Learning Cardiac Latent Representations in Vectorcardiogram Space

Alan Wee-Chung Liew; Bosong Huang; Ming Jin; Panzhen Zhao; Patricia Lee; Shirui Pan; Wei Jin; Zengxiang Li

arxiv: 2605.31249 · v1 · pith:NRSIY7TBnew · submitted 2026-05-29 · 💻 cs.LG · cs.AI

Learning Cardiac Latent Representations in Vectorcardiogram Space

Bosong Huang , Panzhen Zhao , Zengxiang Li , Patricia Lee , Wei Jin , Alan Wee-Chung Liew , Ming Jin , Shirui Pan This is my paper

Pith reviewed 2026-06-28 23:35 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords vectorcardiogramself-supervised learningECG representation learningdomain generalizationcardiac signalslatent representations

0 comments

The pith

Learning representations in vectorcardiogram space reduces redundancy and improves generalization over twelve-lead ECG methods.

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

The paper claims that twelve-lead ECG representation learning introduces redundancy because each lead is a different projection of the same underlying cardiac electrical activity. Shifting the learning process into the three-dimensional vectorcardiogram space, via a self-supervised framework called LVCG, produces view-invariant features instead of lead-specific artifacts. This change yields stronger performance on downstream tasks and markedly better robustness when input distributions shift between domains. A reader would care if the claim holds because ECG data is ubiquitous in medicine, and removing spurious correlations could make learned models more reliable for diagnosis and report generation.

Core claim

By learning view-invariant latent VCG representations rather than lead-specific artifacts, LVCG minimizes redundancy and improves generalization. LVCG generally outperforms ECG-space baselines across tasks, demonstrating enhanced robustness and generalization, especially in domain shift settings.

What carries the argument

LVCG, the self-supervised framework that learns unified latent representations directly in the physically grounded Frank VCG space instead of the projected twelve-lead ECG space.

If this is right

LVCG produces higher accuracy than ECG-space methods on standard cardiac tasks.
The method shows particular gains in robustness when the data distribution changes across recording sites or patient groups.
View-invariant features learned in VCG space avoid spurious correlations tied to individual lead placements.
The reduction in redundancy lowers the risk that models overfit to artifacts rather than true cardiac signals.

Where Pith is reading between the lines

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

The same principle of moving from multiple surface projections to an underlying physical latent space could apply to other multi-sensor biomedical signals.
If VCG reconstruction from standard ECG leads proves stable, LVCG could be inserted into existing clinical pipelines without new hardware.
Domain-shift gains suggest the method may help when models trained on hospital data are deployed in wearable or ambulatory settings.

Load-bearing premise

The Frank VCG model supplies a latent space that already contains less redundancy than the twelve-lead ECG projections.

What would settle it

A controlled experiment in which LVCG is trained and tested on the same tasks and domain-shift scenarios yet fails to match or exceed ECG-space baseline performance would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.31249 by Alan Wee-Chung Liew, Bosong Huang, Ming Jin, Panzhen Zhao, Patricia Lee, Shirui Pan, Wei Jin, Zengxiang Li.

**Figure 1.** Figure 1: ECG and VCG as complementary views of cardiac electrical activity. (Left) Multi-lead ECG provides multiple linear projections of the same underlying cardiac field from different lead directions. (Right) Frank vectorcardiogram (VCG) offers a 3D representation of this field, and lead geometry defines a mapping between VCG space and ECG space. such as atrial fibrillation and supraventricular tachycardia (Liu … view at source ↗

**Figure 2.** Figure 2: ECG self-supervised learning via multi-lead reconstruction in the LVCG framework. (a) Beat-level processing yields beat patches and rhythm embeddings. (b) VCG space transformation performs lift and project for cross-lead reconstruction. (c) The VCG beat encoder and temporal module produce structure and dynamic embeddings, which are concatenated with rhythm embeddings as the ECG embeddings for downstream ta… view at source ↗

