Encoding of Demographic and Anatomical Information in Chest X-Ray-based Severe Left Ventricular Hypertrophy Classifiers

Basudha Pal; Muhammad Umair; Rama Chellappa

REVIEW 3 major objections 6 minor 15 references

Encoding of Demographic and Anatomical Information in Chest X-Ray-based Severe Left Ventricular Hypertrophy Classifiers

T0 review · 3 major / 6 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 Severe left ventricular hypertrophy can be classified directly from chest X-rays by end-to-end deep networks, with the best model reaching AUROC 0.816 and AUPRC 0.803 on a balanced test set.

desk verdict A reasonable direct-classification study whose headline AUPRC comparison is invalid because the test set was rebalanced to 50% prevalence while the baseline was not; worth a major revision, not a desk reject. read the letter →

arxiv 2506.03192 v1 pith:UFKYNVVH submitted 2025-05-31 eess.IV cs.AIcs.CV

Basudha Pal , Rama Chellappa , Muhammad Umair This is my paper

classification eess.IVcs.AIcs.CV

keywords severeleftventricularhypertrophychestX-rayclassificationdeeplearningVisionTransformermutualinformationneuralestimationmodelinterpretabilityclass-imbalancedevaluationattributeencoding

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 proposes that severe left ventricular hypertrophy (SLVH) can be detected directly from chest X-rays by end-to-end deep classifiers, without first estimating echocardiographic measurements or supplying patient demographics. On a class-balanced subset of the CheXchoNet dataset, a vision transformer reaches AUROC 0.816 and AUPRC 0.803, exceeding the AUROC 0.79 and far exceeding the AUPRC 0.19 of the prior regression-plus-thresholding pipeline. The paper also uses mutual-information neural estimation to show that these classifiers, trained only on binary SLVH labels, internally encode clinically relevant attributes in a consistent order: age, sex, septal wall thickness, posterior wall thickness, and, weakly, ventricular internal diameter. If correct, this would make SLVH screening feasible from a cheap, widely available imaging modality while avoiding the error cascades and demographic confounding of multi-stage pipelines.

What carries the argument

The machinery has two parts. The classification framework is a direct end-to-end pipeline: a pretrained ResNet-18 or ViT backbone takes a 256 by 256 chest X-ray and outputs a binary SLVH prediction, with the ViT encoder frozen and only its head trained. The interpretability machinery is MINE, a neural estimator of mutual information I(F;A) between internal feature vectors F and an attribute A, built on the Donsker-Varadhan lower bound; the resulting per-layer expressivity score measures how much demographic or anatomical information is entangled in the learned representation. A third component is the construction of a class-balanced train, validation, and test split from CheXchoNet, sampling equal numbers of positive and negative CXRs while keeping the original split proportions.

What would settle it

Take the trained ViT and ResNet-18 models and evaluate them on the original imbalanced CheXchoNet test split (or an external CXR-echo cohort), computing AUROC and AUPRC. If the ViT's AUPRC falls toward the baseline's 0.19 or its AUROC drops below 0.79, the direct-classification advantage shown on the balanced subset does not transfer to natural prevalence.

Watch

Extended reading notes

Core claim

The central claim is that SLVH can be classified from chest X-rays directly, using only image input and binary outcome labels, and that the resulting models are both accurate and clinically interpretable. The best-performing model, a vision transformer pretrained with masked autoencoding, achieves AUROC 0.816 [0.781, 0.850] and AUPRC 0.803 [0.755, 0.849] on a balanced test set, while a ResNet-18 achieves AUROC 0.760 [0.718, 0.802] and AUPRC 0.731 [0.669, 0.786]; the baseline that regresses anatomical measurements and then thresholds reaches AUROC 0.79 but AUPRC only 0.19. The paper further claims that expressivity analysis via MINE reveals an attribute-encoding hierarchy of age greater than sex greater than IVSDd greater than LVPWDd greater than LVIDd across both architectures, consistent with clinical reasoning, even though age, sex, and anatomical measurements were never provided as inputs.

