Singh, Vinit Veerendraveer2023-02-232023-02-232022https://udspace.udel.edu/handle/19716/32348The success of Convolutional Neural Networks can be primarily attributed to considering the underlying grid representation of images and the local neighborhoods around pixels. Another critical factor for their success is to implicitly and hierarchically attend to task-critical regions in the images. However, extending techniques and components from image-based Convolutional Neural Networks to higher dimensional data is non-trivial. In addition, deeper Convolutional Neural Networks are required to attend to the task-critical regions implicitly, thus increasing their computational complexity. ☐ The goal of this dissertation is to present novel techniques and components for neural networks that harness the data’s underlying representation for computer vision tasks. It also presents attention modules that aid the neural network in explicitly attending to regions that are most likely to improve their performance on predictive modeling problems such as object classification and object segmentation. I first present an attention mechanism for feature maps of neural networks that takes advantage of the underlying grid representation of feature maps. When it was introduced to pre-trained Convolutional Networks’ bottlenecks, consistent improvements in accuracy were observed for image classification. Second, I extended dilated convolutions from the image domain to 3D mesh representations. Then, I utilized dilated mesh convolutions to build novel attention and pooling mechanisms in neural networks for mesh classification. State-of-the-art results were achieved. Lastly, I present an attention-inspired and graph-representation-based data augmentation approach that generalizes to n-dimensional space. Utilizing this augmentation approach with neural networks improved their performance for image classification and 3D shape analysis tasks.Attention-aided neural networksData representationObject classificationObject segmentationThesis1371056999https://doi.org/10.58088/6vhq-7r732022-09-21en