Common and Unique Feature Learning for Data Fusion
- Publication Type:
- Thesis
- Issue Date:
- 2020
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
In today’s era of big data, information about a phenomenon of interest is available from multiple acquisitions. Data captured from each of these acquisition frameworks are commonly known as modality, where each modality provides information in a complementary manner. Despite the evident benefits and plethora of works on data fusion, two challenging issues persist, 1) feature representation: how to exploit the data diversity that multiple modalities offer, and 2) feature fusion: how to combine the heterogeneous information for better decision making.
To address these challenges, this thesis presents a significantly improved model of two widely utilised fusion techniques, a) early fusion: combining features from multiple modalities for joint prediction, and b) late fusion: combining modality-specific predictions at the decision level. I illustrate how both these techniques have their own specific limitations, with late fusion unable to harness the inter-modality benefits, and the reliance of early fusion on a single model causing failure when information from any modality is futile. To overcome these drawbacks, I developed novel multimodal systems that performs feature extraction and feature fusion in a consolidated frameworks. Technically, I designed feature extraction schemes to capture both unique information from individual modalities and common information from multimode representations. I then combine these two kinds of information for supervised prediction, by designing efficient fusion schemes that enable this frameworks to perform information discovery and feature fusion simultaneously.
In this thesis, I also demonstrated the benefits of fusing both the common and unique information in supervised learning and validate the significance of the developed techniques on multimodal, multiview, and multisource datasets. The designed methods leverage the multimodal benefits by creating additional diversity, and obtain a more unified view of the underlying phenomenon for better decision making.
Please use this identifier to cite or link to this item: