Overview

"The focus of work in this thesis is on the problem of Distance Metric Learning (DML) with limited supervision, i.e., we seek to learn distance metrics with very few labeled examples (semi-supervised DML) or no labeled examples at all (unsupervised DML). The DML is fundamentally related to the problem of embedding learning, and hence both the problems are studied interchangeably. Though the recent, state-of-the-art DML approaches have shown promise of success on a variety of machine learning tasks, they all make use of class labels or manual annotations, to obtain constraints for DML. However, obtaining class labels for large datasets may not be feasible in many tasks, thus limiting the practical applications of these approaches. To further emphasize the importance of looking beyond labeled data, we begin the thesis by borrowing the analogy recently stated by Yann LeCun 1: ""Assuming intelligence to be a cake, supervised learning can be seen as the icing and Reinforcement Learning (RL) can be seen as the cherry. It is unsupervised learning that forms the bulk of the cake!"""

Full Product Details

9798224740444

Table of Contents

Reviews

Author Information

Tab Content 6

Customer Reviews

Recent Reviews

Countries Available