Enhancing Dimensionality Reduction Techniques for Deep Neural Network Visualization

Deep Neural Networks have achieved impressive results in a wide range of applications over the past few years, being responsible for many advances in computational technology. However, debugging and understanding the inner workings from these models is a complex task, as there are often millions of variables involved in every decision. Aiming to solve this problem, researchers from the fields of Visual Analytics and Explainable Artificial Intelligence have proposed several approaches to visualize and explain different aspects of DNN models. One of such approaches is the use of Dimensionality Reduction techniques for hidden layer output visualization, which has been employed in literature with relative success. However, there are certain limitations to applying these techniques in this context that need to be addressed, such as the visual comparison between multiple multidimensional projections. Furthermore, the particular characteristics of this domain can be taken into account to generate specialized visual representations that are more informative. This doctorate thesis shows the process of investigating problems and opportunities in DNN visualization using dimensionality reduction and the development of improved visualization methods for this domain. (AU)