Carl Allen. Variational classification. arXiv preprint arXiv:2305.10406, 2023 b

    2023 arXiv

  2. [10]

    Theory and evaluation metrics for learning disentangled representations

    Kien Do and Truyen Tran. Theory and evaluation metrics for learning disentangled representations. arXiv preprint arXiv:1908.09961, 2019

    1908 arXiv

  3. [11]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  4. [12]

    Sawtooth factorial topic embeddings guided gamma belief network

    Zhibin Duan, Dongsheng Wang, Bo Chen, Chaojie Wang, Wenchao Chen, Yewen Li, Jie Ren, and Mingyuan Zhou. Sawtooth factorial topic embeddings guided gamma belief network. In International Conference on Machine Learning, pp.\ 2903--2913. PMLR, 2021

    2021

  5. [13]

    A non-negative vae: the generalized gamma belief network

    Zhibin Duan, Tiansheng Wen, Muyao Wang, Bo Chen, and Mingyuan Zhou. A non-negative vae: the generalized gamma belief network. arXiv preprint arXiv:2408.03388, 2024

    2024 arXiv

  6. [14]

    Contextual dropout: An efficient sample-dependent dropout module

    Xinjie Fan, Shujian Zhang, Korawat Tanwisuth, Xiaoning Qian, and Mingyuan Zhou. Contextual dropout: An efficient sample-dependent dropout module. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=ct8_a9h1M

    2021

  7. [15]

    Sidiropoulos

    Xiao Fu, Kejun Huang, and Nicholas D. Sidiropoulos. On identifiability of nonnegative matrix factorization. IEEE Signal Processing Letters, 25: 0 328--332, 2017. URL https://api.semanticscholar.org/CorpusID:3483466

    2017

  8. [16]

    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, pp.\ 1050--1059. PMLR, 2016

    2016

  9. [17]

    Nonnegative matrix factorization

    Nicolas Gillis. Nonnegative matrix factorization. SIAM, 2020

    2020

  10. [18]

    Partial identifiability for nonnegative matrix factorization

    Nicolas Gillis and R \'o bert Rajk \'o . Partial identifiability for nonnegative matrix factorization. SIAM Journal on Matrix Analysis and Applications, 44 0 (1): 0 27--52, 2023

    2023

  11. [19]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pp.\ 1321--1330. PMLR, 2017

    2017

  12. [20]

    Card: Classification and regression diffusion models

    Xizewen Han, Huangjie Zheng, and Mingyuan Zhou. Card: Classification and regression diffusion models. ArXiv, abs/2206.07275, 2022. URL https://api.semanticscholar.org/CorpusID:249674651

    2022 arXiv

  13. [21]

    HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma

    Jeff Z. HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma. Provable guarantees for self-supervised deep learning with spectral contrastive loss. In Neural Information Processing Systems, 2021. URL https://api.semanticscholar.org/CorpusID:235367888

    2021

  14. [22]

    Variational bayesian last layers

    James Harrison, John Willes, and Jasper Snoek. Variational bayesian last layers. arXiv preprint arXiv:2404.11599, 2024

    2024 arXiv

  15. [23]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

    2016

  16. [24]

    Probabilistic backpropagation for scalable learning of bayesian neural networks

    Jos \'e Miguel Hern \'a ndez-Lobato and Ryan Adams. Probabilistic backpropagation for scalable learning of bayesian neural networks. In International conference on machine learning, pp.\ 1861--1869. PMLR, 2015

    2015

  17. [25]

    Non-negative matrix factorization with sparseness constraints

    Patrik O Hoyer. Non-negative matrix factorization with sparseness constraints. Journal of machine learning research, 5 0 (9), 2004

    2004

  18. [26]

    Non-negative matrix factorization revisited: Uniqueness and algorithm for symmetric decomposition

    Kejun Huang, Nicholas D Sidiropoulos, and Ananthram Swami. Non-negative matrix factorization revisited: Uniqueness and algorithm for symmetric decomposition. IEEE Transactions on Signal Processing, 62 0 (1): 0 211--224, 2013

    2013

  19. [27]

    Being bayesian about categorical probability

    Taejong Joo, Uijung Chung, and Minji Seo. Being bayesian about categorical probability. In International Conference on Machine Learning, 2020. URL https://api.semanticscholar.org/CorpusID:211171911

    2020

  20. [28]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun (eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015. URL http://arxiv.or...

