Dynamic Label Smoothing for Fine-Grained Dataset

Abstract

This paper addresses the classification task of deep learning model. We provide a proposed method to improve Label Smoothing based regularization technique between loss and data. We mathematically show that label smoothing performs poorly on fine-grained data. Based on mathematical evidence, we propose Dynamic Label Smoothing, a new Label Smoothing based on regularization between loss and data. Experimental results show that Dynamic Label Smoothing provides more accurate visualization than the previous research.