Pith. sign in

REVIEW 4 major objections 6 minor 54 references

Enhancing Interpretability Through Loss-Defined Classification Objective in Structured Latent Spaces

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Adding a cluster-structuring loss to cross-entropy improves accuracy and interpretability on three image benchmarks.

desk verdict A plausible loss-modification recipe with consistent but overstated gains; the headline numbers are inflated by test-set λ selection and the 'first' claim is wrong, yet the core idea is worth a careful referee. read the letter →

arxiv 2412.08515 v1 pith:PXMFKXPD submitted 2024-12-11 cs.LG cs.AI

Daniel Geissler , Bo Zhou , Mengxi Liu , Paul Lukowicz This is my paper · ORCID
classification cs.LGcs.AI
keywords LatentBoostdistancemetriclearningloss-definedclassificationspaceinterpretabilitySilhouettescoreweightedsumlossclusterseparationsupervised
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

Latent Boost adds a distance-metric term to the standard cross-entropy loss so that, during supervised training, each class's latent representations are pulled into compact clusters and pushed apart from other classes. The paper reports that this weighted-sum loss consistently outperforms both the plain baseline and the classic Magnet loss across Fashion MNIST, CIFAR-10, and CIFAR-100, with relative accuracy gains of 2 to 3 percent, training-time reductions of 13 to 21 percent, and higher Silhouette scores indicating better separated latent clusters. If true, the method offers a low-cost way to make neural network latent spaces more interpretable without sacrificing classification performance.

What carries the argument

The load-bearing object is the Latent Boost loss in Equation (11), a cluster-based distance-metric term built on the Magnet loss. It computes, for each sample, the negative log of the ratio between the sample's distance to its own cluster centroid (scaled by that cluster's variance) and the sum of its distances to all other cluster centroids (scaled by beta). Three modifications carry the improvement: PCA dimensionality reduction on the latent vectors before the distance computation (retaining components up to 95 percent cumulative explained variance), per-cluster variance sigma-squared replacing the single global variance of Magnet loss, and dynamic alpha (exponential decay) and beta (linear decay) schedules that shift the training focus from intra-class compactness to inter-class separation.

What would settle it

Train the same model with Latent Boost on a dataset whose class-conditional latent structure is deliberately non-spherical, for example classes arranged along curved manifolds or with two well-separated sub-clusters per class. If accuracy and Silhouette scores do not improve over the plain cross-entropy baseline, or if they degrade relative to a version using per-class full covariance, the hyper-spherical-cluster assumption is the load-bearing premise.

Watch

Extended reading notes

Core claim

The central claim is that inserting cluster-structure information into the classification objective, rather than treating the loss as a function of individual samples only, improves both accuracy and latent interpretability. Concretely, the authors redefine the Magnet loss with per-cluster variances, a PCA projection of latent vectors before distance computation, and epoch-dependent alpha and beta schedules that first tighten intra-class clusters and later increase inter-class separation, then combine the resulting Latent Boost loss with cross-entropy via a lambda-weighted sum. On the three benchmark datasets the method reaches its best results at lambda between 0.5 and 0.75, and the authors report that Latent Boost consistently outperforms the baseline and the classic Magnet loss in accuracy, Micro-F1, and Silhouette score while reducing the number of training epochs.

Load-bearing premise

Latent Boost assumes each class's latent representation forms a roughly hyper-spherical, compact cluster around a single centroid, so that one variance per cluster and Euclidean distances to centroids describe the structure worth optimizing; on datasets with elongated, hierarchical, or heavily overlapping clusters the loss should distort rather than structure the latent space.

Editorial extensions

If this is right

  • The same lambda-weighted combination can be applied to other base losses such as contrastive, triplet, and N-pair, with Magnet-style cluster terms showing the most robust gains in the paper's comparisons.
  • Because the loss uses only latent vectors and cluster means, it can be plugged into any differentiable classifier without changing the architecture or requiring labels beyond the classification task.
  • The reduced epochs and stable convergence suggest lower energy consumption and computational cost for training, an explicit goal of the authors.
  • The improvement in Silhouette scores implies that the latent space becomes more amenable to downstream clustering, retrieval, and visualization tasks.
  • The method's benefit shrinks on high-complexity datasets like CIFAR-100, indicating that the gains are tied to the separability of class clusters in the chosen latent layer.

