With convolutional neural networks revolutionizing the computer vision field it is important to extend the capabilities of neural-based systems to dynamic and unrestricted data like graphs. Doing so not only expands the applications of such systems, but also provide more insight into improvements to neural-based systems.
Currently most implementations of graph neural networks are based on vertex filtering on fixed adjacency matrices. Although important for a lot of applications, vertex filtering restricts the applications to vertex focused graphs and cannot be efficiently extended to edge focused graphs like social networks. Applications of current systems are mostly limited to images and document references.
Beyond the graph applications, this work also explored the usage of convolutional neural networks for intelligent character recognition in a novel way. Most systems define Intelligent Character Recognition as either a recurrent classification problem or image classification. This achieves great performance in a limited environment but does not generalize well on real world applications. This work defines intelligent Character Recognition as a segmentation problem which we show to provide many benefits.
The goal of this work was to explore alternatives to current graph neural networks implementations as well as exploring new applications of such system. This work also focused on improving Intelligent Character Recognition techniques on isolated words using deep learning techniques. Due to the contrast between these to contributions this documents was divided into Part I focusing on the graph work, and Part II focusing on the intelligent character recognition work.
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Petroski Such, Felipe, "Deep Learning Architectures for Novel Problems" (2017). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus