pith. sign in

arxiv: 2604.27564 · v1 · submitted 2026-04-30 · 💻 cs.LG

Learning from a single labeled face and a stream of unlabeled data

Branislav Kveton , Michal Valko This is my paper

Pith reviewed 2026-05-07 09:26 UTC · model grok-4.3

classification 💻 cs.LG
keywords one-class classificationface recognitionsingle labeled imageunlabeled data streamnon-parametric modelauthenticationmachine learning
0
0 comments X

The pith

A non-parametric one-class model from one labeled face image and an unlabeled data stream achieves 90 percent recall at near-zero false positives.

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

The paper examines the problem of recognizing a single person from one labeled image when no labeled examples of any other person exist. It treats the task as one-class classification and develops an algorithm that builds a non-parametric model of the target face by combining the single labeled image with a stream of unlabeled images presumed to come mostly from the same person. This setting matches real device-authentication scenarios where negatives are unavailable yet camera data arrives continuously. The resulting method is tested on images of 43 individuals and reaches 90 percent recall with almost no false positives, exceeding the strongest baseline by more than 25 percent.

Core claim

We formalize single-person face recognition with one labeled image and no negatives as a one-class classification problem. We propose and analyze an algorithm that learns a non-parametric model of the target face from the labeled image plus a stream of unlabeled data. On a dataset of 43 people the method recognizes the target 90 percent of the time at nearly zero false positives, a gain of more than 25 percent over the best baseline. A full sensitivity study supplies practical rules for choosing the algorithm's parameters.

What carries the argument

Non-parametric model of the target face distribution that integrates one labeled image with the unlabeled data stream to estimate the positive class without any negative examples.

If this is right

  • Device authentication systems can use everyday camera streams to improve single-image recognition without collecting labeled negatives.
  • Parameter guidelines from the sensitivity analysis let practitioners tune the model for different data volumes and quality.
  • The same one-class construction may transfer to other biometrics or anomaly-detection tasks that receive continuous unlabeled streams.

Where Pith is reading between the lines

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

  • Continuous online updating of the model as fresh unlabeled frames arrive could let the recognizer adapt to gradual changes in appearance.
  • The one-class stream approach may generalize to voice or gait authentication on personal devices where negatives are likewise scarce.
  • Measuring how performance scales with the fraction of target images inside the unlabeled stream would identify safe operating regimes for deployment.

Load-bearing premise

The unlabeled stream contains mostly images of the target person, so the positive distribution can be estimated reliably without any negative examples from other people.

What would settle it

Apply the algorithm to a version of the dataset in which the unlabeled stream is deliberately contaminated with many faces from other people and check whether recall falls below the baseline.

Figures

Figures reproduced from arXiv: 2604.27564 by Branislav Kveton, Michal Valko.

Figure 1
Figure 1. Figure 1: An illustration of the face manifold tracked by OMT. view at source ↗
Figure 2
Figure 2. Figure 2: Images and faces in the VidTIMIT dataset. view at source ↗
Figure 3
Figure 3. Figure 3: Representative faces learned by OMT for Person 1, 15, 22, and 42. The four leftmost faces are the labeled examples. view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of the NN and OMT recognizers that are view at source ↗
Figure 6
Figure 6. Figure 6: Varying the generalization radius R in OMT. For each value R, we report the ROC curve, the computation time, and the cover radius r. TPR for R = 0.25, many of these positives can be classified correctly at nearly zero false positives. So the generalization radius of R = 0.25 is too restrictive. At low FPRs, the TPR for R = 0.3 is higher than the TPR for R = 0.35. This trend can be explained as follows. Bar… view at source ↗
read the original abstract

Face recognition from a single image per person is a challenging problem because the training sample is extremely small. We consider a variation of this problem. In our problem, we recognize only one person, and there are no labeled data for any other person. This setting naturally arises in authentication on personal computers and mobile devices, and poses additional challenges because it lacks negative examples. We formalize our problem as one-class classification, and propose and analyze an algorithm that learns a non-parametric model of the face from a single labeled image and a stream of unlabeled data. In many domains, for instance when a person interacts with a computer with a camera, unlabeled data are abundant and easy to utilize. This is the first paper that investigates how these data can help in learning better models in the single-image-per-person setting. Our method is evaluated on a dataset of 43 people and we show that these people can be recognized 90% of time at nearly zero false positives. This recall is 25+% higher than the recall of our best performing baseline. Finally, we conduct a comprehensive sensitivity analysis of our algorithm and provide a guideline for setting its parameters in practice.

