Pith. sign in

REVIEW 4 major objections 5 minor 43 references

A Machine Learning-based Anomaly Detection Framework in Life Insurance Contracts

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read An ensemble of autoencoders detects all four injected anomalies in two insurance-like contract datasets, with moderate false positives.

desk verdict The headline claim that autoencoder ensembles are more accurate is plausible but not established: the 4/4 detection counts come from parameters and thresholds tuned on the injected anomalies, at uncontrolled false-positive rates. read the letter →

arxiv 2411.17495 v1 pith:DUKSRJVP submitted 2024-11-26 stat.AP cs.LG

Andreas Groll , Akshat Khanna , Leonid Zeldin This is my paper · ORCID
classification stat.APcs.LG MSC 62H3062P05
keywords anomalydetectionlifeinsurancecontractsautoencoderensemblevariationalIsolationForestunsupervisedlearningreconstructionerrorcontract-levelanomalies
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 tries to establish that an ensemble of autoencoders is a practical unsupervised tool for finding anomalous life insurance-style contracts where labeled examples are unavailable. It benchmarks classical proximity-based and tree-based detectors against autoencoders and variational autoencoders on two open-source health insurance datasets, each augmented with four manually planted contract-level anomalies. Its central claim is that the autoencoder ensemble detects all four anomalies in both datasets while flagging a moderate share of normal contracts, whereas the classical methods miss anomalies, take impractically long, or flag far too many points. If the claim is right, insurers can screen large contract databases without labels by averaging per-contract reconstruction errors and reviewing only the contracts above a threshold.

What carries the argument

The load-bearing mechanism is the ensemble autoencoder driven by reconstruction error. Each autoencoder takes a standardized, one-hot encoded contract vector, compresses it through a low-dimensional bottleneck, and reconstructs it using mean squared error loss; a network trained only on unlabeled contracts learns to reproduce normal feature combinations and produces high error on atypical contracts. Three architectures are trained so that no single choice of layer sizes decides the outcome, and their per-point reconstruction errors are averaged into one score. A threshold on that averaged score (0.5 here, and 0.7 for the variational autoencoder on the larger dataset) separates normal contracts from anomalies. The paper attributes the method's stability across the two datasets to this averaging step.

What would settle it

Run the same three-model autoencoder ensemble on both cleaned datasets with no injected anomalies and count how many ordinary contracts exceed the chosen threshold; if the false-positive count approaches the reported 224 and 705, the method is responding to dataset structure rather than to genuine anomalies.

Watch

Extended reading notes

Core claim

The discovery the paper argues for is that an ensemble of autoencoders, built from three feed-forward networks with deliberately varied hidden-layer and latent sizes, detects every one of the four manually inserted contract-level anomalies in both datasets. The ensemble's reconstruction errors are averaged per point, and a fixed threshold marks a contract as anomalous; on the smaller dataset (990 rows after insertion) it flags 224 points, and on the larger dataset (24,010 rows) it flags 705. The variational autoencoder ensemble also finds all four anomalies, but it flags more than half of the smaller dataset (512 of 990 points), so the paper identifies the plain autoencoder ensemble as the stronger result. Against this, nearest-neighbor, k-means, DBSCAN, HDBSCAN, and one-class SVM all miss at least one anomaly on the smaller set or fail completely on the larger set, with DBSCAN and HDBSCAN not finishing within five hours and k-means detecting none of the four anomalies. Isolation Forest finds all four on both datasets but labels 342 of 990 and 1,974 of 24,010 points as anomalous, which the paper treats as a reliability concern.

Load-bearing premise

The load-bearing premise is that the four manually inserted anomalies, designed as atypical feature combinations, represent the real anomalies a life insurer would need to find; the entire ranking of methods rests on detecting exactly these four points.

Editorial extensions

If this is right

  • If the central claim is correct, insurers can use reconstruction-error ensembles as an unsupervised first-pass screen, sending only the flagged contracts to human review.
  • Classical density-based methods such as DBSCAN and HDBSCAN should be deprioritized for large contract portfolios, since they did not finish on the larger dataset within the allotted five hours.
  • Isolation Forest remains the fastest classical alternative, but its low Anomaly Score and high flag counts mean its output needs additional vetting before use.
  • The paper's own proposed extensions—larger ensembles, hybrid classical-plus-deep pipelines, and GPU-based training—are direct ways to test whether the result scales beyond these two datasets.

