With the recent advancements of machine learning, especially deep learning, we have seen fast-growing applications of these intelligent systems in various domains. However, the increasing complexity of these systems makes it very challenging to explain or interpret their reasoning process, which limits their adoption in critical decision-making scenarios. In the meantime, visualization has been effectively applied to support the understanding and analyzing of complex systems and large data collections. In this thesis, we study how to make machine learning systems explainable for human users using visualizations.

We first propose a user-model interaction framework for describing and categorizing the explainable machine learning problem. Then we discuss the role of visualization i...[ Read more ]

With the recent advancements of machine learning, especially deep learning, we have seen fast-growing applications of these intelligent systems in various domains. However, the increasing complexity of these systems makes it very challenging to explain or interpret their reasoning process, which limits their adoption in critical decision-making scenarios. In the meantime, visualization has been effectively applied to support the understanding and analyzing of complex systems and large data collections. In this thesis, we study how to make machine learning systems explainable for human users using visualizations.

We first propose a user-model interaction framework for describing and categorizing the explainable machine learning problem. Then we discuss the role of visualization in explainable machine learning, including How, Where, and Why visualization could be used to help explain What parts of the machine learning pipeline to Whom. We also summarize the recent research advances in this field.

We then grounded our study of different aspects of the explainable problem on specific applications: 1) how can visualization help explain the inner working mechanisms of deep learning models for model developers and researchers? 2) how can we explain the behavior of a model for non-expert users with little knowledge in machine learning? 3) how can explainability help expert users in various application domains to incorporate domain knowledge into the model? We experiment with these ideas under a human-in-the-loop setting and include preliminary evaluation results in this thesis. At last, we discuss our ongoing and future research as well as open questions in visualization for explainable machine learning.