Geometric Regularization for Long-Tailed Semi-Supervised Learning via Gaussian Feature Bridges
GBC introduces class-conditional Gaussian feature bridges, improving long-tailed SSL by enforcing geometric consistency and feature interpolation, boosting tail-class accuracy.
Key Findings
Methodology
GBC employs a dynamic Prototype Atlas to store diverse class exemplars, constructing Gaussian feature bridges in latent space between unlabeled samples and class anchors. Along these paths, a bridge consistency loss enforces prediction alignment with geometrically interpolated targets. The method integrates BridgeMix, a confidence-guided feature interpolation strategy, to enhance sample diversity and robustness. Experiments on CIFAR10-LT and ImageNet-LT datasets demonstrate significant improvements in tail-class performance, with theoretical analysis confirming the regularization effect on decision boundary smoothness and local Lipschitz continuity.
Key Results
- On CIFAR10-LT, GBC improves tail-class accuracy by 3.5%, and on ImageNet-LT, achieves a 2.8% increase in Top-1 accuracy over baselines like SimPro. Incorporating BridgeMix further stabilizes performance under noisy pseudo-labels and class imbalance, validating the geometric regularization approach.
- Across different architectures (ResNet-50, ViT) and feature layers, GBC consistently outperforms existing methods, confirming its generalizability. Theoretical proofs show that bridge consistency enforces smooth prediction evolution and decision boundary stability, enhancing robustness.
- Results indicate that the class-conditional Gaussian feature bridges effectively mitigate confirmation bias and semantic drift, especially in extreme long-tail scenarios, leading to more reliable and balanced representations.
Significance
This work addresses fundamental challenges in long-tailed SSL, providing a theoretically grounded and practically effective regularization mechanism. It advances the understanding of feature-level geometric regularization, offering a scalable solution for imbalanced data in real-world applications like medical imaging and financial analysis. By stabilizing feature alignment and decision boundaries, GBC paves the way for more robust, fair, and accurate AI systems in complex environments.
Technical Contribution
The paper introduces a Schrödinger bridge-inspired Gaussian feature bridge mechanism, enabling smooth semantic paths in latent space. It combines this with a dynamic Prototype Atlas for class-aware anchor storage and a confidence-guided BridgeMix strategy for feature interpolation. Theoretical analysis proves that bridge consistency enforces local Lipschitz continuity and decision boundary stability, providing rigorous guarantees for generalization and robustness. This integration of geometric regularization with pseudo-label refinement marks a significant departure from existing methods that focus solely on output-level calibration.
Novelty
This is the first work to incorporate Schrödinger bridge concepts into semi-supervised long-tailed learning, constructing class-conditional Gaussian feature paths for regularization. Unlike prior methods that rely on label correction or class reweighting, GBC emphasizes feature-level geometric smoothness, enabling more stable and scalable learning in highly imbalanced scenarios. The combination of dynamic prototype storage and confidence-guided feature interpolation offers a novel framework that bridges probabilistic theory and practical SSL.
Limitations
- The effectiveness depends on the quality and update frequency of the Prototype Atlas; poor or stale prototypes may impair performance.
- In extremely noisy or highly imbalanced datasets, the stability of Gaussian bridges may degrade, requiring further robustness enhancements.
- Computational overhead increases with the size of the prototype set and the complexity of feature interpolation, posing challenges for real-time deployment.
Future Work
Future research could explore adaptive and non-linear bridge functions, multi-modal extensions, and unsupervised prototype refinement. Integrating self-supervised signals and reducing computational costs will further enhance scalability. Extending the framework to other tasks like detection and segmentation, and testing in real-world noisy environments, are promising directions.
AI Executive Summary
In many real-world scenarios, semi-supervised learning struggles with long-tailed class distributions and noisy pseudo-labels, which often lead to biased models and poor tail-class performance. Traditional approaches, such as pseudo-label refinement or class reweighting, have limited effectiveness when the data imbalance is severe or the pseudo-label noise is high. To address these challenges, this paper introduces the Gaussian Bridge Consistency (GBC) framework, inspired by Schrödinger bridge theory, which constructs class-conditional Gaussian feature paths in the latent space. These paths connect uncertain unlabeled features with reliable class anchors stored in a dynamic Prototype Atlas, enforcing geometric regularization through a bridge consistency loss. This mechanism encourages the model to traverse smoothly from uncertain to confident representations, reducing semantic drift and confirmation bias. Additionally, the authors propose BridgeMix, a confidence-guided feature interpolation strategy that amplifies sample diversity and robustness by interpolating features and anchors based on pseudo-label confidence scores. Extensive experiments on CIFAR10-LT and ImageNet-LT datasets demonstrate that GBC consistently outperforms state-of-the-art methods, especially in tail-class accuracy, with improvements of up to 3.5% in accuracy metrics. Theoretical analysis confirms that the geometric regularization enforces local Lipschitz continuity and decision boundary stability, which are crucial for robustness under long-tailed and noisy conditions. This work significantly advances the understanding of feature-level geometric regularization in semi-supervised learning, providing a scalable and theoretically grounded solution for imbalanced data environments. Its implications extend to real-world applications such as medical diagnosis, financial risk assessment, and autonomous systems, where data imbalance and noise are prevalent. Future directions include exploring non-linear bridges, multi-modal extensions, and more efficient prototype maintenance, aiming to further improve robustness and scalability in complex scenarios.
Deep Dive
Plain Language Accessible to non-experts
想象你在一家工厂里工作，工厂里生产各种不同的产品。有些产品很常见（比如普通的玩具），而有些则很少见（比如特殊的机械零件）。工厂的目标是让每个产品都能被正确识别和生产，但因为少见的产品样本少，工人（模型）很难学会它们的特征。现在，工厂引入了一种新方法，就像在不同产品之间架起一座桥。这个桥用一种特殊的“平滑线”连接你已经知道的常见产品和少见产品的代表样本（锚点）。工人可以沿着这座桥，从模糊的认识逐步变得清晰，最终能准确识别所有产品。这个桥是用一种叫高斯（正态）分布的数学方法模拟的，确保路径平滑连续。这样，工人就能逐步理解少见产品的特征，不会一下子迷糊了。通过不断调整和优化这座桥，工厂的生产效率和产品识别都大大提高，就像用一条平滑的路径，让工人从模糊到清楚，逐步掌握每个产品的特征。
ELI14 Explained like you're 14
想象你在学校里学习不同的动物，有些你很熟悉（比如猫和狗），但有些很少见（比如稀有的鸟）。老师想帮你更好地认识这些少见动物，于是设计了一条“想象之路”。这条路像一座桥，把你已经知道的动物和少见的动物连接起来。你可以沿着这座桥，从你熟悉的动物慢慢变得像少见的动物一样。老师还用一种特别的魔法（叫高斯分布）让你在脑海中看到这条路变得平滑、连续。这样，你就能逐步理解少见动物的特征，不会一下子就迷糊了。这个方法帮助你变得更聪明，也让老师更容易教会你所有动物的样子。就像在学习中画一条平滑的路径，让你从模糊到清楚，逐步掌握所有知识。
Abstract
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.