Reading between the lines

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

  • Editorial extension: the four injected anomalies are the only measure of success, so the reported comparison does not say how the method would behave with real, unlabeled fraud; a separate false-positive calibration on completely normal data is needed.
  • Editorial extension: nothing in the pipeline uses life-insurance-specific information, so the same ensemble should transfer to health, property, or other tabular contract portfolios, though the authors do not claim this.
  • Editorial extension: a temporal holdout design—training on past contracts, scoring later ones, and checking whether confirmed future fraud cases rank high—would be a stronger validation than injected anomalies and would also mimic how an insurer would deploy the system.
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

4 major / 5 minor

Summary. The manuscript proposes an unsupervised anomaly detection framework for life insurance contracts, comparing six classical methods (Nearest Neighbors, k-means, DBSCAN, HDBSCAN, OCSVM, Isolation Forest) with two deep learning approaches (autoencoder and variational autoencoder ensembles) on two health-insurance-derived datasets. Four artificial contract-level anomalies are manually injected into each dataset, and methods are compared by how many of these four points are detected, the total number of points flagged as anomalous, runtime, and an internal score (Silhouette or Anomaly Score). The authors report that the autoencoder ensemble detects all four injected anomalies in both datasets while keeping the false-positive rate 'moderate' (except VAE on Dataset 1), and conclude that an ensemble of autoencoders provides higher accuracy for anomaly detection.

Significance. If the central claim were established, the paper would offer a practical, comparatively evaluated recipe for unsupervised contract-level anomaly screening in insurance, with the useful observation that autoencoder ensembles scale better than DBSCAN/HDBSCAN on larger data. The paper also attempts to automate hyperparameter tuning for several classical methods, which is a worthwhile goal for non-specialist adoption. However, the current evaluation design—label-informed tuning on the injected anomalies, no threshold-selection rule for the autoencoders, and comparisons made at uncontrolled operating points—means the headline 'higher accuracy' claim is not yet supported. The manuscript provides no code or data, which limits reproducibility.

