Pith. sign in

REVIEW 5 major objections 5 minor 22 references

NeuroHD-RA: Neural-distilled Hyperdimensional Model with Rhythm Alignment

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that distilling a trainable ECG encoder into a binary hyperdimensional classifier, aligned to RR intervals and trained with proxy metric loss, beats traditional HDC and classical ML baselines on real ECG benchmarks while…

desk verdict Trainable RR-aligned HDC for ECG is a real but overclaimed contribution: the temporal-dynamics selling point dies in Eq. 11's commutative sum, and the abstract's superiority over 'traditional HDC' is never tested. read the letter →

arxiv 2507.14184 v3 pith:L22GV3X4 submitted 2025-07-12 eess.SP cs.AIcs.LG

ZhengXiao He , Jinghao Wen , Huayu Li , Siyuan Tian , Ao Li This is my paper · ORCID
classification eess.SPcs.AIcs.LG
keywords hyperdimensionalcomputingECGclassificationsleepapneadetectionRR-intervalsegmentationproxymetriclearningneuraldistillationedgeinterpretability
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

This paper tries to establish that hyperdimensional computing, usually limited by handcrafted random encoders, can be made competitive for real ECG diagnostics by replacing the fixed encoder with a small trainable neural encoder and aligning representations to cardiac cycles. The method, NeuroHD-RA, segments the ECG into RR intervals, projects each block into a high-dimensional binary space, sums the block hypervectors, and classifies by cosine similarity against learned class proxies. On Apnea-ECG the authors report 73.09% precision and an F1 score of 0.626, beating logistic regression, random forest, and 1D CNN baselines, while the full model stores in 124.5 KB and infers in about 21.5 ms on a CPU. The significance would be that lightweight, interpretable, symbol-like ECG screening could run on wearables without sacrificing accuracy to classical shallow methods. The paper also claims block-level similarity scores offer temporal localization of abnormal beats.

What carries the argument

The load-bearing machinery is the BinaryLinear hyperdimensional projection layer: a learned linear map $\mathbf{W} \in \mathbb{R}^{D \times d}$ with binarized weights using a straight-through estimator, followed by $\mathbf{z}_{b,t} = \tanh(\mathbf{W}_{\mathrm{bin}}\mathbf{x}_{b,t})$. Each RR-interval block is projected this way, then the per-block hypervectors are summed over time as $Z_b = \sum_t z_{b,t}$, and at inference the signs are taken and cosine similarity against learned class proxies decides the label. The design intentionally omits permutation-based positional encoding, arguing that RR-aligned segmentation itself carries the rhythm structure. A proxy-based contrastive loss organizes the embedding space so class prototypes gain angular separation beyond simple averaging.

What would settle it

Shuffle the order of the RR blocks within each test sequence and measure Apnea-ECG accuracy and F1. If the score does not fall materially, the summation over blocks is ignoring cardiac-cycle order, so the reported advantage over fixed-length encoding cannot be attributed to rhythm-aware temporal dynamics. A second check is to rerun the fixed-length control with the same network capacity, same preprocessing, and matched input length; the 46.88% accuracy baseline is low enough that the comparison depends on that control being fair.

Watch

Extended reading notes

Core claim

The paper's central claim is that a neural-distilled hyperdimensional computing pipeline, whose binary projection weights are learned end-to-end and binarized with a straight-through estimator, can encode RR-aligned ECG blocks into discriminative hypervectors, and that adding proxy-based metric loss to cross-entropy sharpens class separability enough to beat both classical HDC and shallow neural baselines on sequence-level sleep apnea detection. On PTB-XL the same pipeline reaches an F1 score of 0.715 in a binary normal-versus-abnormal setting, trailing the reported CNNs slightly but with a much smaller memory footprint and faster CPU inference. The authors position this as the first integration of RR-guided temporal alignment into HDC and the first use of proxy metric learning in an HDC setting for physiological signals.