Reading between the lines

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

  • A natural extension is to replace the fixed Euclidean centroid with a full covariance or learned metric per class; if the hyper-spherical assumption is the bottleneck, such a change should recover gains on datasets with elongated or hierarchical clusters.
  • The PCA threshold of 0.95 is a hyperparameter; testing the sensitivity of both accuracy and Silhouette score to this threshold would reveal whether the dimensionality reduction is load-bearing or merely a computational convenience.
  • Because the authors keep lambda constant during training, an adaptive schedule that lowers lambda once clusters stabilize could avoid the early instability they observed in dynamic lambda trials.
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 / 6 minor

Summary. The paper proposes Latent Boost, a supervised classification loss that combines a Magnet-style distance-metric loss with cross-entropy in a weighted sum (Eq. 11). The method adds PCA-based dimension reduction of latent vectors, per-cluster variance, and epoch-dependent schedules for the α and β balance terms. Experiments on Fashion MNIST, CIFAR-10, and CIFAR-100, using CNN, VGG-16, and ResNet-50 respectively, report 2–3% accuracy gains, 13–21% faster convergence, and higher Silhouette scores relative to a cross-entropy baseline and to classic Magnet loss.

Significance. If the empirical claims withstand a properly validated hyperparameter selection procedure, Latent Boost is a simple, low-cost addition that yields consistent but moderate accuracy improvements and faster convergence, with an interesting attempt to make latent structure an explicit training objective. The manuscript has strengths: five seeded runs with standard deviations for the main classification tables, a clear ablation isolating the PCA component, public benchmarks, and an explicit discussion of the hyper-spherical cluster assumption. However, the headline gains currently rest on test-set-based selection of the balancing weight λ, and the interpretability evidence is partly circular because the Silhouette score measures the same cluster compactness/separation that the loss is designed to enforce. These issues need to be resolved before the central claims can be credited.

major comments (4)
  1. [Section 5.1, Tables 3 and 4] The best value of λ is selected on the test set: Table 4 is captioned 'best λ selection' on the unseen test dataset, and Table 3 sweeps λ ∈ {0.1, ..., 0.9} with the test-set accuracy used to pick λ = 0.75/0.75/0.5. Selecting the best of nine correlated test results inflates the expected maximum relative to a prespecified or validation-based choice, so the reported gains of 2.56%, 2.98%, and 2.06% accuracy are not established as a fair comparison. Please repeat the evaluation using a validation split (or nested cross-validation) for λ selection, and report the accuracy on the test set for the validation-selected λ. The Silhouette scores in Table 5 inherit the same selection issue.
  2. [Section 5.2 and Table 5] The interpretability claim relies on the Silhouette score, but Table 5 reports a single point estimate per method/dataset with no standard deviation or per-seed values, despite the main experiments using five seeds. Moreover, the Silhouette score rewards exactly the tight intra-class clustering and inter-class separation that the Magnet-style component of Latent Boost is designed to enforce, so the reported improvements are a direct consequence of the training objective rather than an independent validation of interpretability. Please report error bars or per-seed scores, and additionally evaluate with a metric that is less directly aligned with the loss, such as a linear-probe accuracy on the latent features or a class-distinguishability measure.
  3. [Section 4.3, Equation (10)] The dynamic schedules for α and β depend on two hyperparameters, α0 and β0, but their values are never specified in the text, tables, or experiment section. For example, the text says 'starts at a value of 1 + α0' and 'starting from β0', yet no concrete values are given for α0 or β0 in any of the experimental descriptions. Without these values the experiments cannot be reproduced; please state the values used for all three datasets or explain if they are tuned.
  4. [Section 5.1 and Table 4] The sentence 'Latent Boost proves to consistently outperform the baseline and the classic Magnet loss results from the previous experiments of Table 2' is contradicted by the 'Latent Boost w/o PCA' row for Fashion MNIST, which reports accuracy 88.12 ± 0.29, below both the baseline (88.59 ± 0.15) and classic Magnet (89.52 ± 0.34). If the claim refers only to the full Latent Boost, the text should say so explicitly, and the failure of the PCA-free variant on Fashion MNIST should be discussed rather than ignored, especially since this variant isolates the effect of the variance and dynamic schedule modifications.
minor comments (6)
  1. [Equations (4) and (11)] The Magnet loss in Equation (4) and the Latent Boost loss in Equation (11) have ambiguous bracket structure: the exponent in the numerator and the placement of the sum over clusters in the denominator are not clearly delimited, and the meaning of μ_c^k and σ^2_{C_k} is not fully defined. Please add parentheses or a clarifying sentence describing which terms are summed and where the margin α is subtracted.
  2. [Section 1 and Abstract] The claim that Latent Boost is 'the first method to integrate distance metrics into the classification loss function' is an overstatement, as hybrid losses combining cross-entropy with center loss, triplet loss, or other metric objectives are common in the literature. Please soften this to something like 'to the best of our knowledge, the first to use the Magnet-style loss with dynamic per-cluster variance in a weighted classification objective.'
  3. [Section 6.3] The text says 'Latent Boost requires several normal epochs with only the probabilistic cross-entropy loss to form initial clusters', but the experimental setup section (Section 3.3) does not mention any warmup phase. Please clarify whether such a warmup was used, how many epochs it lasted, and how it is incorporated into the early-stopping and epoch-counting procedure.
  4. [Table 4] The heading 'Improvement (%)' does not indicate that negative values for epoch counts are favourable; consider using 'Epoch reduction (%)' or adding a note that negative values indicate fewer epochs, which is desirable.
  5. [General] There are several typographical and phrasing issues, e.g., 'Oppose to traditional training' in Figure 1, 'the hyperparameter λ' should be defined explicitly as a scalar in Equation (5), and some references (e.g., the citation for Van der Maaten & Hinton) have inconsistent formatting. A careful proofread is needed.
  6. [Section 5.2] The paper defines interpretable classification in terms of cluster separation, which is a sensible operationalization, but the connection between Silhouette score and human-understandable explanations is not argued in detail; a brief justification or a qualitative user-study reference would strengthen the interpretability claim.