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.

Referee Report

1 major / 1 minor

Summary. The manuscript formalizes single-person face recognition as a one-class classification problem and proposes a non-parametric algorithm that builds a model from one labeled image plus an unlabeled data stream. On a 43-subject dataset it reports 90% recall at near-zero false positives, a 25%+ gain over the strongest baseline, together with a parameter-sensitivity study that supplies practical guidelines.

Significance. If the central performance claims survive scrutiny, the work would be useful for authentication settings (personal devices, cameras) where negative examples are unavailable and unlabeled interaction data are plentiful. The non-parametric formulation and explicit use of the unlabeled stream constitute a concrete, falsifiable contribution to the single-image-per-person regime.

major comments (1)
  1. [Evaluation section and §3] Evaluation section (and the one-class formulation in §3): the reported 90% recall at near-zero FP is obtained under the assumption that the unlabeled stream is dominated by the target face. No controlled experiments with varying contamination fractions from other identities are presented, so it is impossible to assess whether the decision threshold (chosen without negative examples) remains stable when the stream contains non-negligible impostor faces. This directly affects the reliability of both the headline numbers and the claimed 25% improvement.
minor comments (1)
  1. [Abstract] The abstract states that a 'comprehensive sensitivity analysis' was performed; the manuscript should explicitly state whether this analysis includes contamination rates or only the algorithm's internal parameters.

Simulated Author's Rebuttal

1 responses · 0 unresolved

Thank you for the thorough review and the valuable feedback on our manuscript. We address the major comment point by point below.

read point-by-point responses

  1. Referee: [Evaluation section and §3] Evaluation section (and the one-class formulation in §3): the reported 90% recall at near-zero FP is obtained under the assumption that the unlabeled stream is dominated by the target face. No controlled experiments with varying contamination fractions from other identities are presented, so it is impossible to assess whether the decision threshold (chosen without negative examples) remains stable when the stream contains non-negligible impostor faces. This directly affects the reliability of both the headline numbers and the claimed 25% improvement.

    Authors: We thank the referee for pointing out this limitation in our evaluation. Our work focuses on the practical setting of personal device authentication, where the unlabeled data stream is expected to be dominated by the target user's face images due to frequent interactions with the device owner. The non-parametric one-class approach is intended for scenarios lacking negative examples. Nevertheless, we acknowledge that evaluating performance under varying degrees of contamination would strengthen the claims. In the revised version, we will add controlled experiments that introduce different fractions of impostor faces into the unlabeled stream and analyze the stability of the automatically chosen decision threshold. We will also discuss how these results affect the reported improvements over baselines. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper formalizes single-image-per-person face recognition as a one-class classification problem and proposes a non-parametric model learned from one labeled image plus an unlabeled data stream. The central claims rest on an empirical evaluation across 43 subjects that reports 90% recall at near-zero false positives (25% above the best baseline), with an accompanying sensitivity analysis for parameter settings. No load-bearing step reduces by construction to a fitted input renamed as prediction, a self-citation chain, or an ansatz smuggled via prior work; the method description and results are presented as independent of the target quantities they claim to predict.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the non-parametric model and one-class formulation are mentioned at high level without details on any fitted quantities or background assumptions.

pith-pipeline@v0.9.0 · 5498 in / 1123 out tokens · 59881 ms · 2026-05-07T09:26:38.431964+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

  1. [1]

    Face recognition from a single image per person: A survey,

    X. Tan, S. Chen, Z.-H. Zhou, and F. Zhang, “Face recognition from a single image per person: A survey,”Pattern Recognition, vol. 39, no. 9, pp. 1725–1745, 2006

    work page 2006

  2. [2]

    Face recognition: A literature survey,

    W.-Y . Zhao, R. Chellappa, P. Phillips, and A. Rosenfeld, “Face recognition: A literature survey,”ACM Computing Surveys, vol. 35, no. 4, pp. 399–458, 2003

    work page 2003

  3. [3]

    Face recognition using Laplacianfaces,

    X. He, S. Yan, Y . Hu, P. Niyogi, and H. Zhang, “Face recognition using Laplacianfaces,”IEEE Transactions Pattern Analysis and Machine Intelligence, vol. 27, no. 3, pp. 328–340, 2005

    work page 2005

  4. [4]

