Date of Award

12-2017

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical and Computer Engineering (Holcomb Dept. of)

Committee Member

Dr. Melissa C. Smith, Committee Chair

Committee Member

Dr. Walter B. Ligon III

Committee Member

Dr. Adam W. Hoover

Abstract

Regardless of the Deep Learning community's continuous advancements, the challenging domain of one-shot learning still persists. While the human brain is capable of learning a new visual concept with ease, sometimes even at a glance, Deep Learning-based techniques show serious drawbacks in handling problems with small datasets. Much of the existing work on one-shot learning employs a variety of sophisticated network algorithms, prior domain knowledge, and data manipulation to address the generalization challenges presented in such problems. In this work, we demonstrate a one-shot learning method that contains three learning networks — a deep sequential generative model, a candidate network, and a Matching Network — thus offering an alternative approach to solving the one-shot classification problem. The proposed framework does not require domain knowledge, making it potentially portable to other domains. We show that our algorithm improves accuracy from 95.5% to 96.1% on the Omniglot dataset 20-way one-shot learning compared to current state-of-the-art.

Share

COinS