Circularity Check

2 steps flagged · score 6.0 of 10

Interpretability claim is a self-scoring objective; reported accuracy gains use test-best λ selection.

  1. self definitional [Section 5.2, Equations (11) and (12)]
    "We define interpretable classification as one that achieves not only the correct assignment of discrete data points to their respective classes but also induces a structured organization of latent representations, where clusters corresponding to each class are maximally separated and distinct. ... To quantify the density of clusters and their separation from each other in the original dimension, we selected the Silhouette Score to measure the quality of the latent representation."

    Equation (11) defines L_LB as a softmax over intra-cluster distance -||r'_n - μ'_{r_n}||^2/(2σ^2_Ck) versus inter-cluster distances -||r'_n - μ'_ck||^2/(2σ^2_Ck)·β; minimizing it is exactly tightening within-class distances and pushing apart class centroids. Equation (12)'s Silhouette score s(i) = (b(i) - a(i))/max(a(i), b(i)) measures the same ratio: a(i) is the mean distance to same-cluster points and b(i) is the mean distance to the nearest other cluster. Thus the reported higher Silhouette scores are not an independent interpretability prediction; they are a direct check that the loss optimized its own target.

  2. fitted input called prediction [Section 5.1 and Table 4 caption]
    "Accuracy (↑), Micro-F1 Score (↑), and epoch duration (↓) between baseline (λ = 0), standard Magnet loss, the Latent Boost without PCA and the final Latent Boost with all features on the unseen test dataset (best λ selection); percentage improvement compares Latent Boost with the baseline for each metric."

    Section 5.1 selects λ per dataset from Table 3 by test accuracy: 'For Fashion MNIST, the best performance is achieved at λ = 0.75 ... on CIFAR-10, λ = 0.75 gives the highest accuracy ... On CIFAR-100, ... best results at λ = 0.5.' Table 4 then reports accuracy, Micro-F1, and epoch counts on that same 'unseen test dataset' under 'best λ selection' and computes the improvement percentages. Because λ was chosen to maximize the very test metrics being reported, the 2.56%, 2.98%, and 2.06% gains are the maximum of nine correlated evaluations rather than an unbiased out-of-sample estimate. The claim to 'consistently outperform' the baseline is therefore statistically forced by the selection procedure, not demonstrated as a prediction.