Load-bearing premise

The load-bearing premise is that summing per-beat hypervectors preserves the rhythm information that makes RR alignment better than fixed-length blocks; if per-block inputs do not include beat order or interval duration, the representation is an orderless bag of beats and the claimed temporal advantage collapses.

Editorial extensions

If this is right

  • If the reported results hold, a binary hyperdimensional model with a 124.5 KB memory can match or beat logistic regression, random forest, and 1D CNNs on sequence-level sleep-apnea detection, making wearable screening feasible.
  • RR-interval alignment can be adopted by other HDC pipelines as a preprocessing prior without changing the symbolic inference machinery.
  • Proxy-based metric loss gives HDC class prototypes better angular separation than simple averaging, which should transfer to other HDC classification tasks.
  • Block-level cosine scores provide a built-in temporal localization trace for abnormal beats, so interpretability does not require a separate explainability model.
  • The claimed efficiency, 21.54 ms CPU inference with binary memory, follows directly from replacing convolutions with one binary feedforward projection and summation.

Reading between the lines

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

  • Editorial inference: because the per-block hypervectors are summed without permutation or binding, the model treats a sequence as a bag of beats unless block order or interval duration is encoded in the input; a shuffle test would separate 'rhythm alignment' from 'better block features'.
  • Editorial inference: the large gap between the fixed-length control (46.88% accuracy) and the RR-block variant (73.09% accuracy) may partly reflect different input lengths or preprocessing rather than rhythm alignment, so re-running the fixed-length control with matched network capacity and the same per-block projection would clarify the comparison.
  • Editorial inference: if RR-interval durations were appended as explicit features to each block embedding, the same architecture could genuinely capture heart-rate variability, an extension the paper motivates but does not implement.
  • Editorial inference: the PTB-XL results show CNNs still lead in absolute accuracy, so the honest claim is not superiority over deep networks but a favorable accuracy-efficiency trade-off; a deployment-focused comparison on microcontroller energy would sharpen the claim.
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

5 major / 5 minor

Summary. The manuscript proposes NeuroHD-RA, a neural-distilled hyperdimensional computing (HDC) model for ECG classification. The pipeline segments ECG into RR intervals, maps each block through a learned binarized projection, sums the block embeddings to form a sequence-level hypervector, and classifies by cosine similarity against learned proxy hypervectors. Training combines cross-entropy with a proxy-based contrastive loss. Experiments are reported on Apnea-ECG and PTB-XL under a binary classification setting, together with CPU inference time and model size measurements. The central claim is that the rhythm-aligned RR-block encoding significantly outperforms traditional HDC and classical ML baselines while retaining a very small binary footprint.

Significance. The paper has useful ingredients: held-out patient-wise evaluation on PTB-XL, a fixed official test protocol for Apnea-ECG, and a concrete efficiency report (124.5 KB binary model, 21.54 ms CPU inference). If the comparative and mechanistic claims were supported, the work would be a meaningful step toward interpretable, edge-deployable ECG classification. However, the central claims are currently under-supported: no classical HDC baseline is evaluated, the rhythm-aware mechanism is not visible in the model equations, and the fixed-length control is under-specified. The paper also does not release code or preprocessing details, which limits reproducibility.

