라벨 노이즈 데이터 제거

Abstract

When labeling datasets, inaccurate labels may be given due to causes such as annotation error or ambiguity in data. Incorrect labeling can confuse the learning model in classifying data and degrade performance, so even after labeling is finished, it increases the reliability of labeling through several inspections. Dataset labeling and inspection of assigned labels are time-consuming and costly. If the model can judge the noisy labeled data itself, it will overcome the previously mentioned costs. In this paper, we present a method to filter out noise label data that degrades the classification performance of the model through a filter model.