major comments (4)
  1. [Sections 5.2, 5.3, Appendix A.2.1, Table 9] The comparison of 'anomalies detected out of 4' is invalid as an accuracy comparison because several methods were explicitly tuned on the four injected anomalies. Appendix A.2.1 states that NN's k is chosen by 'selecting the model that detects the most of the four manually placed anomalies' and that OCSVM's ν is found experimentally in the same way; HDBSCAN parameters are 'manually comparing the results'; Section 5.2 says 'the remaining methods were manually adjusted to detect as many of the four manually inserted anomalies as possible.' For AE/VAE, Table 9 lists thresholds (0.5, 0.5, 0.5, 0.7) without any described selection rule, and Section 5.3 says only that a threshold t is chosen. Any method with a continuous anomaly score can achieve 4/4 by lowering its threshold. The reported 4/4 results are therefore in-sample, label-informed operating points, not evidence of higher accuracy. A threshold-independent comparison (AUC or PR-AUC) or a fixed false-positive-rate comparison is required.
  2. [Section 3] The four manually inserted anomalies are not described beyond the statement that they were 'designed with consideration for the overall structural characteristics of the data' and that they represent 'unique or atypical combinations of features.' No actual values, construction procedure, or generating code is given. Since every comparative conclusion in the paper rests on whether these four points are detected, the lack of a concrete, reproducible description makes it impossible to assess whether the task is meaningful, trivial, or biased toward certain methods. The authors should provide the full specification of these anomalies (e.g., which variables were manipulated and how) and ideally make the datasets available.
  3. [Tables 2 and 3, Sections 5.3 and 5.4] The paper's conclusion that the autoencoder ensemble 'provides higher accuracy' is not supported by the reported operating points. On Dataset 1, the AE flags 224/990 (22.6%) of points as anomalies while NN flags 14/990; both detect some of the four injected anomalies (AE 4, NN 3). Without a stated cost model for false positives versus false negatives, one cannot conclude that AE is more accurate—it might simply be more sensitive at a much higher false-positive rate. The discussion in Section 5.4 acknowledges the high flagging rates but does not correct for this; the phrase 'higher accuracy' is used without any statistical or decision-theoretic justification. At minimum, the authors should report precision-recall curves or set a common FPR (e.g., 1% or 5%) across methods and then compare detection rates.
  4. [Section 4.2 and Table 6] The interpretation of the Isolation Forest's 'Anomaly Score' is internally inconsistent and obscures the comparison. Section 4.2 says the score ranges from −1 to 1, with scores near −1 indicating anomalies; Section 5.2 reports an average Anomaly Score of 0.48/0.46 and calls it 'low,' but 0.48 is near the middle of the stated range and would not be considered low under the paper's own description. It is also unclear how this aggregate score is computed (average over points, over trees, or the sklearn decision_function?). The authors should clarify the score's meaning and why it is interpreted as evidence of unreliability.
minor comments (5)
  1. [Throughout] The manuscript contains frequent typos and grammatical errors, e.g., 'ensamble' (Section 5.3), 'challange', 'acceptabe', 'detecion', 'runnung' (Table 1), 'aproximatly', 'descision', and the footnote 'RDF' should be 'RBF' in Table 6. A careful proofread is needed.
  2. [Section 5.1 and Table 7] Section 5.1 reports that adding BMI increases Dataset 1's column count from 12 to 13, but Table 7 gives the autoencoder input layer as 12 for Dataset 1. Please clarify whether the ID column is dropped, whether BMI is actually used as a feature, and why the input dimension is 12.
  3. [Figure 2] The figure caption mentions a 'secondary y-axis' but the figure as described appears to show bars (proportion of anomalies) and lines (number of manual anomalies detected) with unclear scales. Please improve the caption and axis labels so the reader can interpret the visual comparison.
  4. [Section 3] The choice of health insurance datasets as proxies for life insurance is asserted but not justified with reference to specific structural similarities. A brief discussion of how contract-level anomalies in health insurance translate to life insurance would help readers assess external validity.
  5. [Section 4.1, k-means] The description says the anomaly threshold is 'based on the average distance to the centroid, adjusted by the standard deviation,' but no exact criterion is given. Please state the precise rule used in the implementation.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline 'anomalies detected out of 4' comparison is partly in-sample: several classical methods are explicitly tuned to detect the four injected anomalies, and the AE/VAE thresholds are free parameters with no stated selection rule, so the reported 4/4 results are uncontrolled operating points rather than independent predictions.

  1. fitted input called prediction [Section 5.2; Appendix A.2.1 (Nearest-Neighbors, OCSVM, HDBSCAN)]
    "The remaining methods were manually adjusted to detect as many of the four manually inserted anomalies as possible. ... No scoring mechanism⇒ tuningk by selecting the model that detects the most of the four manually placed anomalies within the dataset."

    The outcome reported as accuracy ('anomaly detected out of 4', Table 1) is the same objective used to select hyperparameters for NN, OCSVM, and HDBSCAN. Appendix A.2.1 confirms that NN's k=3 and OCSVM's nu=0.1 were chosen because they detected the injected anomalies on Dataset 1, and HDBSCAN's min cluster size was found by manually comparing results. Reporting these tuned counts as comparative performance is therefore in-sample: the count is maximized by construction on the tuning data, so it is a selection criterion, not an independent prediction of detection ability.

  2. other [Section 5.3; Table 9 (Autoencoder and Variational Autoencoder thresholds)]
    "Finally, there is an average reconstruction error for each point. From now on we can choose a threshold t above which all the reconstruction errors indicate the presence of anomalies. ... Table 9: Threshold 0.5 0.5 0.5 0.7."

