Graph-based Methods For Zero-shot Learning In Image Classifications
Open Access
- Author:
- Chai, Jiaming
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 21, 2022
- Committee Members:
- Wang-Chien Lee, Thesis Advisor/Co-Advisor
George Kesidis, Committee Member
Chitaranjan Das, Program Head/Chair - Keywords:
- Machine Learning
Zero-shot Learning
Image Classification - Abstract:
- Image classification has been extensively researched recently on large-scale datasets using complex neural network models. Those models, requiring large amount of labelled data in the training sets for effective training, are usually designed for classifying images into known classes (i.e., those used to label images in the training sets). In this thesis, we investigate the problem of image classification for classes with no labels provided in the training sets, known as zero-shot learning problem. Zero-shot learning tends to explore side information so that models can learn without relying only on the actual image data. The side information contains useful features that may be encoded into latent representations (also called embeddings) which are then decoded for classification. In recent years, the class labels (words) and their relationships in WordNet have been exploited as the side information to train classification models for the zero-shot problem. Inspired by the recent research advances, I develop a number of zero-shot image classification models that capture multi-type relationships and relationships typed by meta-paths between words (text labels) in WordNet. We evaluate the proposed models in comparison with the state-of-the-art models using large-scale datasets. Experimental results show that our methods outperform the state-of-the-art models in terms of zero-shot accuracy and top-k accuracy.