PaperlayerView exact revision on arXiv

Source-linked AI summary

Geometric Regularization for Long-Tailed Semi-Supervised Learning via Gaussian Feature Bridges

Hongyang He, Xinyuan Song, Yan Zhong, Daizong Liu, Yanbin Li, Yang-fan He, Wenqiao Zhang

arXiv:2608.20710v1cs.LG

TL;DR

Long-tailed SSL lacks stable supervision when pseudo-labels are noisy and biased toward majority classes. GBC builds Gaussian bridges from unlabeled features to reliable class anchors and adds confidence-guided BridgeMix; the paper reports robust gains under realistic long-tailed SSL settings.

  • Problem

    Long-tailed SSL must handle class imbalance, noisy pseudo-labels, and confirmation bias that destabilize learning, particularly for tail classes.

  • Method

    GBC uses a dynamic Prototype Atlas, class-conditional Gaussian Feature Bridges, bridge consistency targets, and confidence-guided mixing of sample-anchor pairs.

  • Results

    The authors report consistent and significant gains over prior state-of-the-art methods under severe long-tailed conditions, including improved tail-class performance.

  • Takeaways & Limitations

    GBC frames SSL as geometric transport from uncertain features to reliable prototypes, combining empirical robustness with a probabilistic-geometry perspective.

  • Takeaways & Limitations

    The semantic trajectory analysis assumes a monotonically increasing interpolation function g(t) with endpoints g(0)=0 and g(1)=1, while the implemented path uses g(t)=t and truncates t to [0.2, 0.8].

Abstract

from arXiv · show

Real-world semi-supervised learning (SSL) often encounters significant challenges with long-tailed label distributions and noisy pseudo-labels, which hinder generalization and amplify confirmation bias. In this work, we introduce a novel framework, Gaussian Bridge Consistency (GBC), to address these challenges by constructing semantic interpolation paths between unlabeled samples and high-quality class anchors. Our method maintains a dynamic Prototype Atlas that stores a diverse and evolving set of labeled and pseudo-labeled exemplars per class. For each unlabeled instance, GBC forms a class-conditional Gaussian Feature Bridge in the latent space, enabling the student model to traverse a smooth trajectory from uncertain predictions to reliable class prototypes. A bridge consistency loss is applied along this path to enforce alignment with a geometrically interpolated target distribution. Furthermore, we propose BridgeMix, a confidence-aware feature mixing strategy that interpolates both sample and anchor pairs to amplify cross-sample generalization. Extensive experiments on CIFAR10-LT and ImageNet-LT (USB benchmarks) validate the robustness and effectiveness of GBC under realistic long-tailed SSL settings, consistently improving long tail-class performance without sacrificing scalability.

1 Introduction

Long-tailed SSL is destabilized by biased pseudo-labels and confirmation bias, especially for tail classes. GBC addresses this by transporting uncertain features toward reliable anchors through Gaussian bridges and confidence-guided mixing.

  • Long-tailed SSL suffers from noisy pseudo-labels, majority-class bias, and undertraining or semantic drift in tail classes.
  • GBC constructs class-conditional interpolation paths from unlabeled features to class anchors stored in a dynamic Prototype Atlas.
  • Bridge consistency aligns student predictions at intermediate bridge points with geometrically interpolated soft targets, guiding features from uncertainty toward reliable representations.
  • BridgeMix interpolates samples and their anchors using pseudo-label confidence, allowing confident examples to guide uncertain ones without changing the underlying loss design.
  • The framework is theoretically presented as a geometric regularizer that smooths decision boundaries and tightens generalization through reduced hypothesis complexity and distributional discrepancy.
  • Experiments report consistent and significant gains over prior state-of-the-art methods under severe long-tailed conditions.

2 Related Works

Prior SSL methods rely on pseudo-labeling and consistency regularization but commonly assume balanced data. Long-tailed SSL and interpolation methods address parts of the problem, while GBC targets semantic alignment along uncertain sample-to-anchor paths.

  • Semi-Supervised Learning: Classic SSL combines pseudo-labeling and consistency regularization, but pseudo-labels become error-prone and head-class biased in low-label or long-tailed settings.
  • Long-Tailed SSL: Long-tailed SSL methods use reweighting, calibration, refinement, debiasing, or expertization to reduce imbalance and pseudo-label errors.
  • Interpolation-Based Learning and Consistency: Feature interpolation methods such as MixUp and Manifold MixUp regularize training signals but lack an explicit mechanism for aligning uncertain unlabeled features semantically.
  • Interpolation-Based Learning and Consistency: GBC addresses this gap by constructing a Gaussian path between unlabeled samples and class anchors.

3 Method

GBC regularizes realistic long-tailed SSL through class-conditional Gaussian feature bridges, residual feature fusion, bridge-consistency targets, and confidence-guided BridgeMix. Its unified objective combines supervised, unsupervised, and bridge losses.

  • Framework Overview: GBC addresses class-distribution mismatch and noisy pseudo-label drift by enforcing consistency along paths between unlabeled samples and Prototype Atlas anchors.
  • PA Selection and Maintenance: The Prototype Atlas stores class-indexed labeled and high-confidence pseudo-labeled features, while size and cosine-distance constraints preserve quality and diversity.
  • Gaussian Feature Bridging in Latent Space: A Gaussian bridge interpolates an unlabeled feature with a same-class anchor after the teacher confidence exceeds a class-dependent threshold.
  • Gaussian Feature Bridging in Latent Space: The semantic trajectory uses g(t)=t, representing a constant-speed transition from the sample feature to the class anchor, with t truncated to [0.2, 0.8].
  • Bridging and Forward Propagation: Residual fusion injects the bridge feature into the student representation, with ω(t)=4t(1−t) emphasizing mid-bridge states where uncertainty is highest.
  • Bridge Consistency Objective: Bridge consistency trains the student prediction against a geometrically interpolated target distribution along the bridged path.
  • BridgeMix: BridgeMix confidence-guidedly interpolates paired sample features and anchor features, allowing more confident pseudo-labels to guide less certain ones.
  • Final Learning Objective: The final objective combines supervised, standard unsupervised, and bridge consistency losses, with β warmed up during early training.

4 Theoretical Analysis

GBC regularizes SSL geometrically by enforcing smooth prediction transitions along Gaussian feature bridges, while BridgeMix reduces complexity and distributional discrepancy through confidence-guided interpolation. The analysis links these mechanisms to pathwise convergence, local Lipschitz continuity, smoother decision boundaries, and tighter generalization bounds.

  • Geometric regularization: GBC constructs smooth transitions from uncertain unlabeled features to reliable class anchors, producing continuous predictions and stable decision boundaries.BridgeMix complements this feature-level regularization with confidence-guided interpolation between samples and anchors.
  • Bridge construction: Under a monotone bridge function g(t), the feature path moves continuously from the unlabeled sample at t=0 to the class anchor at t=1.The practical implementation sets g(t)=t, yielding linear semantic interpolation in feature space.
  • Pathwise convergence: As Lbridge approaches zero, student predictions converge to a continuous semantic trajectory between uncertain-sample and reliable-anchor predictions.The convergence statement is expressed as the L1 distance between pθ(f_t) and the target path tending to zero for all t∈[0,1].
  • Local stability: Minimizing Lbridge enforces local Lipschitz continuity around the unlabeled sample and anchor, limiting prediction changes under bounded feature perturbations.The bound combines a perturbation term Lδ with an O(Lbridge^1/2) term.
  • Decision-boundary stability: Decreasing Lbridge bounds expected decision-boundary curvature and flattens the boundary, improving robustness to feature noise.The curvature bound is C0 + C1 Lbridge^1/2.
  • Generalization: BridgeMix tightens the true-risk upper bound through empirical risk, Rademacher complexity, sampling error, and distributional discrepancy terms.The stated corollary replaces the discrepancy term with εmix when Δ(D,DBM)≤εmix, while the theory attributes gains to reduced hypothesis complexity and discrepancy.

5 Evaluation results

GBC is evaluated across long-tailed SSL benchmarks, imbalance regimes, feature-injection locations, and ImageNet scales. It performs particularly strongly under imbalanced or shifted distributions, improves tail-class representation quality, and reaches strong ImageNet accuracy while using deep feature layers for bridging.

  • Imbalance regimes: 0.94% and 1.52% gains over SimPro occur in reversed and head-tail settings, respectively, where pseudo-label noise and distribution mismatch are more severe.GBC is reported as best or competitive across most unlabeled class distributions.
  • Imbalance regimes: GBC consistently outperforms feature-mixing alternatives under consistent, reversed, middle, and head-tail imbalance, while SimPro+Manifold MixUp slightly wins under uniform data.The comparison attributes the imbalanced-setting advantage to class-conditional prototypes rather than class-agnostic interpolation.
  • Large-scale benchmarks: 68.6% Top-1 accuracy at 64×64 establishes GBC as a new SOTA on ImageNet-127.At 32×32, GBC reaches 61.9%, surpassing Meta-Expert by 1.6% and SimPro by 2.8%.
  • Large-scale benchmarks: 27.5% Top-1 accuracy on 64×64 ImageNet-1K improves over Meta-Expert by 2.1% and over SimPro by 2.5%.These results support scalability on the full ImageNet-1K dataset.
  • Ablation and representation quality: 92.3% Top-1 accuracy is achieved when the Gaussian Feature Bridge is injected at the penultimate ResNet-50 layer, Stage 4.Stage 2 injection reduces performance by approximately 4.5%, supporting deep semantic feature bridging.
  • Ablation and representation quality: GBC produces more compact, better-separated clusters and larger gains for Medium and Few classes than for Many classes.The t-SNE pattern and Many/Medium/Few breakdown associate bridge regularization with improved tail robustness under imbalance.