    2015 arXiv

  21. [29]

    Stochastic gradient vb and the variational auto-encoder

    Diederik P Kingma and Max Welling. Stochastic gradient vb and the variational auto-encoder. In Second international conference on learning representations, ICLR, volume 19, pp.\ 121, 2014

    2014

  22. [30]

    Semi-supervised learning with deep generative models

    Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. Advances in neural information processing systems, 27, 2014

    2014

  23. [31]

    Variational dropout and the local reparameterization trick

    Durk P Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick. Advances in neural information processing systems, 28, 2015

    2015

  24. [32]

    Stochastic gradient variational bayes for gamma approximating distributions

    David A Knowles. Stochastic gradient variational bayes for gamma approximating distributions. arXiv preprint arXiv:1509.01631, 2015

    2015 arXiv

  25. [33]

    Being bayesian, even just a bit, fixes overconfidence in relu networks

    Agustinus Kristiadi, Matthias Hein, and Philipp Hennig. Being bayesian, even just a bit, fixes overconfidence in relu networks. In International conference on machine learning, pp.\ 5436--5446. PMLR, 2020

    2020

  26. [34]

    Learning multiple layers of features from tiny images

    A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009

    2009

  27. [35]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017

    2017

  28. [36]

    Towards falsifiable interpretability research

    Matthew L Leavitt and Ari Morcos. Towards falsifiable interpretability research. arXiv preprint arXiv:2010.12016, 2020

    2010 arXiv

  29. [37]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521 0 (7553): 0 436--444, 2015

    2015

  30. [38]

    Learning the parts of objects by non-negative matrix factorization

    Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401 0 (6755): 0 788--791, 1999

    1999

  31. [39]

    Preconditioned stochastic gradient langevin dynamics for deep neural networks

    Chunyuan Li, Changyou Chen, David Carlson, and Lawrence Carin. Preconditioned stochastic gradient langevin dynamics for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016

    2016

  32. [40]

    Deep ensembling with no overhead for either training or testing: The all-round blessings of dynamic sparsity

    Shiwei Liu, Tianlong Chen, Zahra Atashgahi, Xiaohan Chen, Ghada Sokar, Elena Mocanu, Mykola Pechenizkiy, Zhangyang Wang, and Decebal Constantin Mocanu. Deep ensembling with no overhead for either training or testing: The all-round blessings of dynamic sparsity. arXiv preprint ...

    2021

  33. [41]

    Energy-based out-of-distribution detection

    Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in neural information processing systems, 33: 0 21464--21475, 2020

    2020

  34. [42]

    Multiplicative normalizing flows for variational bayesian neural networks

    Christos Louizos and Max Welling. Multiplicative normalizing flows for variational bayesian neural networks. In International Conference on Machine Learning, pp.\ 2218--2227. PMLR, 2017

    2017

  35. [43]

    Predictive uncertainty estimation via prior networks

    Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. Advances in neural information processing systems, 31, 2018

    2018

  36. [44]

    Regression prior networks

    Andrey Malinin, Sergey Chervontsev, Ivan Provilkov, and Mark Gales. Regression prior networks. arXiv preprint arXiv:2006.11590, 2020

    2006 arXiv

  37. [45]

    Spade: Sparsity-guided debugging for deep neural networks

    Arshia Soltani Moakhar, Eugenia Iofinova, and Dan Alistarh. Spade: Sparsity-guided debugging for deep neural networks. arXiv preprint arXiv:2310.04519, 2023

    2023 arXiv

  38. [46]

    Identifiable deep generative models via sparse decoding

    Gemma E Moran, Dhanya Sridhar, Yixin Wang, and David M Blei. Identifiable deep generative models via sparse decoding. arXiv preprint arXiv:2110.10804, 2021

    2021 arXiv

  39. [47]

    Evaluating bayesian deep learning methods for semantic segmentation

    Jishnu Mukhoti and Yarin Gal. Evaluating bayesian deep learning methods for semantic segmentation. arXiv preprint arXiv:1811.12709, 2018

    2018 arXiv

  40. [48]

    Reparameterization gradients through acceptance-rejection sampling algorithms

    Christian Naesseth, Francisco Ruiz, Scott Linderman, and David Blei. Reparameterization gradients through acceptance-rejection sampling algorithms. In Artificial Intelligence and Statistics, pp.\ 489--498. PMLR, 2017

    2017

  41. [49]

    Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks

    Anh Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. arXiv preprint arXiv:1602.03616, 2016

    2016 arXiv

  42. [50]

    Benchmarking the neural linear model for regression

    Sebastian W Ober and Carl E Rasmussen. Benchmarking the neural linear model for regression. In Second Symposium on Advances in Approximate Bayesian Inference, 2019

    2019

  43. [51]

