Semi-supervised convolutional extreme learning machine

Abstract

We propose a scheme for training a neural network as an image classifier. The approach includes a very rapid unsupervised feature learning algorithm and a supervised technique. We show that convolving and downsampling clustered descriptors of image patches with each input image can provide more discriminative features compared to both pre-trained descriptors and randomly generated convolutional filters. The implemented algorithm to discover clusters centroids (i.e. k-means clustering) for color images is not restricted to only RGB and we show that the algorithm is appropriate for Lab color representations. We use the centroids for obtaining convolutional features. We also present a high performance extreme learning machine (ELM), which is a method characterized by low implementation complexity, and run-time, to classify the learned features. We show that the combination of the unsupervised feature learning with the ELM outperforms previous related models that use different feature representations fed into an ELM, on the CIFAR-10 and Google Street View House Number (SVHN) datasets.