Recognizing Foods using Deep Neural Networks under Domain Shift

Abstract

Automatic tracking of daily food intake is an efficient method to tackle obesity and micronutrient deficiency. This could be done by developing a method that is able to classify foods using their images. Foods are highly deformable objects with great inter-class similarity and intra-class variation. For these reasons, we need a feature transformation function with ability to learn complex mappings. Deep neural networks possess this property, and they are able to generalize well if they are trained on big and diverse datasets. In absence of a large target dataset, we can train the network on a big and related dataset and adapt the knowledge acquired from this dataset to the target dataset. In this thesis, we formulate the problem of transfer learning and break it down to knowledge adaptation and domain adaptation. Alongside, we explain how to compute uncertainty of prediction in neural network. After studying these two problems, we explain how active learning could help us to improve neural network models with minimum amount of annotation. In the last part of the thesis, we designed a new network and show how to distill knowledge of a bigger network to a smaller network.