Pre-trained convolutional networks and generative statistical models: A comparative study in large datasets

Abstract

This study explored the viability of out-the-box, pre-trained ConvNet models as a tool to generate features for large-scale classification tasks. A juxtaposition with generative methods for vocabulary generation was drawn. Both methods were chosen in an attempt to integrate other datasets (transfer learning) and unlabelled data, respectively. Both methods were used together, studying the viability of a ConvNet model to estimate category labels of unlabelled images. All experiments pertaining to this study were carried out over a two-class set, later expanded into a 5-category dataset. The pre-trained models used were obtained from the Caffe Model Zoo. The study showed that the pre-trained model achieved best results for the binary dataset, with an accuracy of 0.945. However, for the 5-class dataset, generative vocabularies outperformed the ConvNet (0.91 vs. 0.861). Furthermore, when replacing labelled images with unlabelled ones during training, acceptable accuracy scores were obtained (as high as 0.903). Additionally, it was observed that linear kernels perform particularly well when utilized with generative models. This was especially relevant when compared to ConvNets, which require days of training even when utilizing multiple GPUs for computations. © Springer International Publishing AG 2017.