Self-Paced Learning with Statistics Uncertainty Prior

Lihua GUO

IEICE TRANSACTIONS on Information and Systems   Vol.E101-D    No.3    pp.812-816
Publication Date: 2018/03/01
Publicized: 2017/12/13
Online ISSN: 1745-1361
DOI: 10.1587/transinf.2017EDL8169
Type of Manuscript: LETTER
Category: Artificial Intelligence, Data Mining
self-paced learning,  curriculum learning,  uncertainty prior,  simulated annealing,  

Full Text: PDF>>
Buy this Article

Self-paced learning (SPL) gradually trains the data from easy to hard, and includes more data into the training process in a self-paced manner. The advantage of SPL is that it has an ability to avoid bad local minima, and the system can improve the generalization performance. However, SPL's system needs an expert to judge the complexity of data at the beginning of training. Generally, this expert does not exist in the beginning, and is learned by gradually training the samples. Based on this consideration, we add an uncertainty of complexity judgment into SPL's system, and propose a self-paced learning with uncertainty prior (SPUP). For efficiently solving our system optimization function, an iterative optimization and statistical simulated annealing method are introduced. The final experimental results indicate that our SPUP has more robustness to the outlier and achieves higher accuracy and less error than SPL.