2019/10/04 by Rakesh Katuwal, Katuwal, Rakesh, Ponnuthurai Nagaratnam Suganthan +1
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning and ELM #Neural and Evolutionary Computing (cs.NE) #Stochastic Gradient Optimization Techniques
paper · pdf · doi:10.48550/arxiv.1910.01858
openalex publication_date 2019/10/04 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
Extreme learning machine (ELM), which can be viewed as a variant of Random\nVector Functional Link (RVFL) network without the input-output direct\nconnections, has been extensively used to create multi-layer (deep) neural\nnetworks. Such networks employ randomization based autoencoders (AE) for\nunsupervised feature extraction followed by an ELM classifier for final\ndecision making. Each randomization based AE acts as an independent feature\nextractor and a deep network is obtained by stacking several such AEs. Inspired\nby the better performance of RVFL over ELM, in this paper, we propose several\ndeep RVFL variants by utilizing the framework of stacked autoencoders.\nSpecifically, we introduce direct connections (feature reuse) from preceding\nlayers to the fore layers of the network as in the original RVFL network. Such\nconnections help to regularize the randomization and also reduce the model\ncomplexity. Furthermore, we also introduce denoising criterion, recovering\nclean inputs from their corrupted versions, in the autoencoders to achieve\nbetter higher level representations than the ordinary autoencoders. Extensive\nexperiments on several classification datasets show that our proposed deep\nnetworks achieve overall better and faster generalization than the other\nrelevant state-of-the-art deep neural networks.\n