Abstract

Deep learning has achieved classification performance matching or exceeding the human one, as long as plentiful labeled training samples are available. However, the performance on few-shot learning, where the classifier had seen only several or possibly only one sample of the class is still significantly below human performance. Recently, a type of algorithm called meta-learning achieved impressive performance for few-shot learning. However, meta-learning requires a large dataset of labeled tasks closely related to the test task. The work described in this dissertation outlines techniques that significantly reduce the need for expensive and scarce labeled data in the meta-learning phase. Our insight is that meta-training datasets require only in-class samples (samples belonging to the same class) and out-of-class samples. The actual labels associated with the classes are not relevant, as they are not retained in the meta-learning process. First, we propose an algorithm called UMTRA that generates out-of-class samples using random sampling from an unlabeled dataset, and generates in-class samples using augmentation. We show that UMTRA achieves a large fraction of the accuracy of supervised meta-learning, while using orders of magnitudes less labeled data. Second, we note that the augmentation step in UMTRA works best when an augmentation technology specific to the domain is used. In many practical cases it is easier to train a generative model for a domain than to find an augmentation algorithm. From this idea, we design a new unsupervised meta-learning algorithm called LASIUM, where the in- and out-of-class samples for the meta-learning step are generated by choosing appropriate points in the latent space of a generative model (such as a variational autoencoder or generative adversarial network). Finally, we describe work that makes progress towards a next step in meta-learning, the ability to draw the meta-training samples from a different domain from the target task's domain.

Notes

If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu.

Graduation Date

2021

Semester

Summer

Advisor

Boloni, Ladislau

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Degree Program

Computer Science

Format

application/pdf

Identifier

CFE0008680;DP0025411

URL

https://purls.library.ucf.edu/go/DP0025411

Language

English

Release Date

August 2021

Length of Campus-only Access

None

Access Status

Doctoral Dissertation (Open Access)

Share

COinS