CAJ | 학술논문

单隐藏层前馈神经网络(Single-hidden Layer Feedforward Neural Network,SLFN)已经在模式识别、自动控制及数据挖掘等领域取得了广泛的应用,但传统学习方法的速度远远不能满足实际的需要,成为制约其发展的主要瓶颈.产生这种情况的两个主要原因是:(1)传统的误差反向传播方法(Back Propagation,BP)主要基于梯度下降的思想,需要多次迭代;(2)网络的所有参数都需要在训练过程中迭代确定.因此算法的计算量和搜索空间很大.针对以上问题,借鉴ELM的一次学习思想并基于结构风险最小化理论提出一种快速学习方法(RELM),避免了多次迭代和局部最小值,具有良好的泛化性、鲁棒性与可控性.实验表明RELM综合性能优于ELM、BP和SVM.
단은장층전궤신경망락(Single-hidden Layer Feedforward Neural Network,SLFN)이경재모식식별、자동공제급수거알굴등영역취득료엄범적응용,단전통학습방법적속도원원불능만족실제적수요,성위제약기발전적주요병경.산생저충정황적량개주요원인시:(1)전통적오차반향전파방법(Back Propagation,BP)주요기우제도하강적사상,수요다차질대;(2)망락적소유삼수도수요재훈련과정중질대학정.인차산법적계산량화수색공간흔대.침대이상문제,차감ELM적일차학습사상병기우결구풍험최소화이론제출일충쾌속학습방법(RELM),피면료다차질대화국부최소치,구유량호적범화성、로봉성여가공성.실험표명RELM종합성능우우ELM、BP화SVM.
SLFNs(Single-hidden Layer Feed forward Neural networks) have been widely applied in many fields including pattern recognition,automatic control,data mining etc.However,the traditional learning methods can not meet the actual needs due to two main reasons.Firstly,the traditional method is mainly based on gradient descent and it needs multiple iterations.Secondly,all of the network parameters need to be determined by iteration.Therefore,the computational complexity and searching space will increase dramatically.To solve the above problem,motivated by ELM's one-time learning idea,a novel algorithm called Regularized Extreme Learning Ma-chine (RELM) based on structural risk minimization and weighted least square is proposed in this paper.The algorithm not only avoids a number of iterations and the local minimum,but also has better generalization,robustness and controllability than the original ELM.Additionally,experi-mental results have shown that RELM' overall performance is also better than BP and SVM.