**Figure 4.** Figure 4: t-SNE visualization of ECG embeddings on CPSC2018 dataset. Color indicates the primary label. LVCG yields reconstructions that better match the groundtruth embedding distribution and reduce errors across leads, consistent with improved view-consistent reconstruction. Embedding visualization on CPSC2018. CPSC2018 exhibits a substantial distribution shift relative to MIMICIV pretraining data, so it provid… view at source ↗

**Figure 5.** Figure 5: The impact study of input different numbers of leads. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

**Figure 6.** Figure 6: Reconstruction visualization on sample s0012 re from PTB dataset. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗

**Figure 7.** Figure 7: Reconstruction visualization on sample s055 re from PTB dataset. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

**Figure 8.** Figure 8: Reconstruction visualization on sample s0164lre from PTB dataset. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

read the original abstract

Electrocardiography (ECG) is a cornerstone of cardiac assessment, making the learning of informative ECG representations fundamental to tasks ranging from disease diagnosis to clinical report generation. However, existing methods operate almost exclusively in the observable ECG signal space. In practice, the standard twelve-lead ECG represents multiple projections of the same underlying cardiac electrical activity from different spatial orientations. Therefore, representation learning in the ECG space inevitably introduces substantial redundancy, which may lead to spurious correlations and increased risk of overfitting. To address this and motivated by the Frank vectorcardiogram (VCG) model, we propose learning a unified latent representation of cardiac electrical activity directly in the VCG space. We introduce LVCG, the first general self-supervised representation learning framework designed to operate in this physically grounded latent space. By learning view-invariant latent VCG representations rather than lead-specific artifacts, VCG minimizes redundancy and improves generalization. LVCG generally outperforms ECG-space baselines across tasks, demonstrating enhanced robustness and generalization, especially in domain shift settings.

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

The paper tries shifting self-supervised cardiac representation learning into VCG space to cut redundancy from ECG projections, but the central advantage is an untested modeling assumption rather than a demonstrated result.

read the letter

The core move here is to run self-supervised learning directly in the three-dimensional VCG space given by the Frank model instead of the usual twelve-lead ECG. The claim is that this removes lead-specific artifacts and improves robustness, especially under domain shift. LVCG is positioned as the first general framework to do this.

The paper does a clean job laying out why multiple projections of the same source can create redundancy and spurious correlations. Treating the VCG as the underlying representation is a reasonable physical motivation and distinguishes the work from standard ECG-space SSL.

The soft spot is that nothing in the abstract shows the VCG latents actually carry less redundancy. There is no reported measurement of mutual information, effective dimension, or reconstruction fidelity that would confirm the claimed reduction. The Frank dipole is an approximation; when it breaks down the supposed advantage disappears and any gains could just be architecture or training differences. Without methods, data details, or ablations it is impossible to tell whether the reported outperformance is real.

This is for people already working on cardiac signal ML who want to try a physically motivated input space. A reader in that subfield could extract the idea and test it themselves.

I would send it to peer review so the experiments can be checked properly, but the current text leaves the main claim hanging on an assumption.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes LVCG, the first self-supervised representation learning framework operating directly in Frank vectorcardiogram (VCG) space rather than twelve-lead ECG space. It claims that the physically grounded VCG model yields view-invariant latent representations with lower redundancy than lead-specific ECG projections, thereby reducing spurious correlations and improving generalization and robustness, with LVCG outperforming ECG-space baselines across tasks especially under domain shift.

