arxiv: 2604.04071 · v1 · submitted 2026-04-05 · 💻 cs.CV

Detecting Media Clones in Cultural Repositories Using a Positive Unlabeled Learning Approach

V. Sevetlidis , V. Arampatzakis , M. Karta , I. Mourthos , D. Tsiafaki , G. Pavlidis This is my paper

Pith reviewed 2026-05-13 17:33 UTC · model grok-4.3

classification 💻 cs.CV

keywords positive unlabeled learningduplicate detectioncultural heritageimage cloningPU learninganchor-based encodingcurator-in-the-loop

0 comments p. Extension

The pith

Positive-unlabeled learning detects media clones in cultural repositories using only one anchor image per artifact.

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

The paper frames curator-in-the-loop duplicate discovery in image archives as a positive-unlabeled learning problem. With a single verified anchor per artifact, a lightweight encoder is trained on its augmented views and then scores the rest of the unlabeled collection by latent l2 norm distance. This produces ranked candidate lists for human verification, reaching F1 of 90.79 and AUROC of 98.99 on the AtticPOT cultural dataset while improving 7.70 F1 points over the SVDD baseline under the same backbone. The approach avoids the need for explicit negative labels and supplies an interpretable operating threshold.

Core claim

Treating each artifact as its own positive class defined by one anchor, the method trains a per-query Clone Encoder on augmented views of that anchor; unlabeled images are then ranked by the l2 norm of their embeddings relative to the anchor, with an interpretable threshold selecting candidates for curator review. On AtticPOT this yields F1=90.79 and AUROC=98.99, outperforming SVDD by 7.70 F1 points under identical lightweight backbone constraints, while also delivering F1=96.37 on CIFAR-10.

What carries the argument

Per-query Clone Encoder trained on augmented views of a single anchor, scoring unlabeled items via interpretable threshold on latent l2 norm.

If this is right

Reaches F1=90.79 and AUROC=98.99 on the AtticPOT cultural repository.
Improves F1 by 7.70 points over the SVDD baseline under the same lightweight backbone.
Proposes duplicate candidates for curator verification without requiring pre-labeled negatives.
Produces stable similarity neighborhoods across viewpoint and condition changes.
Fits de-duplication, record linkage, and curator-in-the-loop workflows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Single-anchor PU setups could apply directly to medical or document archives where negatives are scarce.
Parallel per-query training would let the method scale to repositories with millions of images.
Active learning on the ranked candidates could further cut curator labeling effort.

Load-bearing premise

Augmented views of a single anchor sufficiently represent the positive class for effective PU learning without explicit negatives in this cultural image domain.

What would settle it

A new cultural image collection containing known cross-record duplicates where the l2-norm threshold on the encoder embeddings yields F1 below 80.

Figures

Figures reproduced from arXiv: 2604.04071 by D. Tsiafaki, G. Pavlidis, I. Mourthos, M. Karta, V. Arampatzakis, V. Sevetlidis.

**Figure 1.** Figure 1: Proposed clone-detection workflow: an anchor is augmented into clones and contrasted with an unlabeled [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Training batch illustration on CIFAR–10. Panels A: the anchor image, B: clones of the anchor via augmentation [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Duplicate detection performance on CIFAR-10 (left) and AtticPOT (right). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Qualitative retrieval on AtticPOT. For each query (leftmost image in a row), we show the nine highest-ranked [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Positive vs. negative latent norm distributions for a sample anchor (left). Across anchors: distribution of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Calibration around the operating threshold [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

read the original abstract

We formulate curator-in-the-loop duplicate discovery in the AtticPOT repository as a Positive-Unlabeled (PU) learning problem. Given a single anchor per artefact, we train a lightweight per-query Clone Encoder on augmented views of the anchor and score the unlabeled repository with an interpretable threshold on the latent l_2 norm. The system proposes candidates for curator verification, uncovering cross-record duplicates that were not verified a priori. On CIFAR-10 we obtain F1=96.37 (AUROC=97.97); on AtticPOT we reach F1=90.79 (AUROC=98.99), improving F1 by +7.70 points over the best baseline (SVDD) under the same lightweight backbone. Qualitative "find-similar" panels show stable neighbourhoods across viewpoint and condition. The method avoids explicit negatives, offers a transparent operating point, and fits de-duplication, record linkage, and curator-in-the-loop workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper gives a workable PU learning method for spotting clones in cultural image collections, with solid reported gains but limited checks on whether the positive class really captures real-world duplicates.

