SMARTCOMP 2022

Tutorial 1: Make machine learning more efficient using external knowledge: A guide to transfer learning

Date: Friday June 24, 2022
Time: 10:30-12:00 EEST

Abstract

Transfer learning (TL) has attracted much attention as a machine learning methodology for efficient adaptation to various real-world data using external knowledge. In this tutorial, we will first introduce problem settings of TL. For homogeneous settings in which source and target domains share the same input spaces but the data distribution differs, we explain TL methods, e.g., instance-based, parameter-based, and feature-based methods. Then, we explain extension of feature-based methods to heterogeneous settings in which source and target domains have different input spaces. For some typical methods, we will also show input and output examples.

Schedule

Presentation: 75 mins + QA: 10 mins

Topic Duration Speaker(s)
Basis of transfer learning (TL) 15 mins Kota Matsui (Nagoya University)
Scenarios and approaches of TL 10 mins Mori Kurokawa (KDDI Research, Inc.)
Homogeneous TL settings:
   Instance-based TL methods 10 mins Mori Kurokawa (KDDI Research, Inc.)
   Parameter-based TL methods 10 mins Zhi Li (Osaka University)
   Feature-based TL methods 10 mins Kei Yonekawa (KDDI Research, Inc.)
Heterogeneous TL settings:
   Feature-based TL methods for heterogeneous settings 10 mins Kei Yonekawa (KDDI Research, Inc.)
Future directions of TL 10 mins Mori Kurokawa (KDDI Research, Inc.); Kota Matsui (Nagoya University)
Q&A 10 mins All

Organizer

Mori Kurokawa (KDDI Research, Inc.)

Speakers
Kota Matsui Nagoya University Graduate School of Medicine
Kota Matsui is a lecturer at Nagoya University Graduate School of Medicine, Department of Biostatistics. He received his Ph.D. from Nagoya University in 2017. His research focuses on the development of machine learning methods for small-size data and its application to medical and material problems. Especially he is interested in active learning and transfer learning for small-data problem.
Zhi Li Osaka University

Zhi Li received B.E. degree from Nankai University, Tianjin, China, in 2017, and the M.E. degree in information systems engineering from Osaka University, Osaka, Japan, in 2021. He is currently pursuing the Ph.D degree in information systems engineering from Osaka University, under the supervision of Prof. T. Hara. His research interests include data mining, deep learning-based recommender system and multi-modal learning-based sentiment analysis.

Kei Yonekawa KDDI Research, Inc.
Kei Yonekawa received the B.S. and M.S. degrees from the Department of Electrical Engineering, University of Tokyo, in 2012 and 2014, respectively. In 2014, he joined KDDI Corporation, where he was engaged in the operation of the infrastructure of cloud service. In 2015, he joined KDDI Research, Inc., where he is currently an researcher. His research interests include machine learning, data mining, transfer learning, and MLOps.
Mori Kurokawa KDDI Research, Inc.
Mori Kurokawa received the B.S. and M.S. degrees from the Faculty of Science and Technology, Keio university, in 2005 and 2007, respectively. In 2007, he joined KDDI Corporation and he is currently head of Integrated Machine Learning Laboratory in KDDI Research, Inc. His research interests include machine learning, transfer learning, knowledge graph embedding, recommender system, and quantum computing.
Audience expectation and prerequisites

Tutorial 2: Gesture recognition

Date: Friday June 24, 2022
Time: 14:00-17:30 EEST

Abstract

Gesture recognition is a natural and device-free way of non-verbal communications. It has many use cases in different applications ranging from smart-homes to human-robot interactions. Different modalities can be used to realize a gesture-based interface from RGB cameras to LiDARs and Radars. Radars, specifically mmWave radars, are robust to different lighting and weather conditions, they have small form factor, and given the high frequency range, they can detect subtle movements. In this tutorial session, first, we will discuss the basic principles of mmWave Frequency-Modulated Continuous-Wave (FMCW) radars. Then, the data representation (Point Cloud) of the radar will be introduced and the challenges of processing this type of data will be discussed. Finally, using graph-based neural networks, we will train a model to recognize gestures by processing the point clouds generated by the radars. The in-person participants will have the opportunity to work with the radars that they will be provided to record real-world data and process them using the pipeline we will introduce during the tutorial.

Organizer
Dariush Salami Aalto University
Dariush Salami received his BSc and MSc degrees from Shahid Beheshti University and Amirkabir University of Technology in Software Engineering in 2016 and 2019, respectively. He is currently a Marie Skłodowska Curie fellow in ITN-WindMill project and a PhD researcher at the department of communications and networking at Aalto University. He is mainly focused on Machine Learning for Wireless Communications and Sensing especially in mmWave range using Frequency-Modulated Continuous-Wave (FMCW) radars for Human Centric Sensing (HCS).
Stephan Sigg Aalto University
Stephan Sigg received his M.Sc. degree in computer science from TU Dortmund, in 2004 and his Ph.D. degree from Kassel University, in 2008. Since 2015 he is an assistant professor at Aalto University, Finland. He is a member of the editorial board of the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies as well as of the Elsevier journal of Computer Communications. He has served as a TPC member of renowned conferences including IEEE PerCom, IEEE ICDCS, etc. His research interests include Ambient Intelligence, in particular, Pervasive sensing, activity recognition, usable security algorithms for mobile distributed systems.
Audience expectation and prerequisites