major comments (5)
  1. [Abstract and Tables I-II] The abstract claims that the model 'significantly outperforms traditional HDC and classical ML baselines,' but Tables I and II contain no classical HDC baseline. The fixed-length block variant is still a neural-distilled HDC model, not a traditional random-projection HDC encoder of the kind described in Section III-A (e.g., quantized item memory with cyclic permutation). Without such a baseline, the headline comparative claim is unsupported.
  2. [Section III-D2, Eq. (11)] The rhythm-aware claim is contradicted by the aggregation operation. In Eq. (11), Z_b = sum_{t=1}^T z_{b,t}, which is commutative, and the same W_bin is applied to every block with no documented positional code, permutation, or RR-duration feature. The text does not state that x_{b,t} includes RR interval length or beat order. As written, the representation is an orderless bag of RR-aligned blocks, and Section V concedes that no sequence-level dynamics beyond bundling are incorporated. To substantiate the claimed 'heart rate variability and temporal dynamics,' the paper must either include such information in the block input or aggregation, or revise the claim.
  3. [Section IV-A, Table I] The fixed-length block control that motivates the rhythm-alignment advantage is under-specified and internally inconsistent. The reported F1 of 0.524 ± 0.020 is not compatible with the listed precision (40.26%) and recall (78.30%), since 2PR/(P+R) ≈ 53.2%. No block length, number of blocks, or training protocol is given for this variant. As the control underlying a central claim, it needs full specification and recomputation.
  4. [Section IV-A, Table II] The PTB-XL table header says 'Normal vs Apnea,' but PTB-XL does not contain an apnea diagnostic class, and the text defines the task as normal versus abnormal. This discrepancy makes the reported PTB-XL numbers ambiguous and needs to be corrected before the results can be interpreted.
  5. [Section IV-B, Table I] The word 'significantly' is used without any significance testing. On F1, the proposed method (0.626 ± 0.123) overlaps with AlexNet (0.610 ± 0.021), and the precision interval overlaps with logistic regression. The paper should either report paired statistical tests or use more qualified comparative language.
minor comments (5)
  1. [Abstract] The abstract reports '73.09% precision,' but Table I lists accuracy as 73.09% and precision as 73.68%; the abstract should state accuracy.
  2. [Section IV-B] The text mentions SVM among the conventional models, but no SVM row appears in Table I; please align the text with the table.
  3. [Section III-E, Eqs. (15)-(16)] The proxy-based contrastive loss in Eq. (16) is a softmax over proxy distances, which is close in form to a cross-entropy loss over proxy logits; the paper should clarify what the contrastive term adds beyond the classification head.
  4. [Section IV-A] Key preprocessing details are missing: the R-peak detection method, the construction of RR blocks, the handling of variable RR lengths, and the fixed-length block configuration. Adding these details is necessary for reproducibility.
  5. [Section IV-E] Figure 3 is referenced in the text but does not appear in the provided manuscript; the figure should be included or the reference removed.

Circularity Check

1 steps flagged · score 3.0 of 10

One mechanism-level reduction: the 'proxy-based contrastive' loss (Eq. 16) with L2-normalized cosine embeddings equals the paper's own softmax classifier objective, so the claimed dual-objective proxy supervision is a re-scaled single loss; the benchmark results themselves are held-out and not circular.

  1. other [Section III-B; Section III-E (Eqs. 15-16); Section IV-A]
    "Ltotal = Lcls + λLcontrastive (15) ... Lcontrastive = − log exp(−d(z, py))/ΣCj=1 exp(−d(z, pj)) (16) ... all embeddings and proxies are L2-normalized ... softmax outputs over L2-normalized embeddings with temperature scaling approximate cosine similarity rankings, especially when class vectors (proxies) are also normalized."

    With cosine distance on L2-normalized embeddings/proxies (Section IV-A), exp(−d(z,p_j)) = exp(cos(z,p_j)−1), so Eq. 16 is exactly the softmax over cosine similarities to class vectors — the same normalized-softmax objective the paper's training classifier is described as using ('softmax outputs over L2-normalized embeddings ... approximate cosine similarity rankings, especially when class vectors (proxies) are also normalized,' Section III-B). Hence L_total = (1+λ)L_cls: the claimed 'joint training paradigm that combines classification loss with a contrastive learning' (Section III-E) reduces, by the paper's own equations, to a single re-weighted classification objective.