Load-bearing premise

The reported AUROC and AUPRC are computed on a class-balanced test set containing 50 percent SLVH-positive cases, and the claim depends on that balanced set being representative of the real chest-X-ray population where severe hypertrophy is rare.

Editorial extensions

If this is right

If the ViT result holds on external data, SLVH can be flagged from chest X-rays at the point of care without echocardiography, widening screening to populations that cannot access ultrasound or MRI.
The direct classifier's AUPRC advantage over the regression baseline means it should retain useful precision in imbalanced real-world populations, where the old pipeline's very low AUPRC made positive predictions hard to trust.
The expressivity ordering implies the network learns hypertrophy-relevant wall thickness and demographic context without being told to, so the model's decisions can be audited layer by layer rather than treated as a black box.
Because age and sex show high expressivity in late layers, the framework provides a quantitative target for fairness interventions: one can measure how much demographic information remains after suppression and check whether classification accuracy survives.

Reading between the lines

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

A natural next experiment is to suppress age and sex information in the learned features and measure whether the expressivity hierarchy flattens while classification accuracy holds; the paper's MINE scores make that experiment directly actionable.
The same MINE-based expressivity audit could be applied to other cardiac-imaging classifiers to check for unintended encoding of race, body mass, or scanner site, since the estimator accepts any attribute vector.
Because the observed attribute hierarchy is consistent across a CNN and a transformer, it likely reflects dataset-level clinical structure rather than one architecture's inductive bias; testing on an external CXR-echo cohort would confirm.

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

Quick read: the direct CXR-to-SLVH classifier is a reasonable thing to try, and the paper is honestly written, but the central comparison with Bhave et al. is not valid as reported. The test set was rebalanced to 50% prevalence, while the baseline AUPRC of 0.19 comes from the original imbalanced distribution. At 50% prevalence a random model already gets AUPRC 0.5, so the ViT's 0.80 is not comparable. The AUROC gap (0.816 vs 0.79) also sits within noise. They never re-run the baseline on the same balanced split, so "we beat the two-stage pipeline" is an overstatement.

What is worth keeping: this is the first direct SLVH classification on CXR that I know of, and the balanced subset is a sensible way to train a classifier for a rare condition if you say what you did. The MINE-based expressivity analysis is a nice illustration of a known method in a new domain, and the ordering age > sex > IVSDd > LVPWDd > LVIDd makes clinical sense. The writing is clear and the limitations are not hidden.

The soft spots beyond the metric problem: the MINE results have no null model or significance tests, so a claim like "the model encodes age" is just a magnitude comparison across attributes. Also, because the SLVH label comes from age- and sex-dependent thresholds, the high expressivity of age and sex is partly built into the label; that doesn't invalidate the analysis, but it dampens the surprise. No code or curated split is released, which makes it hard to reproduce the exact split.

Who this is for: someone working on pragmatic CXR screening or on interpretability tools for medical imaging will get useful ideas. As a refereed paper it needs one more round: re-evaluate the baseline on the balanced test set, report prevalence-matched AUPRC, and add a permutation baseline for MINE. I'd send it to review, but with the expectation of a major revision.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes a direct end-to-end classification approach for severe left ventricular hypertrophy (SLVH) from chest X-rays, removing the need for intermediate echocardiographic regression and demographic inputs. Using a balanced subset of the CheXchoNet dataset, the authors fine-tune ResNet-18 and a frozen pretrained ViT encoder with a trainable head, reporting ViT AUROC 0.816 [0.781, 0.850] and AUPRC 0.803 [0.755, 0.849], which they compare with the Bhave et al. baseline AUROC 0.79 and AUPRC 0.19. They also apply Mutual Information Neural Estimation (MINE) to quantify how much layer-wise features encode age, sex, IVSDd, LVPWd, and LVIDd, concluding a consistent attribute expressivity ordering age > sex > IVSDd > LVPWd > LVIDd across architectures.

