Deep and Wide Feature based Extreme Learning Machine for Image Classification

Abstract Extreme Learning Machine (ELM) is a powerful and favorable classifier used in various applications due to its fast speed and good generalization capability. However, when dealing with complex visual tasks, the shallow architecture of ELM makes it infeasible to have good performance when raw image data are directly fed in as input. Therefore, several works tried to make use of deep neural networks (DNNs) to extract features before ELM classification. On the other hand, when the depth of DNN is too deep, the ELM classifier may suffer from overfitting problem. To solve this issue, a novel deep and wide feature based Extreme Learning Machine (DW-ELM) has been proposed in this research work. We show that the overfitting problem can be largely remedied by employing a “widened” convolutional neural network (CNN) for feature extraction, in the sense that the number of feature maps for each convolutional layer is increased by factor of k compared to a reference model, i.e. deep residual networks (ResNets). While the wide design of residual networks has been shown to benefit image classification in terms of accuracy and efficiency, its application for feature extraction is not fully investigated. We provide an extensive experimental study in this work, showing that when combined with ELM that serves as a classifier, using wide ResNets (WRNs) for feature extraction can produce a performance leap on all benchmark datasets compared to a plain end-to-end trained network over a wide range of selections regardless of architecture choices and ELM designs, while normal ResNets as feature extractors do not provide a performance gain. The gap is even larger when fewer training iterations are employed. This indicates that a good feature extractor for ELM must be wide and deep. Experiments conducted on five benchmark datasets (CIFAR-100, CIFAR-10, STL-10, Flower-102 and Fashion-MNIST) have shown significant accuracy enhancement as well as training stability of the proposed DW-ELM. Ablation studies also demonstrate that the ELM classifier is an important component for DW-ELM which enables superior performance compared with SVM based image classification approaches.