Variation-Bounded Loss for Noise-Tolerant Learning

This repository is the official code of the Variation-Bounded Loss for Noise-Tolerant Learning (VBL).

How to use

Benchmark Datasets: The running file is main.py

• dataset: cifar10 | cifar100.
• loss: NCEandVCE, NCEandVEL, NCEandVSL, CE, GCE, etc.
• noise_type: symmetric | asymmetric | dependent (IDN) | human (cifar-n).

Real-World Datasets: The running file is main_real_world.py

• dataset: webvision | clothing1m.
• loss: NCEandVCE, NCEandVEL, NCEandVSL, CE, GCE, etc.

Examples

NCEandVCE for cifar10 with 0.8 symmetric noise:

python main.py --dataset cifar10 --noise_type symmetric --noise_rate 0.8 --loss NCEandVCE    

NCEandVCE for webvision:

python main_real_world.py --dataset webvision --loss NCEandVCE

Reference

For details, please check the paper. If you have used our method or code in your own, please consider citing:

@inproceedings{
wang2026variationbounded,
title={Variation-Bounded Loss for Noise-Tolerant Learning},
author={Jialiang, Wang and Xiong, Zhou and Xianming, Liu and Gangfeng, Hu and Deming, Zhai and Junjun, Jiang and Hangliang, Li},
booktitle={The Fortieth AAAI Conference on Artificial Intelligence},
year={2026},
}

If you have any question, you can contact cswjl@stu.hit.edu.cn