Significance. If the empirical gains are shown to arise specifically from reduced redundancy in the VCG space (rather than architecture or training differences) and the Frank model is sufficiently complete for the evaluated populations, the approach could supply a principled, physically motivated alternative to standard ECG representation learning with better out-of-distribution behavior.

major comments (2)

[Abstract, Introduction] Abstract and Introduction: The central claim that 'VCG minimizes redundancy' and thereby improves generalization rests on the untested modeling assumption that the Frank dipole model provides a lower-redundancy space than 12-lead projections without information loss for task-relevant signals. No direct measurement (mutual information, effective dimensionality, or reconstruction error) comparing learned VCG latents to ECG inputs is supplied to support this attribution.
[Experiments] Experiments/Results: The reported outperformance over ECG-space baselines is presented as evidence of enhanced robustness due to the VCG space, yet the manuscript provides no ablation that isolates the contribution of the VCG representation from differences in the self-supervised objective, network architecture, or training procedure. This leaves open the possibility that gains are not load-bearing on the claimed redundancy reduction.

minor comments (1)

[Methods] Notation: The distinction between the learned latent VCG representation and the input VCG reconstruction could be clarified with consistent symbols across equations and figures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting areas where additional evidence would strengthen our claims. We address each major comment below and agree that revisions are warranted to provide direct measurements and ablations. These will be incorporated in the revised manuscript.

read point-by-point responses

Referee: [Abstract, Introduction] Abstract and Introduction: The central claim that 'VCG minimizes redundancy' and thereby improves generalization rests on the untested modeling assumption that the Frank dipole model provides a lower-redundancy space than 12-lead projections without information loss for task-relevant signals. No direct measurement (mutual information, effective dimensionality, or reconstruction error) comparing learned VCG latents to ECG inputs is supplied to support this attribution.

Authors: We agree that direct quantitative support for the redundancy reduction claim would strengthen attribution of gains to the VCG space. The Frank model is a well-established physical representation of cardiac electrical activity as a dipole, which by construction uses three orthogonal components rather than the redundant projections in 12-lead ECG. While our empirical results (particularly under domain shift) are consistent with this motivation, we will add in revision explicit comparisons: mutual information between learned latents and inputs, effective dimensionality via participation ratio or PCA, and reconstruction error from VCG vs. ECG representations. This will provide the requested measurements. revision: yes
Referee: [Experiments] Experiments/Results: The reported outperformance over ECG-space baselines is presented as evidence of enhanced robustness due to the VCG space, yet the manuscript provides no ablation that isolates the contribution of the VCG representation from differences in the self-supervised objective, network architecture, or training procedure. This leaves open the possibility that gains are not load-bearing on the claimed redundancy reduction.

Authors: We acknowledge that the current experiments compare against published ECG-space baselines without fully matched architectures and objectives, leaving room for confounding factors. To isolate the effect of operating in VCG space, the revised manuscript will include controlled ablations: (1) applying the identical self-supervised objective and network architecture to both VCG and ECG inputs, and (2) additional controls varying only the input representation while holding all other factors fixed. These will directly test whether performance differences arise from the representation space itself. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical claims rest on external comparisons without self-referential reductions.

full rationale

The provided manuscript text contains no equations, derivations, or parameter-fitting steps that reduce any claimed result to its own inputs by construction. The central motivation (Frank VCG model yields lower-redundancy space than 12-lead projections) is presented as a modeling assumption rather than a derived theorem, and performance claims are supported by task comparisons to baselines. No self-citation chains, ansatz smuggling, or renaming of known results appear. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review limits visibility into parameters or assumptions; the central motivation rests on the unexamined validity of the Frank VCG model as a redundancy-reducing latent space.

axioms (1)

domain assumption The Frank vectorcardiogram model accurately captures the underlying cardiac electrical activity as a low-redundancy 3D representation superior to twelve-lead ECG projections.
Invoked in the motivation paragraph to justify shifting representation learning to VCG space.

invented entities (1)

LVCG framework no independent evidence
purpose: Self-supervised learning of view-invariant latent VCG representations
New method introduced to operate in VCG space; no independent evidence provided in abstract.

pith-pipeline@v0.9.1-grok · 5717 in / 1283 out tokens · 18324 ms · 2026-06-28T23:35:42.273824+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 11 canonical work pages · 1 internal anchor

[1]