6 Ablation Study

The ablations show that bridge consistency, stochasticity, class reweighting, BridgeMix, and calibrated hyperparameters jointly affect robustness, while training remains computationally scalable. Sensitivity analyses also reveal dataset-dependent confidence-threshold trade-offs and stable optimization under distribution shift.

  • Distribution-shift ablations: Mean teacher–student KL divergence decreases monotonically across all unlabeled-distribution regimes, indicating progressive alignment without oscillatory dynamics.The reversed curve converges toward the other regimes after approximately 300 epochs in the consistent setting.
  • Distribution-shift ablations: The reversed regime starts with the largest KL divergence but rapidly reaches the lowest level, showing strong performance under distribution shift.Its steep decay contrasts with the slower consistent curve, which stabilizes at a higher divergence.
  • Confidence-threshold sensitivity: CIFAR10-LT accuracy generally increases with τc, whereas CIFAR100-LT performance peaks at τc ≈0.2–0.3 and then declines.The CIFAR100-LT pattern reflects a precision–coverage trade-off as stricter thresholds reduce anchor recall.
  • Component ablations: Removing Lbridge drops accuracy from 92.3% to 85.1%, identifying path-level consistency as the most important component.Removing Gaussian noise lowers accuracy to 90.2%, while hard replacement performs worse than residual fusion.
  • Component ablations: Eliminating class reweighting causes accuracy to decline to 80.5%, highlighting the necessity of tail-aware balancing.BridgeMix consistently improves results across distributions, while a moderate Prototype Atlas size provides the best stability.
  • Hyperparameter sensitivity: The noise level ν = 0.10, Beta parameter α = 2.0, and bridge weight β = 0.75 provide favorable hyperparameter settings.The bridge weight slightly declines beyond 0.75, indicating that bridge alignment remains complementary to the main SSL objective.
  • Computational overhead: GBC increases parameters from 25.6M to 26.1M, raises peak memory by 3%, and increases runtime by only +1.7%.The reported overhead is characterized as negligible while accuracy gains are maintained.

7 Conclusion

GBC reframes semi-supervised learning as geometric transport between uncertain features and reliable class prototypes. The conclusion positions this probabilistic-geometric perspective as a basis for future manifold regularization and distributional interpolation research.

  • Conclusion: GBC reframes semi-supervised learning as geometric transport between uncertain features and reliable class prototypes.It presents noisy pseudo-labeling as a stable representation flow under long-tailed and noisy conditions.
  • Conclusion: The framework connects optimal transport and consistency learning through a probabilistic-geometry perspective.The paper identifies dynamic manifold regularization, multimodal and temporal interpolation, and continuous-time consistency fields as future directions.

A Theoretical Analysis

The analysis presents GBC as a geometric regularizer and BridgeMix as a distribution-level regularizer, linking smooth feature paths and confidence-guided interpolation to robustness and generalization.

  • Geometric regularization: GBC enforces smooth transitions from uncertain unlabeled features to reliable anchors, yielding continuous predictions and stable decision boundaries.The framework operates at the feature level by regularizing semantic trajectories in latent space.
  • Pathwise convergence: Under the bridge-consistency setting, minimizing Lbridge yields convergence to a continuous semantic trajectory between uncertain and reliable anchor predictions.The convergence statement assumes continuously differentiable predictions and targets along the bridge.
  • Perturbation stability: Minimizing Lbridge enforces local Lipschitz continuity, bounding prediction variation around unlabeled and anchor features.The bound applies under a gradient-norm assumption along the bridge path.
  • Decision-boundary stability: Decreasing Lbridge bounds expected decision-boundary curvature and thereby improves robustness to feature noise.The curvature bound has the form C0 + C1 Lbridge^1/2.
  • Distribution-level generalization: BridgeMix tightens the true-risk bound by reducing effective hypothesis complexity and distributional discrepancy between mixed and empirical distributions.The result is formalized through a generalization bound involving empirical risk, Rademacher complexity, concentration, and distributional discrepancy.

B Proof of Lemma 4- Lemma 6

The proofs establish that bridge consistency produces pathwise convergence, local prediction stability, and smoother decision boundaries under the stated regularity assumptions.