full rationale

The self-citations in the paper (refs. 2 and 19) are not load-bearing: they are contextual background and do not supply a uniqueness theorem, an ansatz, or a central premise. The core engineering contribution — a weighted-sum combination of a Magnet-style latent loss with cross-entropy — is not circular by itself. However, the interpretability claim is circular in a definitional sense: the paper defines interpretable classification as maximally separated and distinct clusters, trains with a loss whose explicit terms are intra-class compactness and inter-class separation, and then evaluates interpretability with the Silhouette score, which is a direct scalarization of exactly those two quantities. This makes the reported Silhouette improvements an expected consequence of the optimization objective rather than an independent validation. Additionally, the headline accuracy and convergence gains are reported after choosing λ on the test set, so the magnitude of the improvement is inflated by construction; a validation-based λ selection would be needed to support the 'consistently outperform' claim. These two issues affect the central empirical claims, but the method itself remains concrete and reproducible, so a score of 6 reflects partial circularity rather than a fully forced derivation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The main free parameters are the weighting lambda, the initial alpha offset, the PCA threshold, and the stability constant. The method relies on domain assumptions about hyper-spherical clusters and PCA preserving distance structure. No new physical or explanatory entities are introduced; the Latent Boost loss is a mathematical construction rather than an invented entity.

free parameters (5)
  • lambda = 0.75 (Fashion-MNIST, CIFAR-10), 0.5 (CIFAR-100)
    Weights the distance loss against cross-entropy; best value selected from the test set across 0.1 to 0.9.
  • alpha0 = not reported
    Controls the initial strength of the intra-cluster margin in the exponential decay schedule (Equation 10).
  • beta0 = 1.0 (stated as initialization)
    Initial inter-cluster balancing factor in the linear decay schedule (Equation 10).
  • T (PCA variance threshold) = 0.95
    Cumulative explained variance threshold for choosing the number of PCA components; set based on initial investigations.
  • epsilon = 1e-8
    Stability constant added to the denominator to prevent division by zero.
assumptions (4)
  • domain assumption Class-conditional latent representations are approximately hyper-spherical.
    The Magnet-based loss uses a single per-cluster variance and Euclidean centroid distances, which only fully describe spherical Gaussian-like clusters. Acknowledged as a limitation in Section 6.4.
  • domain assumption PCA preserves the distance information relevant to the loss.
    The loss is computed in a PCA-reduced subspace; if PCA discards the dimensions that separate classes, the loss will not properly structure the latent space (Section 4.1).
  • ad hoc to paper The proposed dynamic alpha and beta schedules are beneficial.
    The exponential decay of alpha and linear decay of beta are introduced without theoretical justification and appear to be tuned empirically (Section 4.3).
  • domain assumption Euclidean distance in the latent space is the appropriate similarity measure.
    All distance losses use Euclidean distances, implicitly assuming that Euclidean closeness in the latent space corresponds to semantic similarity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Interpretability Through Loss-Defined Classification Objective in Structured Latent Spaces." pith.science (2026). https://pith.science/paper/PXMFKXPD

@misc{pith2026241208515,
  author       = {Pith},
  title        = {Pith review of: Enhancing Interpretability Through Loss-Defined Classification Objective in Structured Latent Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXMFKXPD}},
  note         = {Machine review of arXiv:2412.08515}
}
read the original abstract

Supervised machine learning often operates on the data-driven paradigm, wherein internal model parameters are autonomously optimized to converge predicted outputs with the ground truth, devoid of explicitly programming rules or a priori assumptions. Although data-driven methods have yielded notable successes across various benchmark datasets, they inherently treat models as opaque entities, thereby limiting their interpretability and yielding a lack of explanatory insights into their decision-making processes. In this work, we introduce Latent Boost, a novel approach that integrates advanced distance metric learning into supervised classification tasks, enhancing both interpretability and training efficiency. Thus during training, the model is not only optimized for classification metrics of the discrete data points but also adheres to the rule that the collective representation zones of each class should be sharply clustered. By leveraging the rich structural insights of intermediate model layer latent representations, Latent Boost improves classification interpretability, as demonstrated by higher Silhouette scores, while accelerating training convergence. These performance and latent structural benefits are achieved with minimum additional cost, making it broadly applicable across various datasets without requiring data-specific adjustments. Furthermore, Latent Boost introduces a new paradigm for aligning classification performance with improved model transparency to address the challenges of black-box models.

