Browsing by Author "Neocosmos, Kibidi"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Image Classification with Graph Neural Networks
    (Stellenbosch : Stellenbosch University, 2022-04) Neocosmos, Kibidi; Bah, Bubacarr; Stellenbosch University. Faculty of Science. Dept. of Mathematical Sciences.
    ENGLISH SUMMARY: Convolutional neural networks (CNNs) are a prominent and ubiquitous part of machine learning. They have successfully achieved consistent state-of-the-art performance in areas such as computer vision. However, they require large datasets for such achievements. This is in stark contrast to human-level performance that demands less data for the same task. The question naturally arises as to whether it is possible to develop models that require less data without a significant decrease in performance. I n t his thesis, we address the above question from a different perspective by investigating whether a richer data structure could result in more learning from fewer training examples. We explore the idea by constructing images as graphs – a structure that naturally contains more information about an image than the standard tensor representation. We then use graph neural networks (GNNs) to leverage the graph structure and perform image classification. We found that the graph structure did not enable GNNs to perform well given less data. However, during the process of experimentation, we discovered that the graph topology as well as node features significantly influence performance. Furthermore, some of the proposed GNN models were not able to effectively utilize the graph structure.