Deep Extreme Learning Machines for Classification

Abstract

We present a method for synthesising deep neural networks using Extreme Learning Machines (ELMs) as a stack of supervised autoencoders. We show that the network achieves comparable performance to an ELM with a single hidden layer with a size equal to the total number of hidden-layer neurons in the deep network. The main advantage of our method is in its significantly reduced network training time and memory usage. These favourable properties suggest that our method can be applied to a resource-constrained hardware implementation to increase the network performance.