Abstract
Great advancements in recent years on machine learning and artificial intelligence (AI) development contributed by huge progresses in machine computation as well as data availability have drawn new waves of complex applications across multiple industries. As machine learning models going larger, deeper, and higher performance, elements of operation or decision remain at highly non-linear black box status which create many trust and acceptance issues at user level. This thesis studies the topic of explainable AI from perspectives of various deep neural network (DNN) architectures or deep learning applications. Initial discussion includes some of the difficulties driving the need of explainable AI, and general approaches available that address different aspects of machine learning for DNN. A first investigation on one explainable AI technique which interprets network classification result by Layer-wise Relevance Propagation (LRP) was conducted to demonstrate effectiveness of interpretation in explaining convolutional neural network (CNN). The investigation involved a novel method of integrating LRP analysis within a machine learning loop to create new augmented training data based on analyzed relevancy of each input data features towards the final network classification result. A new LRP automation toolchain was subsequently developed to seamlessly recreate an auxiliary network for relevance back-propagation. This toolchain enabled in-depth analysis of several deep neural network architectures specifically for image classification application. Through the analysis, a few interesting observations concerning network bias, latent ski player features, and feature sign/magnitude were uncovered. Furthermore, the analysis demonstrated some limitations of current LRP interpretation method for such deep networks including the need of preset of multiple decomposition rules tuned to specific network architecture or layers, and a gap in term of sensitivity comparing to basic gradient methods. With better understanding from analysis and observations, this thesis further proposes a new unified decomposition method based on an enhanced relevance model targeting at improving robustness across different DNN architectures and models with clean accurate interpretations. Finally, the thesis discusses some potential applications including adversarial attacks