Mislabeled Samples Adjustment Based on Self-paced Learning Framework
Zhongtao Huang, Xiaojuan Li, Lingzhu Deng, Kaizhen Wei, Yunfeng Sui · 2021 7th International Conference on Computer and Communications (ICCC) · 2021
Human have ability to detect and correct the mistakes in life events. For instance, it could be easy for a child to find a cat image incorrectly labeled with a dog label. However, supervised learning network directly learns the mapping between features and labels even though some labels are obviously wrong. Label noise surely increases the complexity of model and undermine the model performance. Enlightened by self-paced learning (SPL) framework which can learn samples in the order of complexity, we propose one iteration-based framework called MSASL which can discriminate and correct the possibly mislabeled samples. The approach can not only achieve the data cleansing task, but also ensure the performance of the model.