Non-iterative and fast deep learning: multilayer extreme learning machines (Q2005317)

This paper focuses primarily on studying non-iterative and fast deep learning: multilayer extreme learning machines (ELM). The deep residual network (ResNet) was proposed for easing the training of deep neural networks, in which stacked layers were fitted with residual mappings instead of desired underlying mappings. In the proposed framework, ML-ELM-1 is implemented for extracting features of all the modalities separately, and the extracted discriminative features run through a joint dimension reduction by weakly paired maximum covariance analysis. Compared with linear weakly paired approaches, the proposed framework can achieve better performance with better robustness. ELM theory is first introduced in Section 2. Stacked ELM autoencoder (ELM-AE), residual ELM, and local receptive field-based ELM (ELMLRF) are then reviewed in Sections 3, 4, and 5, respectively. Selected applications are highlighted in Section 6. Finally, discussions are given in Section 7, followed by conclusions in Section 8.