基于栈式结构的深度极限学习机

Deep Extreme Learning Machine Based on Stacking Structure

针对极限学习机无法学习到原始数据中蕴含的知识,以及其隐节点设置不当引起的过拟合等问题,提出一种快速训练的深度栈式极限学习机(FT DSELM).栈式结构以极限学习机为基础单元,通过级联多个基础单元,将初始样本与当前基础单元的决策信息逐层融合,充分挖掘子分类器中的分类知识,并引入dropout思想,以增加分类器集成的多样性.基于基准数据集的实验表明,FT DSELM不仅能深化对原始信息的理解,还具有较好的识别性能和极快的学习速度.

In order to alleviate the problem that ELM cannot learn the knowledge implied in the original data and overfitting caused by improper assignment of hidden nodes,a fast training deep stacking extreme learning machine(FT DSELM)is proposed.The stacking structure regards ELM as the basic building module and cascades multiple modules.The original instances and the decision of the current module are fused in a layerwise manner so that the classification knowledge in the sub classifier is fully extracted.Additionally,the theory of dropout is also introduced to improve the diversity of ensemble structure.According to the experimental result,FT DSELM not only develops a deeper insight into the original information,but also achieves satisfactory recognition performance and superb learning speed.