full rationale

The central empirical results are not circular: the model is trained end-to-end without fitting test labels and is scored on held-out external splits (Apnea-ECG official test set; PTB-XL patient-wise 5-fold cross-validation), so no fitted parameter is renamed as a prediction, and the accuracy/F1 numbers stand as independent evidence. No self-citation is load-bearing: the authors cite no work of their own; the neural-distillation result invoked (Reference [18]) is external, and no uniqueness theorem is imported. Two concerns are flagged but are not circularity. First, the claimed 'rhythm-aligned ... temporal dynamics' (Abstract; Section III-A) is unsupported by the paper's own Eq. 11, an order-invariant sum Z_b = Σ_t z_b,t, and Section V explicitly concedes 'our current implementation does not incorporate temporal attention or sequence-level dynamics beyond bundling,' while Section III-A concedes 'permutation may be omitted'; this is a claim-evidence gap (a correctness risk), not a reduction of the result to its inputs. Second, the abstract's claim to 'significantly outperform traditional HDC' is unverified by Tables I-II, which contain no random-projection HDC baseline — an experimental omission, not a circular step. The single genuine reduction found is the proxy-loss/CE identity detailed in the step above: with L2-normalized embeddings and cosine distance, Eq. 16 is the same normalized softmax objective as the classifier loss, so the 'joint' objective is a single re-scaled loss and the ablation's attribution to a distinct proxy-based mechanism is not supported by the stated equations. This affects the stated design mechanism but not the benchmark measurements. Overall score 3: one partial, mechanism-level reduction; the headline results remain independent and externally benchmarked.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or abstract entities are introduced; the contributions are architectural configurations. The free parameters are hyperparameters chosen by hand or validation, and the axioms are domain heuristics about ECG segmentation, summation-as-bundling, and proxy interpretability. None are derived from first principles.

free parameters (3)
  • D (hypervector dimensionality) = 10,000
    Set in Section IV.A to 'strike a balance between accuracy and symbolic representational capacity'; no sweep or theoretical derivation is reported.
  • lambda (contrastive loss weight) = 0.5
    Set 'empirically based on validation performance' in Section IV.A; the joint loss in Eq. 15 depends directly on this choice.
  • RR block input dimension d = not stated (inferred ~100 from 124.5 KB model size at D=10,000)
    The per-block input length and any embedded RR features are unspecified, yet they define the projection layer W in R^(D x d) in Eq. 9.
assumptions (4)
  • domain assumption ECG can be reliably segmented into RR intervals and each interval carries discriminative class information.
    Section III.C builds the entire pipeline on RR-interval blocks; the method assumes robust R-peak detection and that a per-beat representation is sufficient for apnea and disease classification.
  • ad hoc to paper Summing binarized block embeddings approximates HDC bundling and yields valid hypervectors for cosine similarity.
    Eqs. 10-13 replace the classical sign of a bundle with an unbounded integer sum; the paper provides no analysis that this preserves HDC's theoretical properties.
  • domain assumption Cross-entropy plus proxy-based contrastive loss improves generalization over cross-entropy alone.
    The joint loss in Eq. 15 is motivated empirically; no ablation isolates the proxy term's contribution, and the proxy loss in Eq. 16 is functionally similar to the CE loss over normalized embeddings.
  • ad hoc to paper Learned class proxies serve as interpretable HDC class prototypes.
    Section III.D.4 interprets cosine similarity to proxies as blockwise explainability, but the proxies are optimized for classification, so the visualization reflects the model's own decision surface rather than an independent physiological signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeuroHD-RA: Neural-distilled Hyperdimensional Model with Rhythm Alignment." pith.science (2026). https://pith.science/paper/L22GV3X4

@misc{pith2026250714184,
  author       = {Pith},
  title        = {Pith review of: NeuroHD-RA: Neural-distilled Hyperdimensional Model with Rhythm Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L22GV3X4}},
  note         = {Machine review of arXiv:2507.14184}
}
read the original abstract

We present a novel and interpretable framework for electrocardiogram (ECG)-based disease detection that combines hyperdimensional computing (HDC) with learnable neural encoding. Unlike conventional HDC approaches that rely on static, random projections, our method introduces a rhythm-aware and trainable encoding pipeline based on RR intervals, a physiological signal segmentation strategy that aligns with cardiac cycles. The core of our design is a neural-distilled HDC architecture, featuring a learnable RR-block encoder and a BinaryLinear hyperdimensional projection layer, optimized jointly with cross-entropy and proxy-based metric loss. This hybrid framework preserves the symbolic interpretability of HDC while enabling task-adaptive representation learning. Experiments on Apnea-ECG and PTB-XL demonstrate that our model significantly outperforms traditional HDC and classical ML baselines, achieving 73.09\% precision and an F1 score of 0.626 on Apnea-ECG, with comparable robustness on PTB-XL. Our framework offers an efficient and scalable solution for edge-compatible ECG classification, with strong potential for interpretable and personalized health monitoring.

