Data Augmentation by Feature Space Profiling and Mining for Building Powerful Models with Very Little Data

Abstract

Data are most crucial and essential building component for any data mining and AI applications exist. More significantly, deep learning approaches require massive datasets. We know that the theory and algorithms have been around for quite a while however the ability to process the right amounts of data brought us to the recent breakthroughs in the field. A challenge comes up in a case of a small dataset, comparing to the required training data required. However, mostly, getting this data are neither an easy nor a cheap task, many annotating services take advantage of the problem and charge for tagging data-sets campaigns, those could cost hundreds of dollars easily and yet with an uncertain quality. As the task of generalization at hand, we wondered how to exploit the minimal data we have and still have an AI system to learn well. In this paper, we overview methods for solving the problem and suggest solutions in order to overcome the challenge.