TrainRef: Learning Distributional Labels with Minimal Reference for Noisy Data Curation
Date: 2025.05.05
Embedded Files
Contents
Contents
Summary of Notations
Proof of Convergence
More examples: [link] (mentioned in Sec. 4 Experiment in the paper)
Reliable soft labels: soft labels generated by TrainRef for ambiguous samples.
Reliable confidence: more reliable model prediction after fine-tune the model on the soft labels obtained from TrainRef.
Out-of-distribution(OOD) detection: Some examples detected as OOD samples by TrainRef.
Implementation details: [link] (mentioned in Sec. 4 Experiment)
Introduction
Introduction
We present TrainRef, a training-time data curation method that refines dataset quality by leveraging a reference set and an improved embedding space. In this way, we unify learning with noisy labels and confidence calibration to ensure more reliable model predictions.
Learning with noisy labels: TrainRef detects and corrects mislabeled samples, allowing the model to learn robustly despite imperfect supervision.
More reliable confidence: In high-stakes applications like autonomous driving, medical diagnosis, and finance, simply predicting a category is not enough—decisions must consider uncertainty. TrainRef improves confidence calibration, ensuring models provide meaningful probability estimates rather than just single predictions.
This approach enhances both model robustness and trustworthiness, making it suitable for real-world scenarios where understanding uncertainty is as important as making predictions.
Summary of Notations
Summary of Notations
Proof of Convergence
Proof of Convergence
As mentioned in the Sec. 2 Problem Statement in the paper, we give a formal proof of convergence of our TrainRef to ease the constraints of assumptions of perfection.
Page updated
Google Sites
Report abuse