Jürgen Bernard, Computer Science at UBC, Vancouver
Enhancing Interactive Machine Learning
December 10th, 2019, 4:00 pm CET
Room: Science Park Building 3, Room 063
Abstract:
Machine Learning (ML) is a key technology in the era of data-centered research and applications.
Given the availability of data, a series of data-driven learning tasks have successfully been formalized in the past and made ML-capable. As a result, many automated ML-based solutions already ease our everyday life or gain other types of promising benefits.
Interactive Machine Learning is a principle that may be applicable for the plethora of data-driven learning tasks for which formalizations and automations could not be achieved so far. Interactive ML explicitly includes the involvement of humans in the iterative and incremental ML process, following the idea to combine the strengths of both humans and machines to create more effective and efficient solutions. There is considerable evidence that many remaining data-driven learning challenges can be tackled with interactive ML principles. Example include data-oriented challenges like heterogeneous data, dirty data, uncertain data, or unlabeled data, as well as process-oriented challenges such as data preprocessing, model building, model parameterization, model quality assessment, model uncertainty assessment, model refinement, model explanation and interpretation, or model comparison. These performance-driven challenges are accompanied by at least two human-centered goals worth to be pursued: 1) creating personalized ML solutions tailored to the information need of individual users or application tasks, as well as 2) making ML applicable for larger user groups far beyond the group of data science experts.
This talk motivates the human-centered and interactive perspective of ML. Using two of my research branches as examples, I will show how users can be enabled to play an active role in the ML process and be an integral part of highly interactive and iterative training and validation cycles. The first example is the Segmentation and Labeling of Multivariate Time Series (VISSECT), a visual analytics approach that investigates the interplay between ML models, model parameters, and uncertainty in a systematic way. The second example includes Visual-Interactive Labeling (VIAL), an approach to solve the data labeling problem. The data labeling problem refers to the lack of labels in the substantial majority of datasets existing today. Finding solutions to label data effectively will make today’s unlabeled data feasible for supervised ML such as classification and regression.
About the Speaker:
Jürgen Bernard is a postdoctoral research fellow at the University of British Columba, Vancouver. He has joined the InfoVis group, led by Professor Tamara Munzner in 2019. Jürgen Bernard studied Computer Sciences with focus on Computer Graphics and Bio Technology at the University of Technology of Darmstadt. He received his PhD Degree in 2015, when he was with Fraunhofer IGD. His thesis “Exploratory Search in Time-Oriented Primary Data” was awarded with the Hugo-Geiger Prize for excellent PhD theses, granted by the Fraunhofer Foundation. In 2016, Jürgen Bernard started as a Post-doc researcher at TU Darmstadt at Interactive Graphics Systems Group (Computer Science), leading his own research group (Visual-Interactive Machine Learning). In 2017, he received the Dirk Bartz Prize for Visual Computing in Medicine, granted by the Eurographics Association (EG).