Significance. If validated, the direct classification approach is potentially clinically valuable because chest X-rays are more accessible than echocardiography or MRI, and avoiding a multi-stage regression pipeline could reduce error propagation. The paper's strengths include the use of a public dataset, a transparent evaluation with confidence intervals, and the use of MINE as a concrete interpretability tool for probing internal representations. The layer-wise expressivity analysis is a useful direction for model transparency. However, the headline performance comparison and the interpretability claims currently outrun the evidence because of the test-set rebalancing and the lack of null-model calibration, so the significance of the reported numbers is not yet established.

major comments (3)

[Section 1.1, Table 1] The central comparison in Table 1 is invalid as reported. Section 1.1 states that equal numbers of SLVH-positive and SLVH-negative cases were sampled, so the test set has roughly 50% prevalence, while the baseline AUPRC of 0.19 from Bhave et al. was computed on the original low-prevalence distribution. AUPRC depends on prevalence: at 50% prevalence a chance classifier already obtains AUPRC 0.5, so comparing the ViT's 0.803 to the baseline's 0.19 confounds model quality with test-set construction. The manuscript must either re-evaluate the Bhave baseline on the identical balanced test set, report all AUPRCs at a common prevalence (including the original), or drop the claim that the proposed models achieve the best performance on both AUROC and AUPRC. Until this is fixed, the stated contribution of matching or beating the two-stage baseline is not supported.
[Section 1.3, Fig. 2] The MINE expressivity analysis reports absolute mutual information estimates but provides no null model, permutation test, or baseline comparison. Because MINE lower-bound estimates are biased and depend on network capacity, training stochasticity, and feature dimensionality, the reported differences among attributes and layers cannot be interpreted without a reference, such as the MI of the same features with shuffled attribute values or the MI of random features. The claim of a consistent hierarchy age > sex > IVSDd > LVPWd > LVIDd is therefore not statistically supported.
[Section 1.1, Introduction, Fig. 2] The interpretability conclusion is partly circular. The SLVH labels are constructed by applying echocardiographic thresholds that explicitly depend on age and sex (as the Introduction notes in citing clinical guidelines), so any classifier trained on those labels is expected to encode age and sex. The high MINE expressivity for age and sex in deeper layers therefore confirms the label-generation criteria rather than demonstrating that the model discovers clinically meaningful structure independent of demographics. The authors should either condition the MINE analysis on age/sex subgroups, train a comparison model on demographic-adjusted labels, or explicitly reframe the result as expected encoding of the label-generation process.

minor comments (6)

[Section 1.2, Abstract] The abstract and results state that pretrained models are fine-tuned, but Section 1.2 reports that the ViT encoder remains frozen and only the classification head is trained; this is linear probing, and the terminology should be made consistent.
[Section 1.3, Eq. (1)] The notation in Equation (1) is introduced only partially; the distributions P_FA and P_F⊗P_A should be defined explicitly in the text before the equation.
[Fig. 2] The x-axis in Fig. 2 is labeled as 'Layer' but the specific layers used for each architecture (e.g., which ResNet stage or which ViT block index) are not specified; this should be clarified so the layer-wise comparison is reproducible.
[Algorithm 1] The step 'after a particular epoch' is vague; the authors should state which checkpoint or epoch is used for feature extraction and whether the choice is based on validation AUROC.
[Throughout] There are minor typographical and formatting inconsistencies, including 'L VIDDd' instead of 'LVIDd', inconsistent use of 'SL VH' versus 'SLVH', and mixed CI formats in Table 1 (0.76–0.81 versus 0.72, 0.80); these should be unified.
[Section 1.1] The manuscript does not describe how the negative cases were sampled from the original dataset, nor does it report whether the balanced train/validation/test splits are patient-disjoint; this information is needed to assess generalizability.

Circularity Check

1 steps flagged · score 4.0 of 10

MINE 'expressivity' result is partly circular because SLVH labels are themselves age/sex-thresholded echo measurements; the main CXR classifier is not circular, but the AUPRC comparison with the baseline is confounded by prevalence.

self definitional [Introduction; Section 1.3; Fig 2]
"Consequently, prior work [6] adopts an indirect approach of first performing regression to estimate intermediate anatomical variables such as IVSDd, L VIDd, and L VPWDd, followed by thresholding to determine SL VH status. Additionally, since, clinical thresholds for defining SL VH depend on age and gender... MINE helps us confirm that the model attends to this underlying structure implicitly."

