SVMnet: Non-parametric image classification based on convolutional SVM ensembles for small training sets

Hunter Goddard, Lior Shamir

Keywords: machine learning, support vector machine, convolutional neural network

Abstract: Deep convolutional neural networks (DCNNs) have demonstrated superior power in their ability to classify image data. However, one of the downsides of DCNNs for supervised learning of image data is that their training normally requires large sets of labeled "ground truth" images. Since in many real-world problems large sets of pre-labeled images are not always available, DCNNs might not perform in an optimal manner in all real-world cases. Here we propose SVMnet -- a method based on a layered structure of Support Vector Machine (SVM) ensembles for non-parametric image classification. By utilizing the quick learning of SVMs compared to neural networks, the proposed method can reach higher accuracy than DCNNs when the training set is small. Experimental results show that while "conventional" DCNN architectures such as ResNet-50 outperform SVMnet when the size of the training set is large, SVMnet provides a much higher accuracy when the number of "ground truth" training samples is small.

One-sentence Summary: SVMnet is a new machine learning model for image classification inspired by convolutional neural networks and designed for classification tasks where few labeled training samples are available.