Machine learning algorithms for sparse supervision

Abstract

The recent advances in artificial intelligence technology have been attributed to the exponential increase in the amount of data. On the other hand, however, it is becoming more difficult to categorize or label the data for training of artificial intelligence model. In this sense, machine learning algorithms in low-supervised situation with little or no labeled data are becoming increasingly important. This thesis proposes sparse supervision machine learning algorithms using only a few labeled data or utilizing unlabeled data. In the first part of thesis, we focus on the few-shot learning method that trains the model to classify new classes with only a few labeled data. The proposed few-shot learning model has the ability to quickly adapt to new tasks through a linear projection of the feature space. In the second part, we focus on the few-shot segmentation, which aims to conduct semantic segmentation for a new class with only a small amount of data. We propose a few-shot segmentation method utilizing existing semantic segmentation model by transforming the unfamiliar novel feature into more comprehensible form based on the few labeled data. Finally in the third part, we focus on the self-supervised learning method that trains the useful representation using the unlabeled data. We propose an contrastive self-supervised learning method utilizing the nonlinear feature transformation via the self-attention technique to overcome the limitation of the existing contrastive self-supervised learning methods.