Reduce Consumption in Learning: from Data, Model, and Task

Abstract

While deep learning approaches have demonstrated impressive results in a wide variety of visual recognition tasks, there are several key factors still stop it from general usages: 1) the needs of having large amount of training data cost expensively ; 2) the models with great performance are usually too heavy to be fitted into mobile or edge devices; 3) models or data have low generalizability across different tasks. In this talk I will introduce several tools aiming for reducing the consumption in deep learning from aforementioned perspectives, from my own tunnel view. If time permits, I will close by listing several exciting research topics that my research group is working on.