基于微分同胚优化极端学习机的人脸识别

摘要

极端学习机(ELM)以其快速高效和良好的泛化能力在模式识别领域得到了广泛应用.然而当前的ELM及其改进算法并没有充分考虑到隐层节点输出矩阵对极端学习机泛化能力的影响.通过实验发现激活函数选取不当及数据维数过高将导致隐层节点输出值趋于零,使得输出权值矩阵求解不准,降低ELM的分类性能.为此,提出一种微分同胚优化的极端学习机算法.该算法结合降维和微分同胚技术提高激活函数的鲁棒性,克服隐层节点输出值趋于零的问题.为验证所提算法的有效性使用人脸数据进行实验.实验结果表明所提算法具有良好的泛化性能.%Extreme learning machine (ELM) has been widely applied in the field of pattern recognition for its efficient and good generalization ability.However, the current ELM and its improved algorithm have not considered the effect of hidden layer nodes' output matrix on the generalization ability of extreme learning machine.Through experiments we find that when the activation function is improperly selected and the data sample dimension is too high, it will result in output value of hidden layer node tending to zero.It comes to make the solution of output weight matrix inaccurate and reduce the classification performance of ELM.In order to solve these problems, an optimized extreme learning machine algorithm based on diffeomorphism is proposed.The algorithm combines techniques of diffeomorphism and dimensionality reduction to improve the robustness of activation functions and overcome the problem that the output value of hidden layer nodes tends to zero.In order to evaluate the validity of the proposed algorithm, face data is used to implement experiments.Experimental results show that the proposed algorithm has a good generalization performance.