EchoAlign: Bridging Generative and Discriminative Learning under Noisy Labels
Yuxiang Zheng , Zhongyi Han , Yilong Yin
Noisy labels severely hinder the accuracy and generalization of machine learning models, especially when caused by ambiguous instance features that complicate reliable annotation. Existing approaches, such as transition-matrix-based label correction, struggle to capture complex relationships between instances and noisy labels, limiting their effectiveness in such scenarios. We present EchoAlign, a framework that bridges generative and discriminative learning under noisy labels. Instead of correcting labels, EchoAlign treats noisy labels () as accurate and modifies corresponding instances (X) to align with them. The framework integrates two components: (1) EchoMod employs controllable generative models to adjust instance features while preserving key instance-level structural cues (e.g., shape and edges) and avoiding excessive distortion; and (2) EchoSelect addresses distribution shifts by retaining a reliable subset of original instances, guided by feature similarity between original and modified samples. This generative-discriminative interplay enables robust learning even in highly noisy settings. Experiments on three benchmark datasets show that EchoAlign outper-forms state-of-the-art methods in most evaluated settings. Under 30% instance-dependent noise, EchoSelect retains nearly twice as many correctly labeled samples as competing approaches while maintaining 99% selection accuracy, highlighting the robustness and effectiveness of EchoAlign.
Learning from Noisy Labels / Controllable Generative Models / Instance Modification / Feature Alignment / Sample Selection / Robust Machine Learning
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(47404KB)
