Combining Distance to Class Centroids and Outlier Discounting for Improved Learning with Noisy Labels

Farooq Ahmad WaniMaria Sofia BucarelliFabrizio Silvestri

Mar 2023

摘要

In this paper, we propose a new approach for addressing the challenge oftraining machine learning models in the presence of noisy labels. By combininga clever usage of distance to class centroids in the items' latent space with adiscounting strategy to reduce the importance of samples far away from all theclass centroids (i.e., outliers), our method effectively addresses the issue ofnoisy labels. Our approach is based on the idea that samples farther away fromtheir respective class centroid in the early stages of training are more likelyto be noisy. We demonstrate the effectiveness of our method through extensiveexperiments on several popular benchmark datasets. Our results show that ourapproach outperforms the state-of-the-art in this area, achieving significantimprovements in classification accuracy when the dataset contains noisy labels.

展开全部

图表提取

论文十问

参考文献

被引用

提问与回答

暂无人提供速读十问回答

被引用

笔记

问答

领域