Training a Mentee Network by Transferring Knowledge from a Mentor Network

Abstract

Automatic classification of foods is a challenging problem. Results on ImageNet dataset shows that ConvNets are very powerful in modeling natural objects. Nonetheless, it is not trivial to train a ConvNet from scratch for classification of foods. This is due to the fact that ConvNets require large datasets and to our knowledge there is not a large public dataset of foods for this purpose. An alternative solution is to transfer knowledge from already trained ConvNets. In this work, we study how transferable are state-of-art ConvNets to classification of foods. We also propose a method for transferring knowledge from a bigger ConvNet to a smaller ConvNet without decreasing the accuracy. Our experiments on UECFood256 dataset show that state-of-art networks produce comparable results if we start transferring knowledge from an appropriate layer. In addition, we show that our method is able to effectively transfer knowledge to a smaller ConvNet using unlabeled samples.