The paper's own description makes SLVH a thresholded function of the echo measurements IVSDd, LVIDd, and LVPWDd, with age/sex-dependent thresholds. The MINE analysis then estimates mutual information between the classifier's internal features and exactly those attribute values (age, sex, IVSDd, LVPWDd, LVIDd). A model trained to reproduce the binary target is optimizing a function of those very attributes, so nonzero MI for age/sex/IVSDd/LVPWDd is guaranteed by label construction, not an independent discovery about what CXR features encode. The reported ordering age>sex>IVSDd>LVPWDd>LVIDd is therefore a restatement of the label-generation process, making the 'expressivity reveals clinically meaningful encoding' claim partially self-referential.

full rationale

The central claim—direct SLVH classification from CXRs—is empirical and not circular: the model is trained on external labels and images, no fitted parameter is renamed as a prediction, and no load-bearing result rests on a same-author citation. The MINE interpretability claim, however, is partially circular because the SLVH labels are themselves derived from the anatomical and demographic attributes the MINE analysis 'discovers' are encoded. This is a genuine by-construction reduction for the interpretability finding, but it does not invalidate the measured AUROC/AUPRC numbers for the classifier itself. The AUPRC comparison between the imbalanced-distribution baseline (0.19) and the balanced-subset models (0.73-0.80) is a prevalence confound rather than circularity, since AUPRC depends on class prevalence and the baseline is not re-evaluated on the balanced subset; this should be treated as a benchmarking concern, not as a circular derivation. Overall, a partial circularity score of 4 reflects the self-referential nature of the expressivity analysis while acknowledging that the main classification contribution remains independent.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central evaluation depends on a user-chosen class balance ratio, not on fitted physical constants. The interpretability claim relies on an unvalidated assumption that MINE estimates are reliable across layers, and the ground truth labels are taken as given from the CheXchoNet dataset. No new physical entities are introduced.

free parameters (1)

Test and validation class balance ratio = 0.5 (equal positives and negatives in train, validation, and test subsets)
The balanced subset fixes prevalence at 50%, directly shaping AUPRC and making the comparison with the imbalanced baseline's AUPRC of 0.19 invalid. This is a design choice, not a fitted value, but it is central to the performance claims.

assumptions (3)

domain assumption The CheXchoNet echo-derived SLVH labels are accurate and representative.
All metrics are computed against these labels; label noise or selection bias in the original dataset would propagate. Invoked in Section 1.1.
domain assumption The Donsker-Varadhan lower bound estimated by MINE is tight enough to rank attribute dependencies.
MINE returns a lower bound; without convergence diagnostics or control features, the reported expressivity ordering assumes the bound is comparably tight across attributes and layers. Invoked in Sections 1.3 and 1.4.
domain assumption Chest X-rays in CheXchoNet contain visual correlates of echo-based SLVH labels that transfer to the pretrained backbones.
The entire direct classification approach depends on the paired CXR images carrying sufficient phenotypic information. Invoked throughout Sections 1.1 and 1.2.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Encoding of Demographic and Anatomical Information in Chest X-Ray-based Severe Left Ventricular Hypertrophy Classifiers." pith.science (2026). https://pith.science/paper/UFKYNVVH

@misc{pith2026250603192,
  author       = {Pith},
  title        = {Pith review of: Encoding of Demographic and Anatomical Information in Chest X-Ray-based Severe Left Ventricular Hypertrophy Classifiers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UFKYNVVH}},
  note         = {Machine review of arXiv:2506.03192}
}

read the original abstract

While echocardiography and MRI are clinical standards for evaluating cardiac structure, their use is limited by cost and accessibility.We introduce a direct classification framework that predicts severe left ventricular hypertrophy from chest X-rays, without relying on anatomical measurements or demographic inputs. Our approach achieves high AUROC and AUPRC, and employs Mutual Information Neural Estimation to quantify feature expressivity. This reveals clinically meaningful attribute encoding and supports transparent model interpretation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 12 canonical work pages

