OpenNovelty.org

One-Shot Exemplars for Class Grounding in Self-Supervised Learning

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Self-supervised learningOne-shot exemplarRepresentation learning
Abstract:

Self-Supervised Learning (SSL) has recently achieved remarkable progress by leveraging large-scale unlabeled data. However, SSL pretrains models without relying on human annotation, so it usually does not specify the class space. This inevitably weakens the effectiveness of the learned representation in most downstream tasks that have the intrinsic class structure. In this work, we introduce the new easy setting of One-Shot Exemplar Self-Supervised Learning (OSESSL), requiring only one instance annotation for each class. By introducing this extremely sparse supervision, OSESSL provides the minimum class information to guide the exploration of unlabeled data, achieving significant performance boosts with neglectable annotation cost (i.e., a complexity of O(1)\mathcal{O}(1) w.r.t. the sample size). In this OSESSL setting, we propose a simple yet effective framework that leverages the single-labeled exemplar to build the class-specific prototype for learning reliable representations from the huge unlabeled data. To this end, we also build a novel consistency regularization, which extends the sparse exemplar supervision into the decision boundaries, thus improving the robustness of the learned representation. Extensive experiments on real-world datasets clearly validate the reliability of this simple and practical setting. The proposed approach successfully outperforms the state-of-the-art methods, achieving gains of approximately 3% and 6% kk-NN accuracy on CIFAR-100 and ImageNet-100, respectively.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces One-Shot Exemplar Self-Supervised Learning (OSESSL), requiring exactly one labeled instance per class to guide representation learning from unlabeled data. It resides in the 'One-Shot Exemplar Learning' leaf of the taxonomy, which contains only two papers including this work. This positions the paper in a relatively sparse research direction within the broader 'Minimal Supervision Paradigms' branch, where most work requires either multiple shots per class or operates in semi-supervised settings with richer label budgets.

The taxonomy reveals several neighboring directions that contextualize this work. The 'Few-Shot Learning with Self-Supervised Pretraining' branch contains methods like SSL-ProtoNet and Robust Few-Shot that rely on prototypical networks but typically assume multiple examples per class. The 'Semi-Supervised Learning with Self-Supervision' branch addresses scenarios with larger labeled pools rather than the single-exemplar constraint. The sibling paper 'One-Shot Class Grounding' shares the one-shot setting but may differ in its grounding mechanism versus the prototype-based consistency regularization proposed here.

Among 30 candidates examined, none clearly refute the three main contributions: the OSESSL setting itself (10 candidates examined), the exemplar-guided prototype framework (10 candidates), and the interpolation consistency regularization (10 candidates). This suggests limited direct prior work addressing the exact one-shot exemplar regime with this combination of techniques. However, the search scope is modest—top-K semantic matches plus citation expansion—and does not exhaustively cover all prototype learning or consistency regularization literature, leaving open the possibility of related approaches in unexamined papers.

Based on the limited search, the work appears to occupy a relatively novel position at the intersection of extreme label scarcity and self-supervised learning. The taxonomy structure shows this is a less populated area compared to few-shot or semi-supervised branches. While the analysis covers 30 candidates and identifies no clear refutations, a more comprehensive search might reveal additional related methods, particularly in broader prototype learning or consistency regularization contexts beyond the minimal supervision paradigm.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: self-supervised learning with minimal class-level supervision. The field encompasses a diverse set of strategies for learning robust representations when labeled data is scarce or expensive to obtain. At the top level, the taxonomy organizes work into several major branches: Few-Shot Learning with Self-Supervised Pretraining leverages unlabeled data to improve generalization from a handful of examples per class; Semi-Supervised Learning with Self-Supervision blends limited labels with large unlabeled pools; Weakly Supervised and Pseudo-Labeling Methods exploit noisy or indirect supervision signals; Contrastive Self-Supervised Learning builds discriminative features through instance-level comparisons; Cross-Domain and Domain Adaptation Methods address distribution shifts with minimal target labels; Meta-Learning and Knowledge Distillation transfer learned inductive biases or compressed knowledge; Minimal Supervision Paradigms push the envelope on extremely low-label regimes; and Specialized Self-Supervised Applications tailor these ideas to domain-specific challenges such as medical imaging, remote sensing, or audio analysis. Representative works like S4L[20] and Boosting Few-Shot[7] illustrate how self-supervision can bootstrap few-shot classifiers, while studies such as Medical Image Segmentation[4] and DINOv2 Medical[19] demonstrate the impact of pretraining in specialized domains. A particularly active line of inquiry centers on one-shot and few-shot scenarios where only a single or very few labeled examples per class are available. Within the Minimal Supervision Paradigms branch, One-Shot Class Grounding[0] exemplifies this extreme setting by grounding new classes from a single exemplar, closely aligning with Class-Cognizant Few-Shot[47], which also emphasizes class-aware mechanisms under severe label scarcity. These approaches contrast with methods in the Few-Shot Learning with Self-Supervised Pretraining branch, such as SSL-ProtoNet[27] and Robust Few-Shot[22], which rely on richer pretraining or prototypical networks to handle distribution variability. Meanwhile, works like UNSURE[3] and Twin Class Distribution[10] explore uncertainty quantification and distribution modeling to improve reliability when supervision is minimal. The landscape reveals an ongoing tension between leveraging large-scale unlabeled pretraining versus designing architectures and learning objectives that can extract maximal signal from just one or a few labeled instances, with One-Shot Class Grounding[0] situated at the frontier of this minimal-label regime.

Claimed Contributions

One-Shot Exemplar Self-Supervised Learning (OSESSL) setting

The authors propose a new learning paradigm that requires only a single labeled instance per class to guide self-supervised learning. This setting provides minimal class information with negligible annotation cost (O(1) complexity with respect to sample size) while enabling the model to learn semantically grounded representations.

10 retrieved papers
Exemplar-guided prototype learning framework

The authors develop a framework that constructs class-specific prototypes by augmenting each labeled exemplar with discriminative neighbors from unlabeled data. This approach ensures prototypes are both semantically aligned with true classes and representative of the data distribution, enabling effective knowledge transfer to unlabeled samples.

10 retrieved papers
Exemplar-guided interpolation consistency regularization

The authors introduce a consistency regularization mechanism that propagates exemplar supervision into interpolated decision boundaries. This regularization smooths decision boundaries in uncertain regions and enforces consistency between mixed views and original views, thereby enhancing representation robustness.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

One-Shot Exemplar Self-Supervised Learning (OSESSL) setting

The authors propose a new learning paradigm that requires only a single labeled instance per class to guide self-supervised learning. This setting provides minimal class information with negligible annotation cost (O(1) complexity with respect to sample size) while enabling the model to learn semantically grounded representations.

Contribution

Exemplar-guided prototype learning framework

The authors develop a framework that constructs class-specific prototypes by augmenting each labeled exemplar with discriminative neighbors from unlabeled data. This approach ensures prototypes are both semantically aligned with true classes and representative of the data distribution, enabling effective knowledge transfer to unlabeled samples.

Contribution

Exemplar-guided interpolation consistency regularization

The authors introduce a consistency regularization mechanism that propagates exemplar supervision into interpolated decision boundaries. This regularization smooths decision boundaries in uncertain regions and enforces consistency between mixed views and original views, thereby enhancing representation robustness.

Table of Contents

One-Shot Exemplars for Class Grounding in Self-Supervised Learning | Novelty Validation