    The AE/VAE 4/4 detection result is defined by thresholding continuous reconstruction errors at t, yet the paper gives no rule for selecting t. Because t is a free parameter and performance is reported only as the count of injected anomalies above t, the result is an uncontrolled operating point rather than a prediction: lowering t increases the count by construction. Without a fixed false-positive rate, a matched threshold-selection rule, or a threshold-independent metric such as AUC, '4 out of 4' does not by itself establish higher accuracy.

full rationale

The paper's central accuracy claim is the number of manually inserted anomalies detected in each dataset (Tables 1 and 2). That metric is partially circular because Section 5.2 states that 'the remaining methods were manually adjusted to detect as many of the four manually inserted anomalies as possible,' and Appendix A.2.1 confirms that NN's k and OCSVM's nu were selected by detecting those four points on Dataset 1. Thus the Dataset 1 detection counts for these methods are fitted values, not predictive results. For autoencoders and variational autoencoders, the threshold t is introduced without any stated selection rule; Table 9 lists thresholds (0.5, 0.5, 0.5, 0.7) but not how they were chosen. Since any continuous-score method can achieve 4/4 by lowering its threshold, the AE/VAE counts are not identified as predictions unless the threshold is fixed independently and false-positive rates are matched. The comparison is further weakened by the very different total anomaly counts in Table 3 (e.g., AE flags 224/990 on Dataset 1 while NN flags 14/990), so the 'higher accuracy' conclusion is not established without a matched-FPR or threshold-independent metric. No load-bearing self-citation chain or imported uniqueness theorem is present; the circularity is confined to the evaluation protocol. Score 6 reflects that the headline detection-count comparisons reduce, at least in part, to in-sample tuning and free thresholds, while the method implementations and runtime comparisons retain independent content.

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

The paper relies on domain assumptions about dataset suitability and synthetic anomaly representativeness, plus standard assumptions for unsupervised anomaly scoring. No new entities are introduced.

free parameters (6)
  • k (Nearest Neighbors) = 3
    Selected for Dataset 1 by detecting the most of the four injected anomalies, then applied to Dataset 2 (Appendix A.2.1).
  • k (k-means) = 19 (Dataset 1), 2 (Dataset 2)
    Selected via grid search optimizing Silhouette Score (Section 5.2).
  • epsilon (DBSCAN) = 1000
    Tuned via grid search using kneeLocator and Silhouette Score (Appendix A.2.1).
  • nu (OCSVM) = 0.1
    Best model on Dataset 1 found experimentally, then applied to Dataset 2 (Appendix A.2.1).
  • Threshold t (AE/VAE) = 0.5 (AE), 0.5/0.7 (VAE)
    Selection procedure not stated; appears chosen to flag the injected anomalies (Section 5.3, Table 9).
  • Number of ensemble models M = 3
    Chosen arbitrarily as 'ensamble learning' with three architectures; no justification for the number (Section 5.3).
assumptions (4)
  • domain assumption Health insurance datasets are sufficiently similar to life insurance contracts for the anomaly detection comparison.
    Section 3 states the datasets are 'similar enough to allow for comparable analysis'.
  • domain assumption The four manually inserted anomalies are representative of real-world anomalies in insurance contracts.
    Section 3 claims they 'simulate complex relationships found in real-world data'.
  • standard math Reconstruction error is a valid anomaly score for autoencoders.
    Section 4.3 assumes low reconstruction error for normal points and high error for anomalies.
  • standard math Silhouette Score and Anomaly Score are valid unsupervised model-selection criteria.
    Used for grid search in Section 5.2 to automate parameter tuning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Machine Learning-based Anomaly Detection Framework in Life Insurance Contracts." pith.science (2026). https://pith.science/paper/DUKSRJVP

@misc{pith2026241117495,
  author       = {Pith},
  title        = {Pith review of: A Machine Learning-based Anomaly Detection Framework in Life Insurance Contracts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUKSRJVP}},
  note         = {Machine review of arXiv:2411.17495}
}
read the original abstract

Life insurance, like other forms of insurance, relies heavily on large volumes of data. The business model is based on an exchange where companies receive payments in return for the promise to provide coverage in case of an accident. Thus, trust in the integrity of the data stored in databases is crucial. One method to ensure data reliability is the automatic detection of anomalies. While this approach is highly useful, it is also challenging due to the scarcity of labeled data that distinguish between normal and anomalous contracts or inter\-actions. This manuscript discusses several classical and modern unsupervised anomaly detection methods and compares their performance across two different datasets. In order to facilitate the adoption of these methods by companies, this work also explores ways to automate the process, making it accessible even to non-data scientists.