Figures

Figures reproduced from arXiv: 2507.14184 by the authors.

Figure 1
Figure 1. Flowmap and discrimination. It also aligns with architectural analogies drawn between HDC and DNNs in [18], and demonstrates that neural-distilled HDC can generalize to real physiological tasks beyond synthetic or vision benchmarks. C. Flowmap We provide an overview of our overall workflow. As illus￾trated in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Neural-symbolic joint training framework with cross-entropy and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Interpretable Visualization of RR-Interval Blocks [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

  1. [1]

    The apnea-ecg database,

    T. Penzel, G. B. Moody, R. G. Mark, A. L. Goldberger, and J. H. Peter, “The apnea-ecg database,” in Computers in Cardiology 2000. Vol. 27 (Cat. 00CH37163). IEEE, 2000, pp. 255–258

    work page 2000

  2. [2]

    Longitudinal evaluation of sleep- disordered breathing and sleep symptoms with change in quality of life: the sleep heart health study (shhs),

    G. E. Silva, M.-W. An, J. L. Goodwin, E. Shahar, S. Redline, H. Resnick, C. M. Baldwin, and S. F. Quan, “Longitudinal evaluation of sleep- disordered breathing and sleep symptoms with change in quality of life: the sleep heart health study (shhs),” Sleep, vol. 32, no. 8, pp. 1049–1057, 2009

    work page 2009

  3. [3]

    Ptb-xl, a large publicly available electro- cardiography dataset,

    P. Wagner, N. Strodthoff, R.-D. Bousseljot, D. Kreiseler, F. I. Lunze, W. Samek, and T. Schaeffter, “Ptb-xl, a large publicly available electro- cardiography dataset,” Scientific data, vol. 7, no. 1, pp. 1–15, 2020

    2020

  4. [4]

    Cardiologist-level arrhythmia detection with convolutional neural networks,

    P. Rajpurkar, A. Y . Hannun, M. Haghpanahi, C. Bourn, and A. Y . Ng, “Cardiologist-level arrhythmia detection with convolutional neural networks,” arXiv preprint arXiv:1707.01836 , 2017

    arXiv 2017

  5. [5]

    Inceptiontime: Finding alexnet for time series classification,

    H. Ismail Fawaz, B. Lucas, G. Forestier, C. Pelletier, D. F. Schmidt, J. Weber, G. I. Webb, L. Idoumghar, P.-A. Muller, and F. Petitjean, “Inceptiontime: Finding alexnet for time series classification,” Data Mining and Knowledge Discovery , vol. 34, no. 6, pp. 1936–1962, 2020

    work page 1936

  6. [6]

    Medformer: A multi- granularity patching transformer for medical time-series classification,

    Y . Wang, N. Huang, T. Li, Y . Yan, and X. Zhang, “Medformer: A multi- granularity patching transformer for medical time-series classification,” arXiv preprint arXiv:2405.19363 , 2024

    arXiv 2024

  7. [7]

    Hyperdimensional computing: An introduction to comput- ing in distributed representation with high-dimensional random vectors,

    P. Kanerva, “Hyperdimensional computing: An introduction to comput- ing in distributed representation with high-dimensional random vectors,” Cognitive computation, vol. 1, pp. 139–159, 2009

    work page 2009

  8. [8]

    A robust and energy- efficient classifier using brain-inspired hyperdimensional computing,

    A. Rahimi, P. Kanerva, and J. M. Rabaey, “A robust and energy- efficient classifier using brain-inspired hyperdimensional computing,” in Proceedings of the 2016 international symposium on low power electronics and design , 2016, pp. 64–69

    work page 2016

Show all 22 references
  1. [9]