Figures

Figures reproduced from arXiv: 2412.08515 by the authors.

Figure 1
Figure 1. Oppose to traditional training, relying on probabilistic loss only, Latent Boost injects distance metric information, obtained from the model’s hidden latent representations, as addition into the training through balanced weighted sum equations. 2 Related Work Distance Metric Learning has emerged as a crucial area in Machine Learning, offering a wide range of techniques aimed at improving performance in tasks of pri… view at source ↗
Figure 2
Figure 2. Flowchart of the Latent Boost approach, summarizing the mathematical steps to embed distance-metric information into the classic probabilistic training. 5.1 Improved Classification Performance To evaluate our approach, we conducted the same experiment, based on our three datasets and model combinations, with the adapted Latent Boost distance metric. The selected λ values mimic the range of the preliminary experiment… view at source ↗
Figure 3
Figure 3. Comparison of baseline (λ=0), standard Magnet loss, and our Latent Boost approach across the three experiment setups. Score s(i) is calculated following Equation (12). a(i) represents the cohesion as to how closely related a data point is to its own cluster, whereas b(i) represents the separation, meaning the distance between a data point to its nearest neighbor cluster. To evaluate the impact of different training … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 44 canonical work pages

  1. [1]

    & Bawany, N

    Qamar, T. & Bawany, N. Z. Understanding the black-box: towards interpretable and reliable deep learning models. PeerJ Comput. Sci. 9, e1629 (2023)

    work page 2023

  2. [2]

    & Kaiserslautern-Landau, R

    Geißler, D., Zhou, B., Lukowicz, P. & Kaiserslautern-Landau, R. Latent inspector: An interactive tool for probing neural network behaviors through arbitrary latent activation. In IJCAI, 7127–7130 (2023)

    work page 2023

  3. [3]

    Bello, H. et al. Towards certifiable ai in aviation: landscape, challenges, and opportunities.arXiv preprint arXiv:2409.08666 (2024)

    arXiv 2024

  4. [4]

    Making machine learning trustworthy

    Eshete, B. Making machine learning trustworthy. Science 373, 743 – 744, DOI: 10.1126/science.abi5052 (2021)

    work page doi:10.1126/science.abi5052 2021

  5. [5]

    ¸ SAHiN, E., Arslan, N. N. & Özdemir, D. Unlocking the black box: an in-depth review on interpretability, explainability, and reliability in deep learning. Neural Comput. Appl. 1–107 (2024)

    work page 2024

  6. [6]

    & van der Schaar, M

    Crabbé, J., Qian, Z., Imrie, F. & van der Schaar, M. Explaining latent representations with a corpus of examples. Adv. Neural Inf. Process. Syst. 34, 12154–12166 (2021)

    work page 2021

  7. [7]

    & Ommer, B

    Esser, P., Rombach, R. & Ommer, B. A disentangling invertible interpretation network for explaining latent representations. 2020 IEEE/CVF Conf. on Comput. Vis. Pattern Recognit. (CVPR) 9220–9229, DOI: 10.1109/cvpr42600.2020.00924 (2020)

    arXiv 2020

  8. [8]

    Lu, C., Wu, Y ., Hernández-Lobato, J. M. & Schölkopf, B. Invariant causal representation learning for out-of-distribution generalization. In International Conference on Learning Representations (2021)

    work page 2021

Show all 54 references
  1. [9]

    Rybakov, S., Lotfollahi, M., Theis, F. J. & Wolf, F. A. Learning interpretable latent autoencoder representations with annotations of feature sets. bioRxiv DOI: 10.1101/2020.12.02.401182 (2020)

    2020 doi

  2. [10]

    & Vargas, D

    Kotyan, S., Chen, P.-Y . & Vargas, D. V . Linking robustness and generalization: A k* distribution analysis of concept clustering in latent space for vision models. arXiv preprint arXiv:2408.09065 (2024)

    2024 arXiv

  3. [11]

    & Song, X

    Wu, T. & Song, X. Towards interpretable object detection by unfolding latent structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 6033–6043 (2019)

