An Adaptive Data cleaning Framework for Noisy Label Detection

Researchers have proposed a new framework for detecting noisy labels in image data and a large-scale benchmark for evaluating data cleaning methods.

A new data-cleaning framework integrates local, global, and learning dynamics cues to identify noisy labels in image datasets, according to a paper on arXiv^[1]. The framework uses multi-metric clustering on a feature space to partition samples into clean and noisy components without requiring manual thresholds or prior knowledge of the noise ratio. Meanwhile, a new benchmark called CleanPatrick has been introduced, utilizing the Fitzpatrick17k dermatology dataset with 496,377 binary annotations collected from 933 medical crowd workers^[2]. The annotations revealed that 4% of the dataset was off-topic, 21% were near-duplicates, and 32% contained label errors. The CleanPatrick benchmark is the first large-scale benchmark for data cleaning in the image domain^[2]. Experiments with the new framework on various datasets, including CIFAR-10 and ImageNet-100, showed high recall across different noise levels, with near-perfect recall (>=98%) on ImageNet-100 at 40% noise^[1].