    Lehdc: Learning-based hyperdi- mensional computing classifier,

    S. Duan, Y . Liu, S. Ren, and X. Xu, “Lehdc: Learning-based hyperdi- mensional computing classifier,” in Proceedings of the 59th ACM/IEEE Design Automation Conference , 2022, pp. 1111–1116

    2022

  2. [10]

    V oicehd: Hyperdi- mensional computing for efficient speech recognition,

    M. Imani, D. Kong, A. Rahimi, and T. Rosing, “V oicehd: Hyperdi- mensional computing for efficient speech recognition,” in 2017 IEEE international conference on rebooting computing (ICRC) . IEEE, 2017, pp. 1–8

    2017

  3. [11]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

    1998

  4. [12]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” arXiv preprint arXiv:1708.07747, 2017

    2017 arXiv

  5. [13]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” University of Toronto, Tech. Rep., 2009

    2009

  6. [14]

    Hdcoin: A proof-of-useful-work based blockchain for hyperdimensional computing,

    D. Ma, S. Zhang, and X. Jiao, “Hdcoin: A proof-of-useful-work based blockchain for hyperdimensional computing,” arXiv preprint arXiv:2202.02964, 2022

    2022 arXiv

  7. [15]

    Efficient biosignal processing using hyperdimensional computing: Network templates for combined learning and classification of exg signals,

    A. Rahimi, P. Kanerva, L. Benini, and J. M. Rabaey, “Efficient biosignal processing using hyperdimensional computing: Network templates for combined learning and classification of exg signals,” Proceedings of the IEEE, vol. 107, no. 1, pp. 123–143, 2018

    2018

  8. [16]

    An emg gesture recognition system with flexible high-density sensors and brain-inspired high- dimensional computing,

    A. Moin, A. Zhou, A. Rahimi, S. Benatti, G. Alexandrov, S. Tamakloe, J. Ting, N. Yamamoto, and J. Rabaey, “An emg gesture recognition system with flexible high-density sensors and brain-inspired high- dimensional computing,” in2018 IEEE International Solid-State Circuits Confe...

    2018

  9. [17]

    Hyperdimensional brain-inspired learning for phoneme recognition with large-scale inferior colliculus neural activities,

    Y . Ni, Y . Yang, H. Chen, X. Wang, N. Lesica, F.-g. Zeng, and M. Imani, “Hyperdimensional brain-inspired learning for phoneme recognition with large-scale inferior colliculus neural activities,” IEEE Transactions on Biomedical Engineering , 2024

    2024

  10. [18]

    Hyperdimensional computing vs. neural networks: Comparing architecture and learning process,

    D. Ma, C. Hao, and X. Jiao, “Hyperdimensional computing vs. neural networks: Comparing architecture and learning process,” in 2024 25th International Symposium on Quality Electronic Design (ISQED). IEEE, 2024, pp. 1–5

    2024

  11. [19]

    Lightweight multi-task hyper- dimensional computing framework driven by binary neural network for sleep apnea detection,

    T. Chen, Y . Liu, G. Liu, and C. Wang, “Lightweight multi-task hyper- dimensional computing framework driven by binary neural network for sleep apnea detection,” in 2024 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 2024, pp. 1–5

    2024

  12. [20]

    Energy-efficient sleep apnea detection using a hyperdimen- sional computing framework based on wearable bracelet photoplethys- mography,

    T. Chen, J. Zhang, Z. Xu, S. J. Redmond, N. H. Lovell, G. Liu, and C. Wang, “Energy-efficient sleep apnea detection using a hyperdimen- sional computing framework based on wearable bracelet photoplethys- mography,” IEEE Transactions on Biomedical Engineering , 2024

    2024

  13. [21]

    Facenet: A unified embed- ding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embed- ding for face recognition and clustering,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 815– 823

    2015

  14. [22]

    No fuss distance metric learning using proxies,

    Y . Movshovitz-Attias, A. Toshev, T. K. Leung, S. Ioffe, and S. Singh, “No fuss distance metric learning using proxies,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 360–368

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.