[1]

Journal of the American College of Cardiology79(17), 263–421 (2022)

Heidenreich, P.A., Bozkurt, B., Aguilar, D., Allen, L.A., Byun, J.J., Colvin, M.M., Deswal, A., Drazner, M.H., Dunlay, S.M., Evers, L.R.,et al.: 2022 aha/ac- c/hfsa guideline for the management of heart failure: a report of the american college of cardiology/american heart association joint committee on clinical prac- tice guidelines. Journal of the Ameri...

work page 2022
[2]

Circulation95(6), 1686–1744 (1997)

Cheitlin, M.D., Alpert, J.S., Armstrong, W.F., Aurigemma, G.P., Beller, G.A., Bierman, F.Z., Davidson, T.W., Davis, J.L., Douglas, P.S., Gillam, L.D.,et al.: Acc/aha guidelines for the clinical application of echocardiography: a report of the american college of cardiology/american heart association task force on practice guidelines (committee on clinical...

work page 1997
[3]

The American journal of medicine107(3), 198–208 (1999)

Heidenreich, P.A., Masoudi, F.A., Maini, B., Chou, T.M., Foster, E., Schiller, N.B., Owens, D.K.: Echocardiography in patients with suspected endocarditis: a cost-effectiveness analysis. The American journal of medicine107(3), 198–208 (1999)

work page 1999
[4]

Cardiovascular Digital Health Journal2(2), 109–117 (2021)

Khurshid, S., Friedman, S.F., Pirruccello, J.P., Di Achille, P., Diamant, N., Ander- son, C.D., Ellinor, P.T., Batra, P., Ho, J.E., Philippakis, A.A.,et al.: Deep 8 learning to estimate cardiac magnetic resonance–derived left ventricular mass. Cardiovascular Digital Health Journal2(2), 109–117 (2021)

work page 2021
[5]

Radiology195(1), 245–246 (1995)

Gurney, J.: Why chest radiography became routine. Radiology195(1), 245–246 (1995)

work page 1995
[6]

European heart journal45(22), 2002–2012 (2024)

Bhave, S., Rodriguez, V., Poterucha, T., Mutasa, S., Aberle, D., Capaccione, K.M., Chen, Y., Dsouza, B., Dumeer, S., Goldstein, J.,et al.: Deep learning to detect left ventricular structural abnormalities in chest x-rays. European heart journal45(22), 2002–2012 (2024)

work page 2024
[7]

European Heart Journal-Cardiovascular Imaging 16(3), 233–271 (2015)

Lang, R.M., Badano, L.P., Mor-Avi, V., Afilalo, J., Armstrong, A., Ernande, L., Flachskampf, F.A., Foster, E., Goldstein, S.A., Kuznetsova, T.,et al.: Recom- mendations for cardiac chamber quantification by echocardiography in adults: an update from the american society of echocardiography and the european associa- tion of cardiovascular imaging. European...

work page 2015
[8]

IEEE transactions on pattern analysis and machine intelligence 35(8), 1798–1828 (2013)

Bengio, Y., Courville, A., Vincent, P.: Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence 35(8), 1798–1828 (2013)

2013

Show all 15 references

[9]

In: International Conference on Machine Learning, pp

Belghazi, M.I., Baratin, A., Rajeshwar, S., Ozair, S., Bengio, Y., Courville, A., Hjelm, D.: Mutual information neural estimation. In: International Conference on Machine Learning, pp. 531–540 (2018). PMLR

2018
[10]

Communications on pure and applied mathematics 28(1), 1–47 (1975)

Donsker, M.D., Varadhan, S.S.: Asymptotic evaluation of certain markov process expectations for large time, i. Communications on pure and applied mathematics 28(1), 1–47 (1975)

1975
[11]

85–92 (2020)

