Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks

[jchanxtarov]

Basic Information

• Authors: Curtis G. Northcutt, Anish Athalye, Jonas Mueller
• Date: 26 Mar 2021
• Published By: ICLR 2021 RobustML and Weakly Supervised Learning Workshops; NeurIPS 2020 Workshop on Dataset Curation and Security

Link

https://arxiv.org/abs/2103.14749

Overview

• Even if in the case of the well-known open datasets such as ImageNet have an average of 3.4% label errors.
• Recent state-of-the-art models are overfitting for these miss-labels.
• For these noisy real-world data, simpler models such as the ResNet-18 perform better.

Others

Reference (for understanding)