Z., and Wu, J

Chen, J., Zheng, X., Yu, H., Chen, D. Z., and Wu, J. Electro- cardio panorama: synthesizing new ecg views with self- supervision.arXiv preprint arXiv:2105.06293, 2021a. Chen, J., Liao, K., Wei, K., Ying, H., Chen, D. Z., and Wu, J. Me-gan: Learning panoptic electrocardio repre- sentations for multi-view ecg synthesis conditioned on heart diseases. InInter...

work page arXiv
[2]

Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation

Chen, X., Xie, S., and He, K. An empirical study of training self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pp. 9640–9649, 2021b. Cho, K., Van Merri ¨enboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., and Bengio, Y . Learn- ing phrase representations using rnn encoder-decoder...

work page internal anchor Pith review Pith/arXiv arXiv
[3]

K., Li, X., and Guan, C

Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C. K., Li, X., and Guan, C. Time-series representation learning via temporal and contextual contrasting.arXiv preprint arXiv:2106.14112,

work page arXiv
[4]

Reading your heart: Learning ecg words and sentences via pre-training ecg language model.arXiv preprint arXiv:2502.10707,

Jin, J., Wang, H., Li, H., Li, J., Pan, J., and Hong, S. Reading your heart: Learning ecg words and sentences via pre-training ecg language model.arXiv preprint arXiv:2502.10707,

work page arXiv
[5]

Yunfei Luo, Yuliang Chen, Asif Salekin, and Tauhidur Rahman

Li, J., Aguirre, A., Moura, J., Liu, C., Zhong, L., Sun, C., Clifford, G., Westover, B., and Hong, S. An electro- cardiogram foundation model built on over 10 million recordings with external evaluation across multiple do- mains.arXiv preprint arXiv:2410.04133,

work page arXiv
[6]

Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement

Liu, C., Wan, Z., Ouyang, C., Shah, A., Bai, W., and Ar- cucci, R. Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement. arXiv preprint arXiv:2403.06659,

work page arXiv
[7]

Guiding masked repre- sentation learning to capture spatio-temporal relationship of electrocardiogram.arXiv preprint arXiv:2402.09450,

Na, Y ., Park, M., Tae, Y ., and Joo, S. Guiding masked repre- sentation learning to capture spatio-temporal relationship of electrocardiogram.arXiv preprint arXiv:2402.09450,

work page arXiv
[8]

M., Saeed, A., and Ma, D

Pham, H. M., Saeed, A., and Ma, D. Boosting masked ecg-text auto-encoders as discriminative learners.arXiv preprint arXiv:2410.02131,

work page arXiv
[9]

From token to rhythm: A multi-scale approach for ecg-language pretraining.arXiv preprint arXiv:2506.21803,

Wang, F., Xu, J., and Yu, L. From token to rhythm: A multi-scale approach for ecg-language pretraining.arXiv preprint arXiv:2506.21803,

work page arXiv
[10]

Nef-net+: Adapting electrocardio panorama in the wild

Zhan, Z., Hu, Y ., Zhan, J., Lian, W., Wu, W., and Chen, J. Nef-net+: Adapting electrocardio panorama in the wild. arXiv preprint arXiv:2511.02880,

work page arXiv
[11]

The coordinate system is: x axis right→left, y axis superior→inferior, and z axis anterior→posterior

using fixed direction vectors from a standard 12-lead ECG geometry. The coordinate system is: x axis right→left, y axis superior→inferior, and z axis anterior→posterior. Limb leads lie in the frontal plane (z= 0 ) with Einthoven-style angles, while precordial leads use an approximate transverse- plane geometry with a small inferior tilt. We use the MIMIC ...

2000
[12]

Dataset Statistics In our study, we utilize a diverse set of ECG datasets for large-scale pre-training and comprehensive evaluation

B.2. Dataset Statistics In our study, we utilize a diverse set of ECG datasets for large-scale pre-training and comprehensive evaluation. Table 9 summarizes the core statistics of these datasets. Following the protocols in HeartLang (Jin et al., 2025), we utilize the full volume of MIMIC-IV-ECG for self-supervised pre-training and evaluate the learned rep...

