Detecting problematic samples in deep learning

Abstract

This study investigates problematic samples in deep learning, that is samples that hinder model generalization. Although deep neural networks are typically trained on large datasets under the assumption that more data yields better performance, this work highlights that not all samples contribute positively and equally. To examine this issue, the study proposes a data pruning evaluation framework that systematically evaluates several methods (that we call problematic sample detectors) to reveal the presence, nature and impact of problematic samples. Specifically, we evaluate two learning-centric detectors (Sample forgetting and the Error-L2-norm) and two data-centric detectors (Confident Learning and k-nearest neighbours distance), from the fields of data valuation, label noise detection and out-of-distribution detection in deep learning. Using controlled evaluations on synthetic data and experiments involving explicitly added noise on MNIST and CIFAR- 10, trained on Multilayer Perceptrons and Convolutional Neural Networks, we analyse the ability of detectors to distinguish various forms of problematic samples. The results reveal that both input-space and label-space distributional shifts can induce problematic samples and that learning-centric and data-centric methods can be used as problematic sample detectors, although each have their limitations. This prompted us to develop a new problematic sample detector that we call the Variance-Weighted Fitness Gap detector to address the shortcomings of existing learning-centric methods. Our findings also show that problematic-sample detection can improve model generalization even in the absence of explicitly added noise on a real-world dataset. Furthermore, we highlight neglected nuances regarding the detection of problematic samples. These include the meaning of sample contribution in the presence or absence of label noise and the confounding effects of dependencies between samples.