Date of Award
Electrical and Computer Engineering
Image classification, Moore-penrose inverse, Representation learning, Subnet neural network
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
With the fast technological advancement in data storage and machine learning, big data analytics has become a core component of various practical applications ranging from industrial automation to medical diagnosis and from cyber-security to space exploration. Recent studies show that every day, more than 1.8 billion photos/images are posted on social media, and 720 thousand hours of videos are uploaded to YouTube. Thus, to handle this large amount of visual data efficiently, image/video classification, object detection/recognition, and segmentation tasks have gathered a lot of attention since the decade. Consequently, the researchers in this domain has proposed various feature extraction, feature learning, and feature encoding algorithms for improving the generalization performance of the aforesaid tasks. For example, the generalization performance of the image classification models mainly depends on the choice of data representation. These models aim at building comprehensive representation learning (RL) strategies to encode the relationship among the input and output attributes from the raw big data.
Existing RL strategies can be divided into three general categories: statistic approaches (e.g. probabilistic-based analysis, and correlation-based measures), unsupervised learning (e.g., autoencoders), and supervised learning (e.g., deep convolutional neural network (DCNN)). Among these categories, the unsupervised and supervised learning strategies using artificial neural networks (ANNs) have been widely adopted. In this direction, several auxiliary ideas have been proposed over the past decade, to improve the learning capability of the ANNs. For instance, Moore-Penrose (MP) inverse is exploited to refine the parameters (weights and biases) of a trained network. However, the existing MP inverse-based RL methods have an important limitation. The representations learned through the MP inverse-based strategies suffer from loosely-connected feature coding, resulting into a poor representation of the objects having lack of discriminative power. To address this issue, this dissertation proposes a set of eight novel MP inverse-based RL algorithms.
The first part of this dissertation from Chapter 4 to Chapter 7 is dedicated to proposing novel width-growth models based on subnet neural network (SNN) for representation learning and image classification. In this part, a novel feature learning algorithm, coined Wi-HSNN is proposed, followed by an improved batch-by-batch learning algorithm, called OS-HSNN. Then, two novel SNNs are introduced to detect extreme outliers for one-class classification (OCC). Finally, a semi-supervised SNN, named SS-HSNN is introduced to extend the strategy from the supervised learning domain to the semi-supervised learning domain.
The second part of this thesis, subsuming Chapter 8 and Chapter 9, focuses on improving the performance of the existing multilayer neural networks through harnessing the MP inverse. Here, a novel weight optimization strategy is proposed to improve the performance of multilayer extreme learning machines (ELMs), where the MP inverse is used to feedback the classification imprecision information from the output layer to the hidden layers. Then, a novel fast retraining framework is proposed to enhance the efficiency of transfer learning of DCNNs.
The effectiveness of the proposed subnet- and retraining-based algorithms have been evaluated on several widely used image classification datasets, such as ImageNet and Places-365. Furthermore, we validated the performance of the proposed strategies in some extended domains, such as ship-target detection, food image classification, camera model identification and misinformation identification. The experimental results illustrate the superiority of the proposed algorithms.
Zhang, Wandong, "ADVANCED REPRESENTATION LEARNING STRATEGIES FOR BIG DATA ANALYSIS" (2022). Electronic Theses and Dissertations. 8911.