Multi-Method Ensemble for Out-of-Distribution Detection

Lucas Rakotoarivony

REVIEW 4 major objections 4 minor 40 references

Multi-Method Ensemble for Out-of-Distribution Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Multiplying five out-of-distribution detectors into a single score beats every one of them alone.

desk verdict Useful ensembling study with a legitimate new score, but the headline SOTA claim is compromised by a duplicated baseline row and the theory is not sound. read the letter →

arxiv 2508.21463 v1 pith:FMPY4QWP submitted 2025-08-29 cs.CV

Lucas Rakotoarivony This is my paper

classification cs.CV

keywords out-of-distributiondetectionensemblescoringfeaturetruncationmultiplicativescoreImageNet-1KCIFARpost-hoccovarianceassumption

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 argues that out-of-distribution (OOD) detection becomes stronger when several existing detectors are multiplied into one score rather than used individually. It proposes MME, a product of two feature-truncation techniques (SCALE and VRA) and five scoring functions (fDBD, PCA, ViM, NME+, and a consistency factor CO+). The theoretical anchor is Proposition 1: under a covariance condition, the product of scoring functions separates in-distribution from OOD data at least as well as the best individual function. Empirically, MME reports an average false-positive rate of 27.57% on ImageNet-1K with a BiT model, about 6 percentage points lower than the best baseline, and top results on CIFAR-10/100 and near/far-OOD splits. If the claim holds, safety-critical systems could get a better detector by ensembling already published methods with no retraining.

What carries the argument

The central object is the multiplicative ensemble score MME(x), a product of five factors: an exponential of SCALE minus ViM(VRA), fDBD(VRA), PCA(VRA), CO+(x), and NME+(x). The identity that carries the argument is Proposition 1, which states that under the covariance domination assumption Cov_in(S_i,S_j) ≥ Cov_out(S_i,S_j), multiplying scoring functions yields a separation V that is at least as large as the maximum individual V. The consistency factor CO+(x) is a piecewise score that rewards agreement between the predicted classes of a logit-based score (MLS) and a distance-based score (NME+); it serves as an empirical proxy for the covariance assumption. Feature truncation (SCALE, VRA) is

What would settle it

On a pre-trained BiT or ViT ImageNet classifier, compute Cov_in and Cov_out for each pair of the five component scores used in MME. If any pair has Cov_out > Cov_in, Assumption 1 fails and Proposition 1 does not apply. Then compare MME's FPR95 against the best individual component on a held-out OOD set (e.g., a subset of ImageNet-C); a component that beats MME, or a pair with Cov_out > Cov_in, would refute the claim.

Watch

Extended reading notes

Core claim

The central claim is that the MME score defined by Eq. (7), MME(x) = exp(SCALE(x) − ViM(VRA(x))) · fDBD(VRA(x)) · PCA(VRA(x)) · CO+(x) · NME+(x), outperforms each individual post-hoc OOD detector it combines, across ImageNet-1K, CIFAR-10 and CIFAR-100, including near-OOD and far-OOD scenarios. The supporting theoretical claim is Proposition 1: for scoring functions S1,…,Sn satisfying Assumption 1 (the covariance of any two scores is at least as large on in-distribution data as on OOD data), the mean separation V(S1·…·Sn) is at least max_i V(Si). The paper relies on a working hypothesis that feature truncation g that increases raw feature separation also increases the separation of any scorin

Load-bearing premise

The load-bearing premise is that any feature truncation that increases the raw in-distribution versus out-of-distribution feature gap will also increase the gap of whatever scoring function is applied to the truncated features; the authors say no general proof exists and their own ablations show the benefit is small or absent for several component scores.

Editorial extensions

If this is right

Existing OOD scoring functions can be combined multiplicatively to produce a detector that is no worse than its best member, without retraining.
Feature truncation methods such as SCALE and VRA can be attached to modern scoring functions, not just simple energy or logit scores, with average gains.
The agreement between classifier-based and distance-based predictions (CO+) is itself a usable OOD signal.
On ImageNet-1K with a BiT backbone, the ensemble reduces the false-positive rate at 95% recall by about 6 percentage points over the strongest individual method.
The gains persist across ViT and BiT backbones and across small- and large-scale benchmarks, indicating the ensemble effect is not specific to one architecture.

