Invited Tutorials

An introduction to deep learning

Paul Hollensen and
Thomas P. Trappenberg
Faculty of Computer Science,
Dalhousie University, Halifax, Canada

Deep learning is a family of methods that have brought machine learning applications to new heights, winning many competitions in vision, speech, NLP, etc. The centrepiece of this method is the learning of many layers of features to represent data at increasing levels of abstraction. While this has immediate consequences for the quality of classification and regression results, there is also some excitement in the AI community with the possibility to fuse such connectionist AI with symbolic AI.
This tutorial is an introduction to deep learning. We will motivate the excitement in this field with a survey of recent state-of-the-art results, and we will outline some of the theory behind representational learning. We will then discuss a small implementation of a convolutional network before discussing some of the software frameworks that make these powerful methods accessible and practical, including how these methods can be accelerated by inexpensive graphics hardware from high-level, productive languages. We will also highlight exemplary examples of published models with freely available implementations that can serve as a starting point for both advancing theory and application to new domains.

Paul Hollensen studied Cognitive Science at the University of Toronto followed by Computer Science at Dalhousie University where he is a Ph.D. candidate. His principal research interests are computational neuroscience and hierarchical machine learning methods. He worked extensively with unsupervised and deep learning methods and several implementation platforms.
Thomas P. Trappenberg received a PhD in Physics from RWTH Aachen University, Germany, and he held research positions at Dalhousie University, Canada, the RIKEN Brain Science Institute, Japan, and Oxford University, England. He is a full professor in Computer Science and author of Fundamentals of Computational Neuroscience, now in its second edition. His research interests are computational neuroscience, machine learning, and neurocognitive robotics.

User-Centered Text Mining

Axel J. Soto and Evangelos E. Milios
Faculty of Computer Science,
Dalhousie University, Halifax, Canada

Historically, text mining methods for extracting “knowledge” from text have increased in sophistication by the incorporation of both statistical learning and symbolic natural language processing. However, in scenarios where a domain user aims to make sense of document collections and derive insights from them, domain knowledge is necessary to inform the analysis, and text mining needs to be complemented by text visualization and user-driven interaction with the analytic process. In this tutorial we introduce Visual Text Analytics as a multi-disciplinary field of research. We cover conceptual and practical methods and tools,
review state-of-the-art systems that integrate text mining with visualization and user interaction, and identify promising research directions for making text mining more user-centered and accessible to users with an interest in domain-specific applications.

Axel J. Soto is a Research Associate and Adjunct Professor with the Faculty of Computer Science at Dalhousie University (Canada). He received his B.Sc. in Computer Systems engineering and his PhD in Computer Science at Universidad Nacional del Sur (Argentina) in 2005 and 2010, respectively. He now conducts research in the area of Visual Text Analytics. Most of his research focuses on the area of multivariate statistics, machine learning and visualization.
Evangelos E. Milios received a diploma in Electrical Engineering from the NTUA, Athens, Greece, and Master’s and Ph.D. degrees in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology. Since July of 1998 he has been with the Faculty of Computer Science, Dalhousie University, Halifax, Nova Scotia, where he has been Associate Dean – Research since 2008. He currently works on modelling and mining of content and link structure of Networked Information Spaces, text mining and visual text analytics.

Rough Sets in KDD

Dominik Slezak
Institute of Mathematics, University of Warsaw, Poland &
Infobright Inc., Poland/Canada

Rough sets provide foundations for a number of methods useful in data mining and knowledge discovery, at different stages of data processing, classification and representation. Thus, it is worth expanding rough set notions and algorithms towards real-world environments.
We attempt to categorize some ideas of how to scale and apply rough set methods in practice. As case studies, we refer to research projects related to text processing, risk management and sensory measurements, conducted by the rough set team at University of Warsaw. We also give
examples of commercial applications, which follow (or may follow) rough set principles in several areas of data analytics. Finally, we demonstrate functionality of some of existing rough set data mining packages, which can be helpful in academic research and for industry purposes.

Dominik Slezak received his Ph.D. in 2002 from University of Warsaw and D.Sc. in 2011 from Institute of Computer Science, Polish Academy of Sciences. In 2005 he co-founded Infobright, where he holds position of chief scientist. He is also associate professor at University of Warsaw. He worked as assistant professor at Polish-Japanese Academy of Information Technology and at University of Regina. He delivered invited talks at over 20 international conferences. He co-organized a number of scientific events, including a role of general program chair of Web Intelligence Congress 2014. He is in editorial board of Springer’s Communications in Computer and Information Science and serves as associate editor in several journals. In 2014-2016 he is responsible for conference sponsorships in IEEE Technical Committee on Intelligent Informatics. In 2012-2014 he served as president of International Rough Set Society. His interests include databases, data mining and soft computing. He co-authored over 150 papers. He is also co-inventor in five granted US patents.