2025
[13]

Evaluation 21,837 18,885 10s 500 12 5/21/44/12 Classification CPSC 2018 (Liu et al.,

2018
[14]

and CPSC 2018 (Liu et al.,

2018
[15]

PTB provides high-resolution signals (1000 Hz) that are ideal for verifying fine-grained morphological recovery

datasets. PTB provides high-resolution signals (1000 Hz) that are ideal for verifying fine-grained morphological recovery. CPSC 2018 contains signals with diverse arrhythmias and varying durations, providing a robust test for view-synthesis generalization. For both datasets, we simulate a view-generation scenario where the model reconstructs the full 12-l...

2018
[16]

Setting AUC (%) F1 (%)∆AUC vs

Nef-Net (Chen et al., 2021a) 782.84 24.6388.6325.22 0.23 LVCG 374.04 23.9491.0622.81 0.21 Table 11.Robustness under simulated electrode perturbation (linear probing). Setting AUC (%) F1 (%)∆AUC vs. baseline (pp) baseline 67.55 27.72 +0.00 noise: 0.01 67.55 27.79 +0.00 noise: 0.05 67.57 27.86 +0.02 noise: 0.1 67.59 27.87 +0.04 noise: 0.2 67.41 27.56−0.14 r...

work page arXiv 2018
[17]

Furthermore, LVCG demonstrates highlabel efficiencyregardless of lead sparsity

This provides strong empirical validation for our latent VCG formulation: as long as the visible leads provide a sufficient spatial span to recover the underlying cardiac electrical field (i.e., at least 3 leads), LVCG can effectively map the signals into a consistent latent space for downstream classification. Furthermore, LVCG demonstrates highlabel eff...

2018

[1] [1]

Z., and Wu, J

Chen, J., Zheng, X., Yu, H., Chen, D. Z., and Wu, J. Electro- cardio panorama: synthesizing new ecg views with self- supervision.arXiv preprint arXiv:2105.06293, 2021a. Chen, J., Liao, K., Wei, K., Ying, H., Chen, D. Z., and Wu, J. Me-gan: Learning panoptic electrocardio repre- sentations for multi-view ecg synthesis conditioned on heart diseases. InInter...

work page arXiv

[2] [2]

Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation

Chen, X., Xie, S., and He, K. An empirical study of training self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pp. 9640–9649, 2021b. Cho, K., Van Merri ¨enboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., and Bengio, Y . Learn- ing phrase representations using rnn encoder-decoder...

work page internal anchor Pith review Pith/arXiv arXiv

[3] [3]

K., Li, X., and Guan, C

Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C. K., Li, X., and Guan, C. Time-series representation learning via temporal and contextual contrasting.arXiv preprint arXiv:2106.14112,

work page arXiv

[4] [4]

Reading your heart: Learning ecg words and sentences via pre-training ecg language model.arXiv preprint arXiv:2502.10707,

Jin, J., Wang, H., Li, H., Li, J., Pan, J., and Hong, S. Reading your heart: Learning ecg words and sentences via pre-training ecg language model.arXiv preprint arXiv:2502.10707,

work page arXiv

[5] [5]

Yunfei Luo, Yuliang Chen, Asif Salekin, and Tauhidur Rahman

Li, J., Aguirre, A., Moura, J., Liu, C., Zhong, L., Sun, C., Clifford, G., Westover, B., and Hong, S. An electro- cardiogram foundation model built on over 10 million recordings with external evaluation across multiple do- mains.arXiv preprint arXiv:2410.04133,

work page arXiv

[6] [6]

Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement

Liu, C., Wan, Z., Ouyang, C., Shah, A., Bai, W., and Ar- cucci, R. Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement. arXiv preprint arXiv:2403.06659,

work page arXiv

[7] [7]