    2019

  4. [12]

    Kulis, B. et al. Metric learning: A survey. F oundations Trends Mach. Learn.5, 287–364 (2013)

    2013

  5. [13]

    & Bilge, H

    Kaya, M. & Bilge, H. S. Deep metric learning: A survey. Symmetry 11, 1066 (2019)

    2019

  6. [14]

    & Hart, P

    Cover, T. & Hart, P. Nearest neighbor pattern classification. IEEE transactions on information theory 13, 21–27 (1967)

    1967

  7. [15]

    Support-vector networks

    Cortes, C. Support-vector networks. Mach. Learn. (1995). 12/14

    1995

  8. [16]

    & Sun, J

    Wang, F. & Sun, J. Survey on distance metric learning and dimensionality reduction in data mining.Data mining knowledge discovery 29, 534–564 (2015)

    2015

  9. [17]

    K., Waiyamai, K., Konstantinos, S

    Lian, H. K., Waiyamai, K., Konstantinos, S. & Haruechaiyasak, C. Discriminative dimension selection for enhancing the interpretability and performance of clustering output. In 2024 16th International Conference on Knowledge and Smart Technology (KST), 178–183 (IEEE, 2024)

    2024

  10. [18]

    Zhou, Y . & He, D. Multi-target feature selection with adaptive graph learning and target correlations.Mathematics 12, 372, DOI: 10.3390/math12030372 (2024)

    2024 doi

  11. [19]

    & Lukowicz, P

    Geißler, D., Zhou, B., Liu, M., Suh, S. & Lukowicz, P. The power of training: How different neural network setups influence the energy demand. In International Conference on Architecture of Computing Systems , 33–47 (Springer, 2024)

    2024

  12. [20]

    A boosting approach for supervised mahalanobis distance metric learning

    Chang, C.-C. A boosting approach for supervised mahalanobis distance metric learning. Pattern Recognit. 45, 844–862, DOI: 10.1016/j.patcog.2011.07.026 (2012)

    2012 doi

  13. [21]

    Weinberger, K. Q. & Saul, L. K. Distance metric learning for large margin nearest neighbor classification. In Journal of Machine Learning Research, vol. 10, 207–244 (2009)

    2009

  14. [22]

    V ., Kulis, B., Jain, P., Sra, S

    Davis, J. V ., Kulis, B., Jain, P., Sra, S. & Dhillon, I. S. Information-theoretic metric learning. In 24th international conference on Machine learning, 209–216 (2007)

    2007

  15. [23]

    & LeCun, Y

    Hadsell, R., Chopra, S. & LeCun, Y . Dimensionality reduction by learning an invariant mapping. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1735–1742 (2006)

    2006

  16. [24]

    & Chen, K

    Wang, Q. & Chen, K. Zero-shot visual recognition via bidirectional latent embedding. Int. J. Comput. Vis. 124, 356–383 (2017)

    2017

  17. [25]

    & Shen, H

    Wang, B., Yang, Y ., Xu, X., Hanjalic, A. & Shen, H. T. Adversarial cross-modal retrieval. In 25th ACM international conference on Multimedia, 154–162 (2017)

    2017

  18. [26]

    Liu, W. et al. Sphereface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , 212–220 (2017)

    2017

  19. [27]

    & Hinton, G

    Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A simple framework for contrastive learning of visual representations. In Proceedings of the International Conference on Machine Learning (ICML) , 1597–1607 (2020)

    2020

  20. [28]

    Do, T.-T. et al. A theoretically sound upper bound on the triplet loss for improving the efficiency of deep distance metric learning. 2019 IEEE/CVF Conf. on Comput. Vis. Pattern Recognit. (CVPR)10396–10405, DOI: 10.1109/CVPR.2019.01065 (2019)

    2019

  21. [29]

    & Philbin, J

    Schroff, F., Kalenichenko, D. & Philbin, J. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 815–823 (IEEE, 2015)

    2015

  22. [30]

    & Leibe, B

    Hermans, A., Beyer, L. & Leibe, B. In defense of the triplet loss for person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 1–8 (2017)

    2017

  23. [31]