Reading between the lines

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

Multiplicative combination is sensitive to score scales; a factor with tiny absolute values could dominate the product, a sensitivity the paper does not explore.
The covariance condition could serve as a practical selection criterion: before ensembling, estimate Cov_in and Cov_out on a validation set and drop any score pair that violates the assumption.
The paper's own ablations show truncation gives little to fDBD, PCA and NME+; a future variant might apply VRA and SCALE only to the factors that actually benefit.
The same multiplicative-ensemble recipe could transfer to other two-population score problems, such as anomaly or novelty detection, after checking the covariance assumption.

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.

Desk Editor's Note

read the letter

Two things to know. The product-ensemble score MME is a genuinely new construction, and the empirical sweep across CIFAR and ImageNet suggests that combining feature truncation with multiple scoring functions buys real robustness. But the headline number is not trustworthy: the BiT row for VRA in Table 3 is a verbatim copy of the ReAct row for every dataset, and the 'Baseline' row in Table 6 is identical to the MME row. That looks like copy-paste, and it directly undercuts the claimed 6-point FPR95 improvement over the best baseline. You cannot compute the margin without knowing what VRA actually achieves on BiT.

What earns credit: the ensembling design is sensible; the ablations in Tables 4, 6, and 7 give a useful map of which components matter; and the covariance analysis in Table 5 is a reasonable empirical handle on the key correlation assumption. The authors are also honest that Hypothesis 1 is unproven and say so explicitly. That is more than many papers do.

The soft spots are real. Proposition 1's proof in Appendix A rescales S2 so that E_out(S2) >= 1, but the scoring function is not invariant to arbitrary rescaling after the fact; the product V(S1*S2) is computed before that rescaling. Assumption 1 is asserted, empirically motivated, but not derived. The authors essentially admit the theory is a working hypothesis, so the paper should stand on the empirical claims alone. Stripping or heavily reworking the theory would be an improvement.

There are also reporting issues: no code, no error bars, hyperparameters T and lambda are chosen per benchmark, and near/far-OOD results appear only as a figure. The duplicated rows lower confidence in the other numbers as well. The method itself is not obviously broken, and the core practical claim—that multiplying a set of diverse OOD scores works better than any one—is plausible and worth a serious look.

Who this is for: anyone doing post-hoc OOD detection who wants a practical ensemble recipe and a rough map of which recent scores combine well. It deserves a serious referee, but the referee should ask for code, a corrected Table 3, and a version that stops pretending the math proves more than it does.

Referee Report

4 major / 4 minor

Summary. The paper proposes a post-hoc OOD detection score, MME, which multiplies several existing detectors (ViM, fDBD, PCA, a newly introduced NME+, and a covariance-motivated CO+) after applying SCALE and VRA feature truncation. A theoretical Proposition 1 claims that the product of scoring functions is at least as good as the best component under a covariance assumption, and Hypothesis 1 asserts that feature truncation improves any scoring function. The method is evaluated on CIFAR-10/100 and ImageNet-1K, including near-OOD and far-OOD benchmarks. The headline result is an average FPR95 of 27.57% on ImageNet-1K with the BiT backbone, reported as about 6% better than the best baseline.

Significance. If the empirical results are correct, the paper offers a simple and practical recipe for improving post-hoc OOD detection by ensembling complementary scoring functions, and the ablation studies over components and hyperparameters are fairly thorough. However, the theoretical argument is not sound, and an obvious anomaly in Table 3 (VRA identical to ReAct on BiT) undermines the headline comparison until verified. The empirical contribution could be valuable, but the current manuscript overstates its theoretical and empirical support.

major comments (4)