Figures

Figures reproduced from arXiv: 2411.17495 by the authors.

Figure 1
Figure 1. An example representation of anomalous and normal points in an [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Proportion of anomalies and detected manual anomalies (secondary [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages

  1. [1]

    Aggarwal, C. C. (2017). Outlier analysis

    work page 2017

  2. [2]

    (2020, August)

    Arvai, K. (2020, August). kneed

    work page 2020

  3. [3]

    Bauder, R. A. and T. M. Khoshgoftaar (2017). Medicare fraud detection using machine learning methods

    work page 2017

  4. [4]

    Wiestler, S

    Baur, C., B. Wiestler, S. Albarqouni, and N. Navab (2019). Deep autoencoding models for unsupervised anomaly segmentation in brain mr images

    work page 2019

  5. [5]

    Giordano, and M

    Callegari, C., S. Giordano, and M. Pagano (2014). Neural network based anomaly detection

    work page 2014

  6. [6]

    Campello, R. J., D. Moulavi, and J. Sander (2021). Advances in knowledge discovery and data mining -- density-based clustering based on hierarchical density estimates. pp.\ 160 -- 172

    work page 2021

  7. [7]

    Campello, R. J. G. B., D. Moulavi, and J. Sander (2013). Density-based clustering based on hierarchical density estimates. pp.\ 160--172

    work page 2013

  8. [8]

    Chalapathy, R., A. K. Menon, and S. Chawla (2018). Anomaly detection using one-class neural networks. CoRR\/ abs/1802.06360

    arXiv 2018

Show all 43 references
  1. [9]

    Banerjee, and V

    Chandola, V., A. Banerjee, and V. Kumar (2009, Jul.). Anomaly detection: A survey. ACM Comput. Surv.\/ 41

    2009

  2. [10]

    Chauhan, S. and L. Vig (2015). Anomaly detection in ecg time signals via deep long short-term memory networks

    2015

  3. [11]

    Pawlowski, M

    Chen, X., N. Pawlowski, M. Rajchl, B. Glocker, and E. Konukoglu (2018). Deep generative models in the real-world: An open challenge from medical imaging. arXiv preprint arXiv:1806.05452\/

    2018 arXiv

  4. [12]

    Cortes, C. and V. N. Vapnik (1995). Support-vector networks. Machine Learning\/ 20 , 273--297

    1995

  5. [13]

    (2024, May)

    da Costa-Luis, C. (2024, May). tqdm: A fast, Extensible Progress Bar for Python and CLI

    2024

  6. [14]

    Edgeworth, F. (1887). Xli. on discordant observations. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science\/ 23

  7. [15]

    Kriegel, and X

    Ester, M., H.-P. Kriegel, and X. X. Jörg Sander (1996). A density-based algorithm for discovering clusters in large spatial databases with noise

    1996

  8. [16]

    Crowley, F

    Ghojogh, B., M. Crowley, F. Karray, and A. Ghodsi (2023). Variational autoencoders. pp.\ 563--576

    2023

  9. [17]

    Jin, and H

    Gomes, C., Z. Jin, and H. Yang (2021, Jul.). Insurance fraud detection with unsupervised deep learning. Journal of Risk and Insurance\/

    2021

  10. [18]

    Harris, C. R. and K. J. Millman (2020). Numpy: A fundamental package for scientific computing with python

    2020

  11. [19]

    Liao, and R

    Hu, W., Y. Liao, and R. V. Vemuri (2003, Jun.). Robust anomaly detection using support vector machines. Proceedings of the International Conference on Machine Learning\/

    2003

  12. [20]

    Patel, P

    Hudekar, P., D. Patel, P. More, M. Dongare, and S. Shirsath (2023). Implementation of anomaly detection using unsupervised machine learning

    2023

  13. [21]

    Hunter, J. D. (2007). Matplotlib: A 2d graphics environment

    2007

  14. [22]

