题目：A Framework for Building Intelligent User Interfaces in Interactive Machine Learning
摘要：Intelligent user interfaces join human and machine strengths. They are critical for usage scenarios where purely relying on human intelligence is laborious and costly, while purely resorting to machine intelligence leads to imprecise and untrustworthy results. Despite the merits of these interfaces, the development and evaluation of them are challenging. The development typically requires software developers to have multi-disciplinary knowledge spanning human-computer interaction, visualization, machine learning, and solid software engineering skills. The evaluation of these interfaces requires human subjects, which is expensive and unscalable. My research in human-computer interaction and data visualization contributes methods to develop, apply, and evaluate intelligent user interfaces. OneLabeler is a visual programming interface that scaffolds the development of intelligent user interfaces for machine-aided data labeling. It supports efficient building of interactive systems that feature mixed human-machine computation workflows. MI3 is a generic machine-aided workflow for classification. It has been instantiated as an intelligent user interface for the digital humanity application of data reconstruction from historical visualizations. Simulation-based evaluation is proposed for efficient task completion time estimation in interfaces that embed machine-aided data labeling workflows.
题目：Deep Learning for Scientific Data Representation and Generation
摘要：Scientific visualization is one of the core components in supporting fundamental sciences. For example, scientists perform numerical simulations and produce 3D scalar and vector data to visualize, analyze, and understand various kinds of natural phenomena, such as climate change and chemical reactions. As the cost of simulations is expensive when time, ensemble, and multivariate are involved and the scientific data are presented in diverse forms including streamline, pathline, stream surface, volume, and isosurface, a core problem is to how to efficiently produce and analyze these diversified data. In this talk, I will explore the possibilities of deep learning solutions, such as convolutional neural networks and generative adversarial networks, for visual representation encoding and scientific data generation. In particular, I will talk about how to use generative models for synthesizing variables for the purpose of data extrapolation. I will also discuss a unified framework for line and surface clustering and selection using autoencoder.
题目：FORSETI and THEMIS: Visual Analysis Environments for Computational Forensics
摘要：Forensic science holds a unique position among all other disciplines due to its important social implications-lying at the intersection of medical science and legal science. Indeed, forensic autopsy reports, which systematically describe autopsy findings, are imperative both for legal medicine and courts. Typically, they are created by medical examiners (MEs) collaboratively working with diagnostic radiologists (DRs). The forensic autopsy reports also serve as underlying legal documents in judicial trials for engaged judicial personnel (JP). A challenging issue is to provide effective computational assistance tools for facilitating the intricate collaborative work involved in the autopsy. For this issue, we design an integrated visual analysis system named FORSETI (forensic autopsy system for e-court instruments) and its auxiliary system called THEMIS (theoretical estimation of meaning of insults). The FORSETI system was designed to empower MEs and DRs to collaborate and author effective forensic autopsy reports by a combination of an extended version of legal medicine mark-up language and autopsy juxtaposition. On the other hand, the THEMIS system helps MEs and DRs as well as JP determine the cause of death by computing the context-sensitive similarities of surface wounds based on the Mathematical Model of Meaning. In this talk, we will clarify the forensic autopsy objectives and their corresponding computational tasks and give an overview of both systems. Prospective research and development issues are also discussed.
题目：A New Assemblage: Image-Data Based Interactive Assemblage of A Human-Machine Reality
摘要：Assemblage is an artistic practice where found or unrelated objects, and different materials are juxtaposed to result in new entities, which often suggest non-linear narratives, poetic meanings, and new symbols. The ideas and methods of assemblage, based on concepts of connectivity, narration, materiality, and chance and choice operations, emerged during the twentieth century in the early stage of modernism and were closely involved in modern art movements like Cubism, Surrealism, and dada. Current technological developments, including image classification, image generation, and human-computer interactions, offer new opportunities to reconsider the concept of assemblage. My research reformulates the concept of assemblage within an AI and interactive spatial visualization context via the questions of authorship, connectivity, and materiality. In this context, I conduct practice-based research and propose a methodology of a new assemblage: creating a new entity in an artist-machine-audiences collaborative manner. These intelligent systems transform image data into a responsive and evolving multi-media real-time configuration. My research, along with my media art projects, contribute to the fields of experimental visualization, interactive art, and AI art at the intersections of the real and the virtual through current methods by which to engage machine vision in relation to human perception and the impact of automation on originality.