[Table 3 (BiT block)] The VRA row is identical to the ReAct row for every dataset and for the average (iNaturalist 91.50/48.65, Places 87.96/54.67, SUN 92.33/38.99, Textures 90.65/50.14, average 90.61/48.11). Two distinct methods should not coincide to two decimals, and the ViT block shows VRA differing from ReAct. This is almost certainly a reporting error. The claimed '6% improvement' compares MME's 27.57 to ASH-S's 33.19; if the correct BiT VRA value is close to its ViT value (27.14), MME is not better than the strongest baseline. Please rerun the experiment, correct the table, and add error bars or release code so the result can be checked.
[Appendix A, Eq. (8)-(11)] The proof of Proposition 1 is not valid. The statement that 'a scoring function is invariant to scaling' is false because V(λS)=λV(S). Scaling S2 so that E_out(S2)≥1 changes the quantity being bounded; an inequality for the scaled product does not imply the original inequality. The recurrence step for n>2 also requires Assumption 1 to hold for composite products, which is not established. Thus the theoretical demonstration in Section 3.3 is unsound.
[Section 3.3, Hypothesis 1] Hypothesis 1 is stated as central, but no proof is offered and the authors admit 'a general proof is elusive'. The empirical support is mixed: Figure 2 shows Mahalanobis degrades with feature truncation, and Table 6 shows fDBD and PCA with SCALE give essentially identical numbers to the baseline (26.19 vs 26.20 and 26.20 vs 26.20 in average FPR95). Since Eq. (7) applies SCALE and VRA before most components, the design rests on this unverified hypothesis. The abstract's claim of a theoretical demonstration that feature truncation and scoring functions can be effectively combined is therefore overstated.
[Section 3.4, Eq. (7)] SCALE(x) is used as a scalar score in Eq. (7), but SCALE is defined in Section 3.1 as a feature-truncation method that produces a feature map. The notation is never defined. Presumably the authors mean an energy/MLS score computed on the SCALE-truncated features. Without a precise definition of SCALE(x), the central formula is not reproducible.

minor comments (4)

[Section 3.3] 'scalability properties' should be 'scaling properties' (or the intended property should be stated precisely).
[Appendix B.2] NME(x) is defined as argmin over class means, so NME(x)/T is a class index; softmax(NME(x)/T) is not a well-defined score. Please specify the distance used and the exact softmax input.
[Section 4.1 vs Appendix C.3] The main text fixes λ=2, but Appendix C.3 states the best AUROC is achieved at λ=1.5. Clarify this discrepancy.
[Section 4.2] 'Improvement of 6%' should be '6 percentage points' since FPR95 is already a percentage.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical MME evaluation is self-contained; the Proposition 1 proof and Table 3 raise correctness concerns, not circularity.

full rationale

The paper's central claim is empirical: MME is a product of existing scoring/truncation methods (ViM, fDBD, PCA, SCALE, VRA, NME+, CO+) evaluated on standard held-out OOD benchmarks. No parameter is fitted to those benchmarks; T and λ are fixed in advance (Section 4.1), and component hyperparameters are taken from the original papers. The comparison is therefore not circular in the sense of a fitted input being called a prediction. The theoretical Proposition 1 (Section 3.3 / Appendix A) does not create circularity: the proof's step assuming 'Eout(S2(x)) ≥ 1' after claiming scaling is WLOG is mathematically invalid because V(S1·λS2)=λV(S1·S2), but this is a proof gap, not a reduction of the conclusion to the assumptions. Hypothesis 1 is explicitly labeled a working hypothesis ('a general proof is elusive'), so it is an acknowledged assumption rather than a smuggled circular premise. The duplicated VRA/ReAct row in Table 3 (BiT) is a likely reporting error that weakens the reported 6% margin over the best baseline, but that is a data-integrity/correctness issue, not a circular derivation. There are no load-bearing self-citations and no renaming of known results as new. Hence circularity score 0.

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

The central method depends on unproven or empirically checked assumptions more than on a self-contained derivation. The product-ensemble formula introduces implicit weighting via component scales, and the theoretical guarantee requires both Assumption 1 and Hypothesis 1, the latter explicitly admitted as unproven.

free parameters (3)

Temperature T for NME+ = T = 0.5 (ImageNet-1K), T = 0.1 (CIFAR)
Controls softmax smoothness of NME+; tuned per benchmark in Section 4.1 and Appendix C.3; performance varies by a few points at extreme values.
CO+ weight lambda = lambda = 2
Controls influence of prediction consistency between NME+ and MLS; ablation in Appendix C.3 reports best AUROC at lambda = 1.5 while the paper uses lambda = 2, so the reported default is not the empirically optimal value.
Relative scales of product components in Eq. (7) = unspecified
Scores from exp(SCALE - ViM), fDBD, PCA, CO+ and NME+ are multiplied without stated normalization; their natural scales act as implicit fixed weights, a hidden degree of freedom that the experiments do not isolate.