    Feature visualization

    Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2 0 (11): 0 e7, 2017

    2017

  44. [52]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  45. [53]

    Listen to interpret: Post-hoc interpretability for audio networks with nmf

    Jayneel Parekh, Sanjeel Parekh, Pavlo Mozharovskyi, Florence d'Alch \'e Buc, and Ga \"e l Richard. Listen to interpret: Post-hoc interpretability for audio networks with nmf. Advances in Neural Information Processing Systems, 35: 0 35270--35283, 2022

    2022

  46. [54]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pp.\ 1135--1144, 2016

    2016

  47. [55]

    The generalized reparameterization gradient

    Francisco R Ruiz, Titsias RC AUEB, David Blei, et al. The generalized reparameterization gradient. Advances in neural information processing systems, 29, 2016

    2016

  48. [56]

    Imagenet large scale visual recognition challenge

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115: 0 211--252, 2015

    2015

  49. [57]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pp.\ 618--626, 2017

    2017

  50. [58]

    Evidential deep learning to quantify classification uncertainty

    Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems, 31, 2018

    2018

  51. [59]

    Kernel implicit variational inference

    Jiaxin Shi, Shengyang Sun, and Jun Zhu. Kernel implicit variational inference. arXiv preprint arXiv:1705.10119, 2017

    2017 arXiv

  52. [60]

    Deep inside convolutional networks: Visualising image classification models and saliency maps

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013

    2013 arXiv

  53. [61]

    Learning structured output representation using deep conditional generative models

    Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015

    2015

  54. [62]

    Collapsed variational bounds for bayesian neural networks

    Marcin Tomczak, Siddharth Swaroop, Andrew Foong, and Richard Turner. Collapsed variational bounds for bayesian neural networks. Advances in Neural Information Processing Systems, 34: 0 25412--25426, 2021

    2021

  55. [63]

    Non-negative contrastive learning

    Yifei Wang, Qi Zhang, Yaoyu Guo, and Yisen Wang. Non-negative contrastive learning. In The Twelfth International Conference on Learning Representations, 2024

    2024

  56. [64]

    Latent derivative bayesian last layer networks

    Joe Watson, Jihao Andreas Lin, Pascal Klink, Joni Pajarinen, and Jan Peters. Latent derivative bayesian last layer networks. In International Conference on Artificial Intelligence and Statistics, pp.\ 1198--1206. PMLR, 2021

    2021

  57. [65]

    Optimizing over a bayesian last layer

    Noah Weber, Janez Starc, Arpit Mittal, Roi Blanco, and Llu \' s M \`a rquez. Optimizing over a bayesian last layer. In NeurIPS workshop on Bayesian Deep Learning, 2018

    2018

  58. [66]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning

    Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8: 0 229--256, 1992

    1992

  59. [67]

    Leveraging sparse linear layers for debuggable deep networks

    Eric Wong, Shibani Santurkar, and Aleksander Madry. Leveraging sparse linear layers for debuggable deep networks. In International Conference on Machine Learning, pp.\ 11205--11216. PMLR, 2021

    2021

  60. [68]

    Understanding neural networks through deep visualization

    Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015

    2015 arXiv

  61. [69]

    Gaussian-process factor analysis for low-dimensional single-trial analysis of neural population activity

    Byron M Yu, John P Cunningham, Gopal Santhanam, Stephen Ryu, Krishna V Shenoy, and Maneesh Sahani. Gaussian-process factor analysis for low-dimensional single-trial analysis of neural population activity. Advances in neural information processing systems, 21, 2008

    2008

  62. [70]

    Whai: Weibull hybrid autoencoding inference for deep topic modeling

    Hao Zhang, Bo Chen, Dandan Guo, and Mingyuan Zhou. Whai: Weibull hybrid autoencoding inference for deep topic modeling. arXiv preprint arXiv:1803.01328, 2018

    2018 arXiv

  63. [71]

    Places: A 10 million image database for scene recognition

    Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40 0 (6): 0 1452--1464, 2017

    2017

  64. [72]

    Beta-negative binomial process and P oisson factor analysis

    Mingyuan Zhou, Lauren Hannah, David Dunson, and Lawrence Carin. Beta-negative binomial process and P oisson factor analysis. In Artificial Intelligence and Statistics, pp.\ 1462--1471. PMLR, 2012

    2012

  65. [73]

    The P oisson gamma belief network

    Mingyuan Zhou, Yulai Cong, and Bo Chen. The P oisson gamma belief network. Advances in Neural Information Processing Systems, 28, 2015

    2015

  66. [74]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  67. [75]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  68. [76]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  69. [77]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.