    & Singh, S

    Movshovitz-Attias, Y ., Toshev, A., Leung, T., Ioffe, S. & Singh, S. No fuss distance metric learning using proxies.2017 IEEE Int. Conf. on Comput. Vis. (ICCV) 360–368, DOI: 10.1109/ICCV .2017.47 (2017)

    2017 doi

  24. [32]

    & Bourdev, L

    Rippel, O., Paluri, M., Dollar, P. & Bourdev, L. Metric learning with adaptive density discrimination. arXiv preprint arXiv:1511.05939 (2015)

    2015 arXiv

  25. [33]

    Yu, L. et al. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 6982–6991 (2020)

    2020

  26. [34]

    & Zafeiriou, S

    Deng, J., Guo, J., Liu, T., Gong, M. & Zafeiriou, S. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16, 741–757 (Springer, 2020)

    2020

  27. [35]

    & Basu, S

    Bilenko, M. & Basu, S. Integrating constraints and metric learning in semi-supervised clustering. In Proceedings of the 21st International Conference on Machine Learning (ICML) , 11–18 (2004)

    2004

  28. [36]

    Ding, C. & Li, T. Adaptive dimension reduction using discriminant analysis and k-means clustering. In 24th international conference on Machine learning, 521–528 (2007)

    2007

  29. [37]

    Liu, G., Lin, Z. & Yu, Y . Robust subspace segmentation by low-rank representation. InProceedings of the 27th international conference on machine learning (ICML-10) , 663–670 (2010)

    2010

  30. [38]

    Lahoti, P. et al. Fairness without demographics through adversarially reweighted learning. Adv. neural information processing systems 33, 728–740 (2020). 13/14

    2020

  31. [39]

    & Douze, M

    Caron, M., Bojanowski, P., Joulin, A. & Douze, M. Deep clustering for unsupervised learning of visual features. In Proceedings of the European Conference on Computer Vision (ECCV) , 132–149 (2018)

    2018

  32. [40]

    Caron, M. et al. Unsupervised learning of visual features by contrasting cluster assignments. Adv. neural information processing systems 33, 9912–9924 (2020)

    2020

  33. [41]

    Lee, K.-H., Chen, X., Hua, G., Hu, H. & He, X. Stacked cross attention for image-text matching. In Proceedings of the European Conference on Computer Vision (ECCV), 201–216 (2018)

    2018

  34. [42]

    & LeCun, Y

    Chopra, S., Hadsell, R. & LeCun, Y . Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR’05) , vol. 1, 539–546 (IEEE, 2005)

    2005

  35. [43]

    Improved deep metric learning with multi-class n-pair loss objective

    Sohn, K. Improved deep metric learning with multi-class n-pair loss objective. In Advances in Neural Information Processing Systems (NeurIPS), 1857–1865 (2016)

    2016

  36. [44]

    & V ollgraf, R

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

    2017 arXiv

  37. [45]

    Krizhevsky, A. et al. Learning multiple layers of features from tiny images. utoronto.edu (2009)

    2009

  38. [46]

    & Zisserman, A

    Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  39. [47]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , 770–778 (2016)

    2016

  40. [48]

    Paszke, A. et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, 8024–8035 (Curran Associates, Inc., 2019)

    2019

  41. [49]

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

    2014 arXiv

  42. [50]

    & Hinton, G

    Van der Maaten, L. & Hinton, G. Visualizing data using t-sne. J. machine learning research 9 (2008)

    2008

  43. [51]

    Rousseeuw, P. J. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J. computational applied mathematics 20, 53–65 (1987)

    1987

  44. [52]

    M., Aliguliyev, R

    Bagirov, A. M., Aliguliyev, R. M. & Sultanova, N. Finding compact and well-separated clusters: Clustering using silhouette coefficients. Pattern Recognit. 135, 109144 (2023)

    2023

  45. [53]

    & Luma, A

    Januzaj, Y ., Beqiri, E. & Luma, A. Determining the optimal number of clusters using silhouette score as a data mining technique. Int. J. Online & Biomed. Eng. 19 (2023)

    2023

  46. [54]

    & Shao, J

    Du, P., Li, F. & Shao, J. Multi-agent reinforcement learning clustering algorithm based on silhouette coefficient. Neuro- computing 127901 (2024). Acknowledgements This work is supported by the European Union’s Horizon Europe research and innovation program (HORIZON-CL4-2021- ...

    2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.