assumptions (3)

domain assumption Assumption 1: Cov_in(S1,S2) >= Cov_out(S1,S2) for any pair of scoring functions used.
Stated in Section 3.1; used to prove Proposition 1. Empirically checked for six pairs on ImageNet-1K in Table 5, but not guaranteed for all pairs and products.
ad hoc to paper Hypothesis 1: feature truncation g improves any scoring function S, i.e., E_in[S(g(z))] - E_out[S(g(z))] >= E_in[S(z)] - E_out[S(z)].
Central to MME design; paper admits in Section 3.3 that no general proof is available and only empirical Figure 2 and Table 6 support it, with weak gains for several components.
ad hoc to paper Scoring functions are invariant to scaling, used in Appendix A to force E_out(S2) >= 1.
Not true for the V functional, which scales linearly; this makes Proposition 1 near-vacuous and is a logical gap in the proof.

invented entities (2)

NME+ score
purpose: Distance-based scoring function derived from the NME classifier, used as a multiplicative factor in MME.
The NME classifier is from Rebuffi et al., but the NME+ adaptation, defined in Section 3.1 and Appendix B, is new to this paper. Its usefulness is demonstrated only on the benchmarks in this paper.
CO+ score
purpose: Consistency score that rewards agreement between NME+ and MLS predictions, used as a multiplicative factor in MME.
New to this paper; introduced as a proxy for covariance in Section 3.1. No external validation outside the reported benchmarks.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Method Ensemble for Out-of-Distribution Detection." pith.science (2026). https://pith.science/paper/FMPY4QWP

@misc{pith2026250821463,
  author       = {Pith},
  title        = {Pith review of: Multi-Method Ensemble for Out-of-Distribution Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FMPY4QWP}},
  note         = {Machine review of arXiv:2508.21463}
}

read the original abstract

Detecting out-of-distribution (OOD) samples is essential for neural networks operating in open-world settings, particularly in safety-critical applications. Existing methods have improved OOD detection by leveraging two main techniques: feature truncation, which increases the separation between in-distribution (ID) and OOD samples, and scoring functions, which assign scores to distinguish between ID and OOD data. However, most approaches either focus on a single family of techniques or evaluate their effectiveness on a specific type of OOD dataset, overlooking the potential of combining multiple existing solutions. Motivated by this observation, we theoretically and empirically demonstrate that state-of-the-art feature truncation and scoring functions can be effectively combined. Moreover, we show that aggregating multiple scoring functions enhances robustness against various types of OOD samples. Based on these insights, we propose the Multi-Method Ensemble (MME) score, which unifies state-of-the-art OOD detectors into a single, more effective scoring function. Extensive experiments on both large-scale and small-scale benchmarks, covering near-OOD and far-OOD scenarios, show that MME significantly outperforms recent state-of-the-art methods across all benchmarks. Notably, using the BiT model, our method achieves an average FPR95 of 27.57% on the challenging ImageNet-1K benchmark, improving performance by 6% over the best existing baseline.

Figures

Figures reproduced from arXiv: 2508.21463 by the authors.

**Figure 2.** AUROC OOD detection performance on the ImageNet-1K benchmark for eight scoring functions combined with different feature truncation methods. Results are averaged across four OOD datasets. 3.2 Motivation Current post-hoc OOD detection methods can be divided into two main categories: feature truncation [24, 25] and scoring functions [8, 16]. Feature truncation methods aim to increase the gap between ID and OOD sampl… view at source ↗

**Figure 3.** Overview of the MME pipeline. First, features are extracted from the input and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Ablation study on the impact of λ and T hyperparameters. The AUROC performance of MME is reported on the ImageNet-1K benchmark for different values of λ (top row) and T (bottom row). The results are shown for the ViT architecture on the left and the BiT architecture o…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 25 canonical work pages

[1]

In or Out? Fixing ImageNet Out-of-Distribution Detection Evaluation

Julian Bitterwolf, Maximilian Mueller, and Matthias Hein. In or out? fixing imagenet out-of-distribution detection evaluation. arXiv preprint arXiv:2306.00826, 2023

work page Pith review arXiv 2023
[2]

Neat: Neural attention fields for end-to-end autonomous driving

Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793–15803, 2021

work page 2021
[3]

Describing textures in the wild

Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606–3613, 2014

2014
[4]

