Bridging the Human-Machine Gap in Applied Machine Learning with Visual Analytics

Machine learning is becoming a ubiquitous toolset for analyzing and making use of large collections of data. Advanced learning algorithms are able to learn from complex data to build models that can tackle artificial intelligence tasks previously thought impossible. As a result, organizations from many domains are attempting to apply machine learning to their data analysis problems. In practice, such efforts can suffer from gaps between the goals of the human and the objective being optimized by the machine, resulting in models that perform poorly in deployment. In this dissertation, I present my thesis that visual analytics systems can improve the performance of deployed models in applied machine learning tasks by allowing the user to compensate for vulnerabilities in learning paradigms. First, I outline how learning paradigms used by machine learning algorithms can miss out on certain aspects of the end goal of the user. Then, I will describe four different visual analytics systems that allow the user to intervene in the learning process across many types of data and models. In these systems, visualizations help users understand how a model performs on different regions of the data. They can also help a user encode their domain expertise to the learning algorithm, to correct for misalignments between the goals of the machine and the needs of the application scenario. This work offers evidence that advanced machine learning algorithms are applied more effectively by involving a domain user in the learning process, using a visual analytics tool as a medium