Extreme learning machine with kernel model based on deep learning

Extreme learning machine (ELM) proposed by Huang et al. is a learning algorithm for single-hidden layer feedforward neural networks (SLFNs). ELM has the advantage of fast learning speed and high efficiency, so it brought into public focus. Later someone developed regularized extreme learning machine (RELM) and extreme learning machine with kernel (KELM). But they are the single-hidden layer network structure, so they have deficient in feature extraction. Deep learning (DL) is a multi-layer network structure, and it can extract the significant features by learning from a lower layer to a higher layer. As DL mostly uses the gradient descent method, it will spend too much time in the process of adjusting parameters. This paper proposed a novel model of convolutional extreme learning machine with kernel (CKELM) which was based on DL for solving problems—KELM is deficient in feature extraction, and DL spends too much time in the training process. In CKELM model, alternate convolutional layers and subsampling layers add to hidden layer of the original KELM so as to extract features and classify. The convolutional layer and subsampling layer do not use the gradient algorithm to adjust parameters because of some architectures yielded good performance with random weights. Finally, we took experiments on USPS and MNIST database. The accuracy of CKELM is higher than ELM, RELM and KELM, which proved the validity of the optimization model. To make the proposed approach more convincing, we compared with other ELM-based methods and other DL methods and also achieved satisfactory results.