Extremely simple activation shaping for out-of-distribution detection

Andrija Djurisic, Nebojsa Bozanic, Arjun Ashok, and Rosanne Liu. Extremely simple activation shaping for out-of-distribution detection. arXiv preprint arXiv:2209.09858, 2022

arXiv 2022
[5]

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, Syl- vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

arXiv 2010
[6]

Revisit pca-based technique for out-of-distribution detection

Xiaoyuan Guan, Zhouwu Liu, Wei-Shi Zheng, Yuren Zhou, and Ruixuan Wang. Revisit pca-based technique for out-of-distribution detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19431–19439, 2023

work page 2023
[7]

Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem

Matthias Hein, Maksym Andriushchenko, and Julian Bitterwolf. Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 41–50, 2019

work page 2019
[8]

A baseline for detecting misclassified and out-of- distribution examples in neural networks

Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of- distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016

arXiv 2016

Show all 40 references

[9]

Scaling out-of-distribution detection for real-world settings

Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Moham- madreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. LUCAS RAKOTOARIVONY: MME 11

1911 arXiv
[10]

Densely connected convolutional networks

Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017

2017
[11]

Big transfer (bit): General visual representation learning

Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Pa...

2020
[12]

Learning multiple layers of features from tiny images

Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

2009
[13]

A simple unified framework for detecting out-of-distribution samples and adversarial attacks

Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems, 31, 2018

2018
[14]

Enhancing the reliability of out- of-distribution image detection in neural networks

Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out- of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017

2017 arXiv
[15]

Fast decision boundary based out-of-distribution detector

Litian Liu and Yao Qin. Fast decision boundary based out-of-distribution detector. arXiv preprint arXiv:2312.11536, 2023

2023 arXiv
[16]

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, 33:21464– 21475, 2020

2020
[17]

Gen: Pushing the limits of softmax-based out-of-distribution detection

Xixi Liu, Yaroslava Lochman, and Christopher Zach. Gen: Pushing the limits of softmax-based out-of-distribution detection. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 23946–23955, 2023

2023
[18]

Exploring uncertainty measures in deep networks for multiple sclerosis lesion detection and segmentation

Tanya Nair, Doina Precup, Douglas L Arnold, and Tal Arbel. Exploring uncertainty measures in deep networks for multiple sclerosis lesion detection and segmentation. Medical image analysis, 59:101557, 2020

2020
[19]

Reading digits in natural images with unsupervised feature learning

Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning , volume 2011, page 4. Granada, 2011

2011
[20]

Nearest neighbor guid- ance for out-of-distribution detection

Jaewoo Park, Yoon Gyo Jung, and Andrew Beng Jin Teoh. Nearest neighbor guid- ance for out-of-distribution detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1686–1695, 2023

2023
[21]

icarl: Incremental classifier and representation learning

Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lam- pert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017

2001
[22]

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: 211–252, 2015. 12 LUCAS RA...

2015
[23]

f-anogan: Fast unsupervised anomaly detection with generative ad- versarial networks

Thomas Schlegl, Philipp Seeböck, Sebastian M Waldstein, Georg Langs, and Ursula Schmidt-Erfurth. f-anogan: Fast unsupervised anomaly detection with generative ad- versarial networks. Medical image analysis, 54:30–44, 2019

2019
[24]

Dice: Leveraging sparsification for out-of-distribution de- tection

Yiyou Sun and Yixuan Li. Dice: Leveraging sparsification for out-of-distribution de- tection. In European conference on computer vision, pages 691–708. Springer, 2022

2022
[25]

React: Out-of-distribution detection with rectified activations

Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. Advances in neural information processing systems, 34:144–157, 2021

2021
[26]

Out-of-distribution detection with deep nearest neighbors

Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, pages 20827–20840. PMLR, 2022

2022
[27]

Appli- cations of machine learning in drug discovery and development

Jessica Vamathevan, Dominic Clark, Paul Czodrowski, Ian Dunham, Edgardo Ferran, George Lee, Bin Li, Anant Madabhushi, Parantu Shah, Michaela Spitzer, et al. Appli- cations of machine learning in drug discovery and development. Nature reviews Drug discovery, 18(6):463–477, 2019

2019
[28]

The inaturalist species classifica- tion and detection dataset

Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classifica- tion and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages...

