【摘要】 基于极限学习机自编码器(extreme learning machine based autoencoder, ELM-AE)和误差反向传播(back propagation, BP)算法，针对ELM提出了一种改进的特征表示方法。首先，使用ELM-AE以无监督的方式学习紧凑的特征表示，即ELM-AE输出权重；其次，利用ELM-AE输出权重来初始化BP神经网络的输入权重，然后对BP网络进行监督训练；最后，用微调的BP网络输入权重初始化ELM的输入权重参数。在MNIST数据集上的实验结果表明，采用BP算法对ELM-AE学习的参数进行约束，可以得到更紧凑且具有判别性的特征表示，有助于提高ELM的性能。更多还原

【Abstract】 An improved feature representation method for extreme learning machine(ELM) was proposed based on extreme learning machine based autoencoder(ELM-AE) and error back propagation(BP) algorithms.Firstly, ELM-AE was used to learn compact feature representation i.e.ELM-AE output weights, in an unsupervised way.Secondly, ELM-AE output weights were used to initialize the input weights of the BP neural network, which was then trained by BP network in a supervised way.Finally, the input weight parameters of ELM were initialized by the input weight of the fine-tuned BP network.The experimental results on MNIST dataset show that using the BP algorithm to constrain the parameters of ELM-AE learning can result in a more compact and discriminative feature representation, which helps to improve the performance of ELM.更多还原