    Eigenfaces vs. Fish- erfaces: Recognition using class specific linear projection,

    P. Belhumeur, J. Hespanha, and D. Kriegman, “Eigenfaces vs. Fish- erfaces: Recognition using class specific linear projection,”IEEE Transactions Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp. 711–720, 1997

    work page 1997

  5. [5]

    Face recognition using eigenfaces,

    M. Turk and A. Pentland, “Face recognition using eigenfaces,” inIEEE Conference on Computer Vision and Pattern Recognition, 1991, pp. 586–591

    work page 1991

  6. [6]

    One-class classification,

    D. Tax, “One-class classification,” Ph.D. dissertation, Tu Delft, 2001

    work page 2001

  7. [7]

    Quantization,

    R. Gray and D. Neuhoff, “Quantization,”IEEE Transactions on Information Theory, vol. 44, no. 6, pp. 2325–2383, 1998

    work page 1998

  8. [8]

    Incremental clustering and dynamic information retrieval,

    M. Charikar, C. Chekuri, T. Feder, and R. Motwani, “Incremental clustering and dynamic information retrieval,” inProceedings of the 29th Annual ACM Symposium on Theory of Computing, 1997, pp. 626–635

    work page 1997

  9. [9]

    Cover trees for nearest neighbor,

    A. Beygelzimer, S. Kakade, and J. Langford, “Cover trees for nearest neighbor,” inProceedings of the 23rd International Conference on Machine Learning, 2006, pp. 97–104

    work page 2006

  10. [10]

    Online semi-supervised learning on quantized graphs,

    M. Valko, B. Kveton, L. Huang, and D. Ting, “Online semi-supervised learning on quantized graphs,” inProceedings of the 26th Conference on Uncertainty in Artificial Intelligence, 2010

    work page 2010

  11. [11]

    Person identification in webcam images: An application of semi-supervised learning,

    M.-F. Balcan, A. Blum, P. P. Choi, J. Lafferty, B. Pantano, M. R. Rwebangira, and X. Zhu, “Person identification in webcam images: An application of semi-supervised learning,” inICML 2005 Workshop on Learning with Partially Classified Training Data, 2005

    work page 2005

  12. [12]

    Semi-supervised learning using Gaussian fields and harmonic functions,

    X. Zhu, Z. Ghahramani, and J. Lafferty, “Semi-supervised learning using Gaussian fields and harmonic functions,” inProceedings of the 20th International Conference on Machine Learning, 2003, pp. 912– 919

    work page 2003

  13. [13]

    Semi-supervised learning with max-margin graph cuts,

    B. Kveton, M. Valko, A. Rahimi, and L. Huang, “Semi-supervised learning with max-margin graph cuts,” inProceedings of the 13th International Conference on Artificial Intelligence and Statistics, 2010, pp. 421–428

    work page 2010

  14. [14]

    Multi-region probabilistic histograms for robust and scalable identity inference,

    C. Sanderson and B. Lovell, “Multi-region probabilistic histograms for robust and scalable identity inference,” inProceedings of the 3rd International Conferences on Advances in Biometrics, 2009, pp. 199– 208

    work page 2009

  15. [15]

    The OpenCV Library,

    G. Bradski, “The OpenCV Library,”Dr. Dobb’s Journal of Software Tools, 2000

    work page 2000

  16. [16]

    Face recognition with one training image per person,

    J. Wu and Z.-H. Zhou, “Face recognition with one training image per person,”Pattern Recognition Letters, vol. 49, no. 14, pp. 1711–1719, 2002

    work page 2002

  17. [17]

    Face recognition from one example view,

    D. Beymer and T. Poggio, “Face recognition from one example view,” inProceedings of the 5th International Conference on Computer Vision, 1995, pp. 500–507

    work page 1995

  18. [18]

    Semi-supervised on-line boosting for robust tracking,

    H. Grabner, C. Leistner, and H. Bischof, “Semi-supervised on-line boosting for robust tracking,” inProceedings of the 10th European Conference on Computer Vision, 2008, pp. 234–247

    work page 2008

  19. [19]

    Online manifold regularization: A new learning setting and empirical study,

    A. Goldberg, M. Li, and X. Zhu, “Online manifold regularization: A new learning setting and empirical study,” inProceeding of European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, 2008

    work page 2008