Guiding masked repre- sentation learning to capture spatio-temporal relationship of electrocardiogram.arXiv preprint arXiv:2402.09450,

Na, Y ., Park, M., Tae, Y ., and Joo, S. Guiding masked repre- sentation learning to capture spatio-temporal relationship of electrocardiogram.arXiv preprint arXiv:2402.09450,

work page arXiv

[8] [8]

M., Saeed, A., and Ma, D

Pham, H. M., Saeed, A., and Ma, D. Boosting masked ecg-text auto-encoders as discriminative learners.arXiv preprint arXiv:2410.02131,

work page arXiv

[9] [9]

From token to rhythm: A multi-scale approach for ecg-language pretraining.arXiv preprint arXiv:2506.21803,

Wang, F., Xu, J., and Yu, L. From token to rhythm: A multi-scale approach for ecg-language pretraining.arXiv preprint arXiv:2506.21803,

work page arXiv

[10] [10]

Nef-net+: Adapting electrocardio panorama in the wild

Zhan, Z., Hu, Y ., Zhan, J., Lian, W., Wu, W., and Chen, J. Nef-net+: Adapting electrocardio panorama in the wild. arXiv preprint arXiv:2511.02880,

work page arXiv

[11] [11]

The coordinate system is: x axis right→left, y axis superior→inferior, and z axis anterior→posterior

using fixed direction vectors from a standard 12-lead ECG geometry. The coordinate system is: x axis right→left, y axis superior→inferior, and z axis anterior→posterior. Limb leads lie in the frontal plane (z= 0 ) with Einthoven-style angles, while precordial leads use an approximate transverse- plane geometry with a small inferior tilt. We use the MIMIC ...

2000

[12] [12]

Dataset Statistics In our study, we utilize a diverse set of ECG datasets for large-scale pre-training and comprehensive evaluation

B.2. Dataset Statistics In our study, we utilize a diverse set of ECG datasets for large-scale pre-training and comprehensive evaluation. Table 9 summarizes the core statistics of these datasets. Following the protocols in HeartLang (Jin et al., 2025), we utilize the full volume of MIMIC-IV-ECG for self-supervised pre-training and evaluate the learned rep...

2025

[13] [13]

Evaluation 21,837 18,885 10s 500 12 5/21/44/12 Classification CPSC 2018 (Liu et al.,

2018

[14] [14]

and CPSC 2018 (Liu et al.,

2018

[15] [15]

PTB provides high-resolution signals (1000 Hz) that are ideal for verifying fine-grained morphological recovery

datasets. PTB provides high-resolution signals (1000 Hz) that are ideal for verifying fine-grained morphological recovery. CPSC 2018 contains signals with diverse arrhythmias and varying durations, providing a robust test for view-synthesis generalization. For both datasets, we simulate a view-generation scenario where the model reconstructs the full 12-l...

2018

[16] [16]

Setting AUC (%) F1 (%)∆AUC vs

Nef-Net (Chen et al., 2021a) 782.84 24.6388.6325.22 0.23 LVCG 374.04 23.9491.0622.81 0.21 Table 11.Robustness under simulated electrode perturbation (linear probing). Setting AUC (%) F1 (%)∆AUC vs. baseline (pp) baseline 67.55 27.72 +0.00 noise: 0.01 67.55 27.79 +0.00 noise: 0.05 67.57 27.86 +0.02 noise: 0.1 67.59 27.87 +0.04 noise: 0.2 67.41 27.56−0.14 r...

work page arXiv 2018

[17] [17]

Furthermore, LVCG demonstrates highlabel efficiencyregardless of lead sparsity

This provides strong empirical validation for our latent VCG formulation: as long as the visible leads provide a sufficient spatial span to recover the underlying cardiac electrical field (i.e., at least 3 leads), LVCG can effectively map the signals into a consistent latent space for downstream classification. Furthermore, LVCG demonstrates highlabel eff...

2018