Multilayer Fisher extreme learning machine for classification

Abstract As a special deep learning algorithm, the multilayer extreme machine (ML-ELM) has been extensively studied to solve practical problems in recent years. The ML-ELM is constructed from autoencoder (ELM-AE), and its generalization performance affected by representation of ELM-AE. However, given label information, unsupervised ELM-AE difficult build discriminative feature space for classification tasks. To address this problem, novel Fisher (FELM-AE) proposed used as component leaning (ML-FELM). FELM-AE introduces criterion into adding regularization term objective function, aiming maximize between-class distance minimize within-class abstract feature. Different ELM-AE, requires class labels calculate loss, so that learned contains sufficient category information complete ML-FELM stacks extract adopts (ELM) classify samples. Experiments on benchmark datasets show extracted more than results are competitive robust comparison with ELM, one-dimensional convolutional neural network (1D-CNN), ML-ELM, denoising (D-ML-ELM), generalized (ML-GELM), hierarchical L21‑norm loss (H-LR21-ELM).