2018
[29]

Open-set recognition: A good closed-set classifier is all you need? 2021

Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need? 2021

2021
[30]

Road anomaly detection by partial image reconstruction with segmenta- tion coupling

Tomas V ojir, Tomáš Šipka, Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino, and Jiri Matas. Road anomaly detection by partial image reconstruction with segmenta- tion coupling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15651–15660, 2021

2021
[31]

Vim: Out-of-distribution with virtual-logit matching

Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4921–4930, 2022

2022
[32]

Sun database: Large-scale scene recognition from abbey to zoo

Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In2010 IEEE computer so- ciety conference on computer vision and pattern recognition, pages 3485–3492. IEEE, 2010

2010
[33]

Scaling for training time and post-hoc out-of-distribution detection enhancement

Kai Xu, Rongyu Chen, Gianni Franchi, and Angela Yao. Scaling for training time and post-hoc out-of-distribution detection enhancement. arXiv preprint arXiv:2310.00227, 2023

2023 arXiv
[34]

Vra: variational rectified activation for out-of-distribution detection

Mingyu Xu, Zheng Lian, Bin Liu, and Jianhua Tao. Vra: variational rectified activation for out-of-distribution detection. Advances in Neural Information Processing Systems, 36:28941–28959, 2023

2023
[35]

Turkergaze: Crowdsourcing saliency with webcam based eye tracking

Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015. LUCAS RAKOTOARIVONY: MME 13

2015 arXiv
[36]

Openood: Benchmarking generalized out-of-distribution detection

Jingkang Yang, Pengyun Wang, Dejian Zou, Zitang Zhou, Kunyuan Ding, Wenxuan Peng, Haoqi Wang, Guangyao Chen, Bo Li, Yiyou Sun, et al. Openood: Benchmarking generalized out-of-distribution detection. Advances in Neural Information Processing Systems, 35:32598–32611, 2022

2022
[37]

Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop

Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015

2015 arXiv
[38]

Openood v1

Jingyang Zhang, Jingkang Yang, Pengyun Wang, Haoqi Wang, Yueqian Lin, Haoran Zhang, Yiyou Sun, Xuefeng Du, Yixuan Li, Ziwei Liu, et al. Openood v1. 5: Enhanced benchmark for out-of-distribution detection. arXiv preprint arXiv:2306.09301, 2023

2023 arXiv
[39]

Out-of-distribution detection based on in-distribution data patterns mem- orization with modern hopfield energy

Jinsong Zhang, Qiang Fu, Xu Chen, Lun Du, Zelin Li, Gang Wang, Shi Han, Dongmei Zhang, et al. Out-of-distribution detection based on in-distribution data patterns mem- orization with modern hopfield energy. In The Eleventh International Conference on Learning Representations, 2022

2022
[40]

confident

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(6):1452–1464, 2017. 14 LUCAS RAKOTOARIVONY: MME A Mathematical proof Here...

2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

In or Out? Fixing ImageNet Out-of-Distribution Detection Evaluation

Julian Bitterwolf, Maximilian Mueller, and Matthias Hein. In or out? fixing imagenet out-of-distribution detection evaluation. arXiv preprint arXiv:2306.00826, 2023

work page Pith review arXiv 2023

[2] [2]

Neat: Neural attention fields for end-to-end autonomous driving

Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793–15803, 2021

work page 2021

[3] [3]

Describing textures in the wild

Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606–3613, 2014

2014

[4] [4]

Extremely simple activation shaping for out-of-distribution detection

Andrija Djurisic, Nebojsa Bozanic, Arjun Ashok, and Rosanne Liu. Extremely simple activation shaping for out-of-distribution detection. arXiv preprint arXiv:2209.09858, 2022

arXiv 2022

[5] [5]

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, Syl- vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

arXiv 2010

[6] [6]

Revisit pca-based technique for out-of-distribution detection

Xiaoyuan Guan, Zhouwu Liu, Wei-Shi Zheng, Yuren Zhou, and Ruixuan Wang. Revisit pca-based technique for out-of-distribution detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19431–19439, 2023

work page 2023

[7] [7]

Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem

Matthias Hein, Maksym Andriushchenko, and Julian Bitterwolf. Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 41–50, 2019

work page 2019

[8] [8]

A baseline for detecting misclassified and out-of- distribution examples in neural networks

Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of- distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016

arXiv 2016