    Witten, T

    James, G., D. Witten, T. Hastie, and R. Tibshirani (2013). An introduction to statistical learning: with applications in r

    2013

  15. [23]

    Liu, F. T., K. M. Ting, and Z.-H. Zhou (2008). Isolation forest

    2008

  16. [24]

    Ramakrishnan, G

    Malhotra, P., A. Ramakrishnan, G. Anand, L. Vig, P. Agarwal, and G. Shroff (2016). Lstm-based encoder-decoder for multi-sensor anomaly detection. arXiv preprint arXiv:1607.00148\/

    2016 arXiv

  17. [25]

    Meier, D

    Masci, J., U. Meier, D. Cire s an, and J. Schmidhuber (2011). Stacked convolutional auto-encoders for hierarchical feature extraction. pp.\ 52--59

    2011

  18. [26]

    McKinney, W. (2021). Pandas: Powerful python data analysis toolkit

    2021

  19. [27]

    Menze, B. H., A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, and J. Kirby (2015). The multimodal brain tumor image segmentation benchmark ( BRATS )

    2015

  20. [28]

    Michelucci, U. (2022). An introduction to autoencoders. CoRR\/ abs/2201.03898

    2022 arXiv

  21. [29]

    Gross, and F

    Paszke, A., S. Gross, and F. Massa (2019). Pytorch: An imperative style, high-performance deep learning library

    2019

  22. [30]

    Pedregosa, F. and G. Varoquaux (2011). scikit-learn: Machine learning in python

    2011

  23. [31]

    Prasad, N. R., S. Almanza-Garcia, and T. T. Lu (2009). Anomaly detection. Cmc -Tech Science Press-\/ 14

    2009

  24. [32]

    Python Software Foundation (2023). Python

    2023

  25. [33]

    Reddy, G. D. (2023). Health insurance dataset 2. https://www.kaggle.com/datasets/gdeepakreddy/insurance. [Accessed 01-03-2024]

    2023

  26. [34]

    (2024, 01)

    Samuels, J. (2024, 01). One-hot encoding and two-hot encoding: An introduction

    2024

  27. [35]

    Schölkopf, B., J. C. Platt, J. Shawe-Taylor, A. J. Smola, and R. C. Williamson (2001, 07). Estimating the Support of a High-Dimensional Distribution . Neural Computation\/ 13\/ (7), 1443--1471

    2001

  28. [36]

    Schölkopf, B., R. C. Williamson, A. Smola, J. Shawe-Taylor, and J. Platt (1999). Support vector method for novelty detection. 12

    1999

  29. [37]

    Mansimov, and R

    Srivastava, N., E. Mansimov, and R. Salakhudinov (2015, 07--09 Jul). Unsupervised learning of video representations using lstms. 37 , 843--852

    2015

  30. [38]

    Sun, J., X. Wang, N. Xiong, and J. Shao (2018). Learning sparse representation with variational auto-encoder for anomaly detection. IEEE Access\/ 6

    2018

  31. [39]

    Medical insurance premium prediction

    Tejashvi (2023). Medical insurance premium prediction. https://www.kaggle.com/datasets/tejashvi14/medical-insurance-premium-prediction. [Accessed 01-03-2024]

    2023

  32. [40]

    Franco-Penya, J

    Wang, F., H.-H. Franco-Penya, J. D. Kelleher, J. Pugh, and R. Ross (2017). An analysis of the application of simplified silhouette to the evaluation of k-means clustering validity

    2017

  33. [41]

    Yao, R., C. Liu, L. Zhang, and P. Peng (2019). Unsupervised anomaly detection using variational auto-encoder based feature extraction

    2019

  34. [42]

    Chen, and Y

    Zhao, M., J. Chen, and Y. Li (2018). A review of anomaly detection techniques based on nearest neighbor

    2018

  35. [43]

    Zimmerer, D., S. A. A. Kohl, J. Petersen, F. Isensee, and K. H. Maier-Hein (2018). Context-encoding variational autoencoder for unsupervised anomaly detection. arXiv preprint arXiv:1812.05941\/

    2018 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.