Self-Correction Method for Automatic Data Annotation

Supervised learning from data, especially using deep neural networks, usually needs tons of labeled data. Considering there are plenty of handwritten samples in the Internet, how to make use of the large amount of dirty data or make them cleaner? This paper presents self-correction learning method to automatically label dirty data. The methods lend inspiration from self-training framework. It trains the classifier iteratively on the dirty data, and shifts the cluster centers, corrects or deletes samples according to the estimated confidence during the iterations. Empirical results demonstrate that the proposed method can effectively improve the quality of the data and reduce great human annotation efforts.