Multi-scale Deep Nearest Neighbors

In this thesis, we aim to learn a deep embedding space suitable for k-NN. Our approach is based on minimizing the leave-one-out 1-NN classification error in the embedding space. Directly optimizing for such a rule is not tractable due to its discontinuous nature. We propose Multi-scale Deep Nearest Neighbour (MsDNN) which is a differentiable loss function that aims to maximize the expected sample margin for every training sample. The output of MsDNN is an embedding space. We evaluate the resulting space from two angles. From the classification view, during testing, we run a k-NN classifier and report the classification accuracy. But classification accuracy does not tell us the entire story about the goodness of an embedding space. Therefore, we run k-means clustering in the embedding space. Analogous to the hierarchical clustering, subclasses might exist on different scales. Our method provides a mechanism to target subclasses in different scales.

Name thesis:degree_name

Master of Computer Science (M.C.S.)

Level thesis:degree_level

Master's

Discipline thesis:degree_discipline

Computer Science

Grantor dc:publisher

Carleton University

Year dc:date.issued

2020

Author dc:creator

• Chauhan, Abhijeet

Statement dc:rights

• Copyright © 2020 the author(s). Theses may be used for non-commercial research, educational, or related academic purposes only. Such uses include personal study, research, scholarship, and teaching. Theses may only be shared by linking to Carleton University Institutional Repository and no part may be used without proper attribution to the author. No part may be used for commercial purposes directly or indirectly via a for-profit platform; no adaptation or derivative works are permitted without consent from the copyright owner.

Language dc:language.iso

en

DOI

10.22215/etd/2020-14357

OAI identifier oai:identifier

oai:carleton.scholaris.ca:20.500.14718/41992