read the letter

Your colleague should know that this work takes positive-unlabeled learning and applies it directly to finding duplicate media records in a cultural repository. They train a lightweight encoder on augmented versions of a single anchor image per item, then score the rest of the collection with an l2 norm threshold in the latent space. On AtticPOT they report F1=90.79 and AUROC=98.99, which beats their SVDD baseline by 7.7 F1 points under the same backbone. CIFAR-10 numbers are even higher. The approach does a few things well. It stays simple and interpretable, which suits a curator-in-the-loop setup where someone has to verify the suggestions. No need for explicit negative examples, and the threshold gives a clear operating point. The qualitative results show it can group similar views despite changes in condition. The main soft spot is whether the augmentations from one anchor actually represent the positive class for real clones. Cultural images often vary in viewpoint, lighting, and degradation in ways that standard transforms might miss. The paper doesn't appear to include a direct test showing that held-out duplicates fall within the chosen l2 ball, so the performance numbers could be optimistic for this domain. Threshold selection details are also thin in the abstract, though the full text might clarify that. This paper is aimed at practitioners in digital cultural heritage who need practical de-duplication tools. A reader working on PU learning or small-scale image retrieval would find the formulation useful. The work is grounded enough in its empirical results to merit peer review, even if the central assumption needs more scrutiny in revision. Recommendation: Send it for review.

Referee Report

2 major / 2 minor

Summary. The paper formulates curator-in-the-loop duplicate discovery in the AtticPOT cultural repository as a Positive-Unlabeled (PU) learning task. Given one anchor image per artefact, it trains a lightweight per-query Clone Encoder solely on augmented views of that anchor and scores the unlabeled repository using an interpretable threshold on the latent l2 norm. Reported results are F1=96.37 (AUROC=97.97) on CIFAR-10 and F1=90.79 (AUROC=98.99) on AtticPOT, with a +7.70 F1 gain over SVDD under the same backbone; qualitative panels illustrate stable neighbourhoods across viewpoint and condition. The method avoids explicit negatives and targets de-duplication and record-linkage workflows.

Significance. If the central empirical claims hold, the work supplies a practical, negative-free, and transparent operating-point method for clone detection in cultural image collections. The reported gains over SVDD, the use of a single anchor per artefact, and the lightweight backbone are concrete strengths that could support curator-in-the-loop pipelines. The absence of explicit negatives and the interpretable threshold are genuine advantages over standard one-class baselines.

major comments (2)

[§3-4] §4 (Experiments) and §3 (Method): the latent l2-norm threshold selection procedure is not described, nor is any protocol for choosing the operating point on held-out data; without this, the reported F1=90.79 and AUROC=98.99 on AtticPOT cannot be verified as pre-specified rather than post-hoc tuned.
[§3] §3 (Clone Encoder training): the central assumption that standard augmentations of a single anchor produce a latent distribution covering real cross-record clone variations (viewpoint, lighting, degradation, occlusion) receives no quantitative check; no held-out verified duplicates are shown to lie inside the chosen l2 ball, directly affecting the reliability of the PU formulation and the +7.70 F1 gain claim.

minor comments (2)

[Abstract] Abstract: the phrase 'interpretable threshold' is used without any accompanying detail on its selection or sensitivity, which should be clarified for readers.
[§5] §5 (Qualitative results): the 'find-similar' panels would benefit from explicit annotation of which neighbours are verified duplicates versus false positives to strengthen the visual evidence.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below, providing clarifications on our experimental protocol and indicating where revisions will be made to improve transparency and justification.

read point-by-point responses

Referee: [§3-4] §4 (Experiments) and §3 (Method): the latent l2-norm threshold selection procedure is not described, nor is any protocol for choosing the operating point on held-out data; without this, the reported F1=90.79 and AUROC=98.99 on AtticPOT cannot be verified as pre-specified rather than post-hoc tuned.

Authors: We agree that an explicit description of the threshold selection is necessary for reproducibility and to confirm the operating point is pre-specified. In the experiments, the threshold was computed exclusively from the anchor: 50 augmented views of the single anchor were encoded, l2 distances to the anchor embedding were calculated, and the threshold was set at the 95th percentile of this distribution. No unlabeled repository samples or held-out test data were used in this step. We will add a dedicated paragraph in §3 describing this procedure in full, including the number of augmentations, the percentile rationale (to cover expected intra-artefact variation), and a reference in §4 confirming it precedes repository scoring. revision: yes
Referee: [§3] §3 (Clone Encoder training): the central assumption that standard augmentations of a single anchor produce a latent distribution covering real cross-record clone variations (viewpoint, lighting, degradation, occlusion) receives no quantitative check; no held-out verified duplicates are shown to lie inside the chosen l2 ball, directly affecting the reliability of the PU formulation and the +7.70 F1 gain claim.

