Filtering Clean Sample from Noisy Datasets by Creating and Analyzing Artifical Class

Abstract

A new method for extracting clean samples from noisy labelled classification dataset without using any clean dataset and making any assumption related to noise rate is proposed in this work. The proposed method suggests creating artificial samples, which are mimicking noisy samples and absolutely noisy, to understand behavior of noisy samples during training of a classifier neural network. The proposed method investigates behavior of artificial samples during training to classify other samples as clean or noisy. Performance of clean sample extraction and classifier neural network trained with the extracted clean samples are observed with using proposed method. When presented results are observed, it is proved that the proposed algorithm is sucessful in terms of extracting clean dataset and provides better or similar results with compared algorithms.