This report aims to study how to train a deep learning based classifier when only large scale noisy dataset is available. In order to overcome dataset noise, a series of training as well as testing methods are proposed, including bootstrapping method …