Algorithms and feature preprocessing for transductive few-shot image classification

Overview & Synopsis

The purpose of this thesis is to investigate one of the most important challenges related to the development of machine and deep learning methods. Namely, our research is conducted in the setting where models make predictions based on a few labeled examples. Particularly in the context of image classification, the goal of this study is to learn a model that can correctly predict class labels based on limited data samples. We firstly discuss the improved performance with the evolution of deep learning methods, and present the problematic of data thriftiness. Secondly, we introduce the standard settings of this problematic and present the related classification methods. We summarize a general pipeline to tackle it. Then we highlight our contributions that address each step in the pipeline, by proposing adaptive methods on the targeted image data whose number is limited by the cost of annotation. Finally, we draw conclusions of our work, along with discussions about the novel challenges as well as potential solutions related to the field.