Reproduce “MixStyle Neural Networks for Domain Generalization and Adaptation” on OpenResearch

About this paper

Kaiyang Zhou, Yongxin Yang, Yu Qiao, Tao Xiang

MixStyle neural networks introduce an efficient, parameter-free feature-level data augmentation method to improve out-of-distribution generalization in deep learning models. The approach implicitly synthesizes novel visual domains by mixing feature statistics, leading to superior performance across various tasks and reducing reliance on extensive data collection.

Key insights

• Feature statistics, specifically channel-wise means and standard deviations in shallow CNN layers, effectively capture image style information which defines visual domains.
• Mixing these instance-level feature statistics provides an efficient mechanism to implicitly synthesize novel domains in the feature space, serving as a powerful data augmentation technique.
• MixStyle diversifies training styles without increasing model capacity or requiring explicit domain alignment, leading to more robust and generalizable representations.

Summary by alphaXiv.