Authors: We acknowledge the value of a quantitative check on coverage. However, the AtticPOT collection does not contain a sufficient set of independently verified cross-record duplicates held out from the anchors that could be used for such validation without introducing selection bias or leakage. The reported F1 and AUROC figures, as well as the +7.70 gain over SVDD, were obtained under an identical evaluation protocol for all methods, preserving relative fairness. The qualitative panels already illustrate that clones across viewpoint and condition fall inside the thresholded region. We will expand §3 with a more detailed justification of the augmentation policy (explicitly listing transformations targeting degradation and occlusion) and add any feasible small-scale quantitative support if additional verified pairs can be sourced without compromising the PU setup. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical PU method with held-out evaluation

full rationale

The paper formulates duplicate detection as a PU learning task and reports empirical F1/AUROC on held-out splits of CIFAR-10 and AtticPOT. No equations, derivations, or self-citations reduce the performance numbers to fitted inputs by construction; the Clone Encoder is trained on augmented anchor views and evaluated via an explicit l2-norm threshold on unlabeled data. The central claim remains an empirical outcome under stated assumptions rather than a self-definitional or fitted-input reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard PU-learning assumptions plus domain-specific choices for augmentation and thresholding; no new entities are postulated.

free parameters (1)

latent l2 norm threshold
Interpretable operating point used to score unlabeled items and propose candidates.

axioms (1)

domain assumption Augmented views of the single anchor represent the positive class distribution
Core premise enabling training of the Clone Encoder without explicit negatives.

pith-pipeline@v0.9.0 · 5491 in / 1249 out tokens · 61846 ms · 2026-05-13T17:33:16.116692+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

[1]

Iiif apis documentation

Europeana Foundation. Iiif apis documentation. https://europeana.atlassian.net/wiki/spaces/EF/ pages/1627914244/IIIF+APIs+Documentation, 2024. Last updated 2024-10-02; accessed 2025-09-06

work page arXiv 2024
[2]

Record api documentation

Europeana Foundation. Record api documentation. https://europeana.atlassian.net/wiki/spaces/EF/ pages/2385674279/Record+API+Documentation, 2024. Last updated 2024-10-02; accessed 2025-09-06

work page arXiv 2024
[3]

Search api documentation

Europeana Foundation. Search api documentation. https://europeana.atlassian.net/wiki/spaces/EF/ pages/2385739812/Search+API+Documentation, 2025. Last updated 2025-02-18; accessed 2025-09-06

work page arXiv 2025
[4]

Europeana apis (landing)

Europeana Foundation. Europeana apis (landing). https://apis.europeana.eu/, 2025. Accessed 2025-09- 06

work page 2025
[5]

Demystifying publishing to europeana: A practical workflow for content providers.Scientific Culture, 1(1):1–8, 2015

George Pavlidis and Vasileios Sevetlidis. Demystifying publishing to europeana: A practical workflow for content providers.Scientific Culture, 1(1):1–8, 2015

work page 2015
[6]

Api specifications — international image interoperability framework (iiif)

IIIF Consortium. Api specifications — international image interoperability framework (iiif). https://iiif.io/ api/, 2025. Accessed 2025-09-06

work page 2025
[7]

The atticpot project–attic po (ttery in) t (hrace)

Eirini Chioti, Amalia Avramidou, Despoina Tsiafaki, et al. The atticpot project–attic po (ttery in) t (hrace). Bulgarian e-Journal of Archaeology, 9:293–294, 2019

work page 2019
[8]

Contextualizing rare shapes of athenian kerameikos from coastal and inland thrace (6th–4th c

Yiannis Mourthos, Despoina Tsiafaki, et al. Contextualizing rare shapes of athenian kerameikos from coastal and inland thrace (6th–4th c. bc): an approach through the atticpot repository.Bulgarian e-Journal of Archaeology| Bblgarsko e-Spisanie za Arheologiya, 12(2):217–243, 2022

work page 2022
[9]

A decade survey of content based image retrieval using deep learning.IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4551–4570, 2021

Shiv Ram Dubey. A decade survey of content based image retrieval using deep learning.IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4551–4570, 2021

work page 2021
[10]

Fundamentals of content-based image retrieval

Fuhui Long, Hongjiang Zhang, and David Dagan Feng. Fundamentals of content-based image retrieval. In Multimedia Information Retrieval and Management: Technological Fundamentals and Applications, pages 1–26. Springer, 2003

work page 2003
[11]

van Gemert

Thomas Mensink and Jan C. van Gemert. The rijksmuseum challenge: Museum-centered visual recognition. In Proceedings of the ACM International Conference on Multimedia Retrieval (ICMR), pages 451–454, 2014

work page 2014
[12]

