Research on Self-paced Learning based on Evolutionary Multi-objective Optimization

• 1、 Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University, Xi'an 710071

Abstract：Self-paced learning (SPL) is a recently proposed machine learning method. SPL imitates the learning process of humans. Humans learn the easy concepts at first and then take more complex knowledge into consideration. Therefore, SPL learns the easy samples at first and then involves more complex ones into training. SPL is beneficial in avoiding bad local minima and in achieving a better generalization results. In this paper, we introduce the background and related researches of self-paced learning. Then we use evolutionary multi-objective optimization to improve self-paced learning for addressing the issues in the current SPL research. Self-paced learning based on evolutionary multi-objective optimization optimizes the loss function and the self-paced regularizer simultaneously and aims to get a set of nondominated solutions. We can use some off-the-shelf tools in multi-objective optimization to find the best trade-off between the two objectives. Experiments on change detection in synthetic aperture radar images demonstrates the effectiveness of the proposed method.

Keywords： Pattern analysis and intelligence system machine learning self-paced learning multi-objective optimization evolutionary algorithm change detection