  • Lemma 4: Pathwise convergence: As Lbridge approaches zero, student predictions converge to a continuous trajectory matching the geometric target path across bridge time.The proof uses KL-to-L1 control and continuity of qt to establish pointwise convergence.
  • Lemma 5: Local Lipschitz continuity: The bridge loss bounds prediction changes under feature perturbations by a term involving the perturbation size and bridge-consistency error.The proof combines a gradient bound with Pinsker’s inequality.
  • Lemma 6: Decision-boundary stability: The expected curvature of the decision boundary is bounded by C0 + C1 Lbridge^1/2.The proof controls the margin Hessian and gradient along the bridge, with Gaussian injection smoothing the loss.

C Proof of Theorem 2

Theorem 2 analyzes BridgeMix through empirical-process bounds, showing that confidence-guided interpolation need not increase complexity and can control distributional discrepancy under anchor-proximity conditions.

  • Complexity control: Mixing does not increase empirical complexity, and it is typically smaller when anchors are less variable or closer to class prototypes.Convexity and affine dependence on soft labels yield the complexity comparison.
  • Discrepancy control: The distributional discrepancy is small when mixed points lie near anchors, corresponding to high confidence and small class spread.Under Lipschitz loss, the discrepancy is controlled by Lz E[1 − Λ]ρ.
  • Corollary 2: Under Δ(D, DBM) ≤ εmix, the true-risk bound replaces the discrepancy term with εmix.This corollary provides an explicit bound when the mixed distribution remains close to the true distribution.
  • BridgeMix construction: The practical construction interpolates sample-anchor pairs using confidence-guided mixing within the original loss formulation.The implementation uses feature mixing, bridge loss weighting, and no additional loss term.

E Additional experiments

Additional experiments show that GBC transfers across architectures and data regimes, improves few-shot and long-tailed performance, remains robust to noise and atlas capacity, and adds less than 2% computational overhead.

  • Backbone adaptation: +2.8 pp is the largest reported CIFAR10-LT gain with ResNet-50, while ViT-B/16 and Vision Mamba-B gain +1.9 and +2.0 pp.The backbone study reports consistent gains across architectures and datasets.
  • Backbone adaptation: GBC maintains gains of +1.1–1.4 pp on ImageNet-127 and up to +2.5 pp on ImageNet-1K at 642.These results extend the observed improvements to large-scale, high-variance SSL benchmarks.
  • Few-shot performance: +0.5–0.9 pp improvements are reported across CIFAR10-LT, CIFAR100-LT, and ImageNet-127 in few-shot settings.The benefits are most pronounced in the 1-shot regime and retain comparable variance.
  • Noise ablation: Moderate Gaussian noise improves KL divergence and anchor recall, while ν=0 remains deterministic but slightly suboptimal.The ablation identifies a modest noise level as the most reliable supervision regime.
  • Efficiency: GBC introduces less than 2% computational overhead compared with FixMatch.The wall-clock comparison evaluates average epoch time and total training time on CIFAR10-LT.
  • Prototype Atlas ablation: A Prototype Atlas capacity of C=64 achieves the best trade-off between representation diversity and stability.Disabling the atlas leaves only EMA prototypes and produces insufficient class diversity and lower accuracy.

F Compute scalability

GBC adds minimal computational overhead while retaining substantial accuracy gains, supporting integration of semantic feature-space bridging into existing SSL pipelines.

  • F Compute scalability: 42.8s per epoch represents only a +1.7% overhead compared to FixMatch on CIFAR10-LT.The comparison uses identical training settings on a single NVIDIA A100 GPU.
  • F Compute scalability: GBC remains efficient despite Gaussian Feature Bridging, projection merging, and geometric target interpolation.These operations are described as lightweight and impose negligible computational burden in the evaluated setting.
  • F Compute scalability: GBC achieves substantial accuracy gains without sacrificing efficiency.The reported gains are referenced in Tables 1–2 alongside the runtime comparison.

G Limitations

GBC has limitations involving prototype-memory stability, the semantic validity of interpolated latent states, latency at larger scales, and scope beyond unimodal visual inputs.

  • G Limitations: Early Prototype Atlas updates may be sensitive to pseudo-label noise when labeled data is extremely scarce.The paper suggests more robust anchor initialization or noise-aware memory correction to stabilize Prototype Atlas dynamics.
  • G Limitations: Gaussian Feature Bridges assume interpolated latent states remain semantically meaningful.Highly entangled or rapidly evolving feature spaces may degrade intermediate representations, motivating adaptive or geometry-aware bridge shaping.
  • G Limitations: Bridging operations incur minor but non-zero latency, with larger-scale or multi-bridge variants potentially requiring more efficient implementations.The paper characterizes the overhead as negligible on CIFAR and ImageNet-127.
  • G Limitations: The current formulation focuses on unimodal visual inputs.Domain-shifted or multimodal extensions may require anchor-domain alignment or modality-specific bridge features.
Loading 2608.20710v1…