Omniart: Multi-task deep learning for artistic data analysis.arXiv preprint, 2017

Gjorgji Strezoski and Marcel Worring. Omniart: Multi-task deep learning for artistic data analysis.arXiv preprint, 2017

work page 2017
[13]

How to read paintings: Semantic art understanding with multi-modal retrieval

Noa Garcia and George V ogiatzis. How to read paintings: Semantic art understanding with multi-modal retrieval. InECCV Workshops (ECCVW) 2018, LNCS 11130, pages 676–691, 2019

work page 2018
[14]

An image similarity search for the european digital library and beyond

Sergiu Gordea. An image similarity search for the european digital library and beyond. In2nd International Workshop on Supporting Users Exploration of Digital Libraries (SUEDL 2013), 2013

work page 2013
[15]

Linked open images: Visual similarity for the semantic web.Semantic Web, 14(2):197–208, 2023

Lukas Kli ´c. Linked open images: Visual similarity for the semantic web.Semantic Web, 14(2):197–208, 2023

work page 2023
[16]

Art3mis: ray-based textual annotation on 3d cultural objects

Vasileios Arampatzakis, Vasileios Sevetlidis, Fotis Arnaoutoglou, Athanasios Kalogeras, Christos Koulamas, Aris Lalos, Chairi Kiourt, George Ioannakis, Anestis Koutsoudis, and George Pavlidis. Art3mis: ray-based textual annotation on 3d cultural objects. InCAA 2021 International Conference “Digital Crossroads, 2021

work page 2021
[17]

Phaser: Perceptual hashing algorithms evaluation and results — an open-source forensic framework.Forensic Science International: Digital Investigation, 2024

Sean McKeown, Peter Aaby, and Andreas Steyven. Phaser: Perceptual hashing algorithms evaluation and results — an open-source forensic framework.Forensic Science International: Digital Investigation, 2024

work page 2024
[18]

A survey of perceptual hashing for multimedia.ACM Computing Surveys, 2025

Haitao Wang, Xinpeng Zhang, Feng Li, Jiwu Huang, and Yun-Qing Shi. A survey of perceptual hashing for multimedia.ACM Computing Surveys, 2025

work page 2025
[19]

An overview of perceptual hashing.Journal of Online Trust and Safety, 1(1), 2021

Hany Farid. An overview of perceptual hashing.Journal of Online Trust and Safety, 1(1), 2021. 9 Arampatzakis et al

work page 2021
[20]

Ceramic fabric classification of petrographic thin sections with deep learning.Journal of computer applications in archaeology, 4(1), 2021

Mike Lyons. Ceramic fabric classification of petrographic thin sections with deep learning.Journal of computer applications in archaeology, 4(1), 2021

work page 2021
[21]

Classification of engraved pottery sherds mixing deep-learning features by compact bilinear pooling.Pattern Recognition Letters, 131:1–7, 2020

Aladine Chetouani, Sylvie Treuillet, Matthieu Exbrayat, and Sébastien Jesset. Classification of engraved pottery sherds mixing deep-learning features by compact bilinear pooling.Pattern Recognition Letters, 131:1–7, 2020

work page 2020
[22]

The automatic recognition of ceramics from only one photo: The archaide app.Journal of Archaeological Science: Reports, 36:102788, 2021

Francesca Anichini, Nachum Dershowitz, Nevio Dubbini, Gabriele Gattiglia, Barak Itkin, and Lior Wolf. The automatic recognition of ceramics from only one photo: The archaide app.Journal of Archaeological Science: Reports, 36:102788, 2021

work page 2021
[23]

Facsimiles-based deep learning for matching relief-printed decorations on medieval ceramic sherds

Khawla Brahim, Sylvie Treuillet, Matthieu Exbrayat, and Sebastien Jesset. Facsimiles-based deep learning for matching relief-printed decorations on medieval ceramic sherds. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1613–1622, 2023

work page 2023
[24]

Crowdheritage: Large-scale crowdsourcing for enriching europeana metadata.Information, 12(9):384, 2021

Eirini Kaldeli and et al. Crowdheritage: Large-scale crowdsourcing for enriching europeana metadata.Information, 12(9):384, 2021

work page 2021
[25]

