Shared talk: Friday, July 6th, 9:00-10:30
Title: Visual Analytics Opens the Black Box of Machine Learning
Abstract: In recent years, machine learning, and in particular deep learning, have made tremendous progresses both on theoretical and practical levels. Currently, machine learning algorithms are used by wide groups of practitioners, ranging from computer scientists, researchers in other exact science branches, technical people in data-intensive industries, and the grand public. However, with its successes, machine learning also has an Achille's heel: The operation of its algorithms is often opaque, and thus hard to understand, predict, and fine-tune. This raises serious problems for the effectiveness, replicability, trustworthiness, and ultimately applicability of machine learning in practical problems.
In this talk, we will discuss a broad range of techniques from information visualization and visual analytics that can be used to open this black box of machine learning for its interested users. Thereby, the general operation, unstable cases, learning behavior, and potential challenges encountered by machine learning techniques become visible, controllable, and actionable upon by its users. We will address questions such as: How to quickly determine if a classification problem is hard? How to understand what a neural network has learned (or not)? How to see which parts of some input data made a neural network take a given decision? And, above all, why do we need to understand deep learning black boxes for our businesses? Examples from real-world usage of visual analytics to help machine learning are presented coming from multiple application domains (image analysis, medicine, software engineering, and physics).
Short bio: Prof. Alexandru C. Telea has obtained his PhD in 2000 in data visualization and software architectures at the University of Eindhoven, the Netherlands. He has worked as assistant professor in visualization and computer graphics at the same university until 2007, when he has been appointed professor in software visualization at the University of Groningen, the Netherlands. His interests cover data, information, and software visualization, and static source code analysis and C/C++ reverse engineering. He has published over 130 papers in international peer-reviewed venues and one textbook in data visualization. He served as co-chair of IEEE EuroVis 2008, IEEE Vissoft 2007, ACM Softvis 2007 and general chair of ACM Softvis 2009. He is recognized as one of the main experts in the field of software visualization. His research has a strongly applied flavor, and he has co-developed several innovative tools and techniques for software visualization in the industry. He is a member of the ACM.