Access Type

Open Access Dissertation

Date of Award

January 2018

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Ming Dong

Abstract

Deep learning methods have achieved great success in the areas of Computer Vision and Natural Language Processing. Recently, the rapidly developing field of deep learning is concerned with questions surrounding how we can learn meaningful and effective representations of data. This is because the performance of machine learning approaches is heavily dependent on the choice and quality of data representation, and different kinds of representation entangle and hide the different explanatory factors of variation behind the data.

In this dissertation, we focus on representation learning with deep neural networks for different data formats including text, 3D polygon shapes, and brain fiber tracts.

First, we propose a topic-based word representation learning approach for text classification. The proposed approach takes global semantic relationship between words over the whole corpus into consideration and encodes the relationships into distributed vector representations with continuous Skip-gram model. The learned representations which capture a large number of precise syntactic and semantic word relationships are taken as input of Convolution Neural Networks for classification. Our experimental results show the effectiveness of the proposed method on indexing of biomedical articles, behavior code annotation of clinical text fragments, and classification of news groups.

Second, we present a 3D polygon shape representation learning framework for shape segmentation. We propose Directionally Convolutional Network (DCN) that extends convolution operations from images to the polygon mesh surface with rotation-invariant property. Based on the proposed DCN, we learn effective shape representations from raw geometric features and then classify each face of a given polygon into predefined semantic parts. Through extensive experiments, we demonstrate that our framework outperforms the current state-of-the-arts.

Third, we propose to learn effective and meaningful representations for brain fiber tracts using deep learning frameworks. We handle the highly unbalanced dataset by introducing asymmetrical loss function for easily classified samples and hard classified ones. The training loss avoids to be dominated by the easy samples and the training step is more efficient. In addition, we learn more effective and meaningful representations by introducing deeper network and metric learning approaches. Furthermore, we propose to improve the interpretability of our framework by inducing attention mechanism. Our experimental results show that our proposed framework outperforms current golden standard significantly on the real-world dataset.

Download

Share

COinS