Employing a learning strategy analogous to human, from the easy to the difficult, better classifiers could be achieved in a complicated pattern classification from different domains, such as computer vision and natural language processing, etc. Curriculum Learning(CL) and Self-paced Learning (SPL) are typical exemplars of the learning strategy from the easy to the difficult. The applications of CL are limited due to the lack of expert or prior knowledge. SPL is easier to use, but it could develop an unreasonable curriculum in early training stages, which leads to an unsatisfactory classifier. Such a situation could be even worse in more complicated datasets (e.g. polluted datasets). To alleviate the above problem encountered by SPL, Weighted Self-Paced Learning with Belief Functions (WSPLBF) is proposed in this paper. In WSPLBF, the evidential uncertainty and the learning loss are jointly used to assist the model in selecting appropriate training samples and the probability based on belief functions is utilized to allocate the weight in SPL. Benefiting from the information calculated from belief functions, WSPLBF can achieve more accurate classifiers. Especially when the dataset is polluted, the proposed method has a 4% improvement in average accuracy compared with traditional SPL. Source codes of the proposed method are available on https://github.com/zhangshixing1994/weight-self-paced-learning-with-belief-functions.git.