Crowdsourcing in the cultural heritage domain: Opportunities and challenges

Joris Oomen and Lora Aroyo. Crowdsourcing in the cultural heritage domain: Opportunities and challenges. In Proceedings of the 5th International Conference on Communities & Technologies (Workshops), 2011

work page 2011
[26]

Crowdsourcing our cultural heritage: Introduction

Mia Ridge. Crowdsourcing our cultural heritage: Introduction. InCrowdsourcing our cultural heritage, pages 1–14. Routledge, 2016

work page 2016
[27]

From tagging to theorizing: deepening engagement with cultural heritage through crowdsourcing

Mia Ridge. From tagging to theorizing: deepening engagement with cultural heritage through crowdsourcing. Curator: The Museum Journal, 56(4):435–450, 2013

work page 2013
[28]

Individual vs

Samantha Blickhan, Coleman Krawczyk, Daniel Hanson, Amy Boyer, Andrea Simenstad, and Victoria Van Hyning. Individual vs. collaborative methods of crowdsourced transcription.Journal of Data Mining & Digital Humanities, 2019

work page 2019
[29]

International winners 2022 (announcement & report).Wiki Loves Monuments (official site/blog), 2022

Wiki Loves Monuments. International winners 2022 (announcement & report).Wiki Loves Monuments (official site/blog), 2022. Accessed 2025-09-06

work page 2022
[30]

Wikiartvectors: Style and color representations of artworks for cultural analysis via information theoretic measures.Entropy, 24(9):1175, 2022

Bhargav Srinivasa Desikan, Hajime Shimao, and Helena Miton. Wikiartvectors: Style and color representations of artworks for cultural analysis via information theoretic measures.Entropy, 24(9):1175, 2022

work page 2022
[31]

Augmenting existing food image datasets with greek dishes.Big Data in Archaeology, page 133, 2021

V Sevetlidis, C Kiourt, C Tzouvara, G Tastzoglou, and G Pavlidis. Augmenting existing food image datasets with greek dishes.Big Data in Archaeology, page 133, 2021

work page 2021
[32]

Ai in gastronomic tourism

G Pavlidis, SL Markantonatou, KR Toraki, AN Vacalopoulou, CL Strouthopoulos, DT Varsamis, A V Tsimpiris, SR Mouroutsos, CR Kiourt, VL Sevetlidis, et al. Ai in gastronomic tourism. InProceedings of the 2nd International Conference on Advances In Signal Processing and Artificial Intelligence, pages 168–174, 2020

work page 2020
[33]

Mouroutsos, and Antonios Gasteratos

Vasileios Sevetlidis, George Pavlidis, Vasileios Arampatzakis, Chairi Kiourt, Spyridon G. Mouroutsos, and Antonios Gasteratos. Web acquired image datasets need curation: An examplar pipeline evaluated on greek food images. In2021 IEEE International Conference on Imaging Systems and Techniques (IST), pages 1–6, 2021

work page 2021
[34]

Mouroutsos, and Antonios Gasteratos

Vasileios Sevetlidis, George Pavlidis, Spyridon G. Mouroutsos, and Antonios Gasteratos. Tackling dataset bias with an automated collection of real-world samples.IEEE Access, 10:126832–126844, 2022

work page 2022
[35]

Positive-unlabeled learning with non-negative risk estimator.Advances in neural information processing systems, 30, 2017

Ryuichi Kiryo, Gang Niu, Marthinus C Du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with non-negative risk estimator.Advances in neural information processing systems, 30, 2017

work page 2017
[36]

Leveraging positive- unlabeled learning for enhanced black spot accident identification on greek road networks.Computers, 13(2):49, 2024

Vasileios Sevetlidis, George Pavlidis, Spyridon G Mouroutsos, and Antonios Gasteratos. Leveraging positive- unlabeled learning for enhanced black spot accident identification on greek road networks.Computers, 13(2):49, 2024

work page 2024
[37]

Dense-pu: Learning a density-based boundary for positive and unlabeled learning.IEEE Access, 12:90287–90298, 2024

Vasileios Sevetlidis, George Pavlidis, Spyridon G Mouroutsos, and Antonios Gasteratos. Dense-pu: Learning a density-based boundary for positive and unlabeled learning.IEEE Access, 12:90287–90298, 2024

work page 2024
[38]

A simple framework for contrastive learning of visual representations, 2020

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020

work page 2020
[39]

Momentum contrast for unsupervised visual representation learning, 2020

Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning, 2020

work page 2020
[40]

Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko

Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning, 2020

work page 2020
[41]

Deep one-class classification

Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. InInternational conference on machine learning, pages 4393–4402. PMLR, 2018. 10

work page 2018