Simplifying Neural Network Training Under Class Imbalance

Ravid Shwartz-Ziv, Micah Goldblum, Yucen Lily Li, C. Bayan Bruss, Andrew Gordon Wilson

Keywords: Class Imbalance, Hyperparameters, Long-Tailed Distributions

TL;DR: We demonstrate that by optimizing the standard components of deep learning pipelines, we can achieve state-of-the-art performance on imbalanced datasets, eliminating the need for specialized loss functions or samplers.

Abstract: Real-world datasets are often highly class-imbalanced, which can adversely impact the performance of deep learning models. The majority of research on training neural networks under class imbalance has focused on specialized loss functions and sampling techniques. Notably, we demonstrate that simply tuning existing components of standard deep learning pipelines, such as the batch size, data augmentation, architecture size, pre-training, optimizer, and label smoothing, can achieve state-of-the-art performance without any specialized loss functions or samplers. We also provide key prescriptions and considerations for training under class imbalance, and an understanding of why imbalance methods succeed or fail.

Submission Number: 13305