Dhar, P., Bansal, A., Castillo, C.D., Gleason, J., Phillips, P.J., Chellappa, R.: How are attributes expressed in face dcnns? In: 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pp. 85–92 (2020). IEEE

2020
[12]

arXiv preprint arXiv:2503.06451 (2025)

Pal, B., Huang, S., Chellappa, R.: A quantitative evaluation of the expressivity of bmi, pose and gender in body embeddings for recognition and identification. arXiv preprint arXiv:2503.06451 (2025)

2025 arXiv
[13]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recogni- tion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016)

2016
[14]

arXiv preprint arXiv:1412.6980 (2014)

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[15]

In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

Xiao, J., Bai, Y., Yuille, A., Zhou, Z.: Delving into masked autoencoders for 9 multi-label thorax disease classification. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 3588–3600 (2023) 10

2023

Pith tools

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

[1] [1]

Journal of the American College of Cardiology79(17), 263–421 (2022)

Heidenreich, P.A., Bozkurt, B., Aguilar, D., Allen, L.A., Byun, J.J., Colvin, M.M., Deswal, A., Drazner, M.H., Dunlay, S.M., Evers, L.R.,et al.: 2022 aha/ac- c/hfsa guideline for the management of heart failure: a report of the american college of cardiology/american heart association joint committee on clinical prac- tice guidelines. Journal of the Ameri...

work page 2022

[2] [2]

Circulation95(6), 1686–1744 (1997)

Cheitlin, M.D., Alpert, J.S., Armstrong, W.F., Aurigemma, G.P., Beller, G.A., Bierman, F.Z., Davidson, T.W., Davis, J.L., Douglas, P.S., Gillam, L.D.,et al.: Acc/aha guidelines for the clinical application of echocardiography: a report of the american college of cardiology/american heart association task force on practice guidelines (committee on clinical...

work page 1997

[3] [3]

The American journal of medicine107(3), 198–208 (1999)

Heidenreich, P.A., Masoudi, F.A., Maini, B., Chou, T.M., Foster, E., Schiller, N.B., Owens, D.K.: Echocardiography in patients with suspected endocarditis: a cost-effectiveness analysis. The American journal of medicine107(3), 198–208 (1999)

work page 1999

[4] [4]

Cardiovascular Digital Health Journal2(2), 109–117 (2021)

Khurshid, S., Friedman, S.F., Pirruccello, J.P., Di Achille, P., Diamant, N., Ander- son, C.D., Ellinor, P.T., Batra, P., Ho, J.E., Philippakis, A.A.,et al.: Deep 8 learning to estimate cardiac magnetic resonance–derived left ventricular mass. Cardiovascular Digital Health Journal2(2), 109–117 (2021)

work page 2021

[5] [5]

Radiology195(1), 245–246 (1995)

Gurney, J.: Why chest radiography became routine. Radiology195(1), 245–246 (1995)

work page 1995

[6] [6]

European heart journal45(22), 2002–2012 (2024)

Bhave, S., Rodriguez, V., Poterucha, T., Mutasa, S., Aberle, D., Capaccione, K.M., Chen, Y., Dsouza, B., Dumeer, S., Goldstein, J.,et al.: Deep learning to detect left ventricular structural abnormalities in chest x-rays. European heart journal45(22), 2002–2012 (2024)

work page 2024

[7] [7]

European Heart Journal-Cardiovascular Imaging 16(3), 233–271 (2015)

Lang, R.M., Badano, L.P., Mor-Avi, V., Afilalo, J., Armstrong, A., Ernande, L., Flachskampf, F.A., Foster, E., Goldstein, S.A., Kuznetsova, T.,et al.: Recom- mendations for cardiac chamber quantification by echocardiography in adults: an update from the american society of echocardiography and the european associa- tion of cardiovascular imaging. European...

work page 2015

[8] [8]

IEEE transactions on pattern analysis and machine intelligence 35(8), 1798–1828 (2013)

Bengio, Y., Courville, A., Vincent, P.: Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence 35(8), 1798–1828 (2013)

2013