Tutorial Session I: 15:00 ~ 16:00, February 21, 2022 (Monday)
Prof. Katsuya Suto
Deep Learning and Its Applications to Radio Map Construction
Abstract:
Radio map plays a key role in decision-making in 6G systems, i.e., resource management for cell-free wireless networks, spatial spectrum sharing, intelligent reflecting surface (IRS). However, it remains an open challenge. Deep learning (especially image-driven deep learning) has been developing as a promising solution to express the complex radio propagation features in the urban area using feature extraction from 3D maps of cities. The approach learns the correlation between the building features and propagation features to recognize the reflection and diffraction by the buildings. By use of rapid advancement of GPU, it achieves high estimation accuracy with low computation time.
The main objective of this tutorial is to provide a fundamental background of deep learning and then show how to address practical challenges in radio map construction. In particular, we first give a tutorial of deep learning used in radio map construction to provide comprehensive knowledge to the audiences. We then give the current research trend together with implementation details to have a better understanding. Finally, we introduce our proposed methods for path loss modeling, spatial interpolation, and spatial extrapolation.
Biography
Katsuya Suto received the B.Sc. degree in computer engineering from Iwate University, Morioka, Japan, in 2011, and the M.Sc. and Ph.D. degrees in information science from Tohoku University, Sendai, Japan, in 2013 and 2016, respectively. He has worked as a Postdoctoral Fellow for Research Abroad, Japan Society for the Promotion of Science, in the Broadband Communications Research Lab., University of Waterloo, ON, Canada, from 2016 to 2018. He is currently an Assistant Professor with the Graduate School of Informatics and Engineering, the University of Electro-Communications, Tokyo, Japan. His research interests include mobile edge computing, cognitive radio, green wireless networking, and deep learning. He received the Best Paper Award at the IEEE VTC2013-spring, the IEEE/CIC ICCC2015, the IEEE ICC2016, and the IEEE Transactions on Computers in 2018.
Tutorial Session II: 14:00 ~ 15:00, February 22, 2022 (Tuesday)
Prof. Yexiang Xue
Knowledge Embeddings to Attack Multi-stage Inference Problems in Reasoning, Learning, and Decision Making
Abstract:
Problems at the intersection of reasoning, optimization, and learning often involve multi-stage inference and are therefore highly intractable. I will introduce a novel computational framework, based on embeddings, to tackle multi-stage inference problems. As a first example, I present a novel way to encode the reward allocation problem for a two-stage organizer-agent game-theoretic framework as a single-stage optimization problem. The encoding embeds an approximation of the agents’ decision-making process into the organizer’s problem. We apply this methodology to eBird, a well-established citizen-science program for collecting bird observations, as a game called Avicaching. Our AI-based reward allocation was shown highly effective, surpassing the expectations of the eBird organizers and bird conservation experts. As a second example, I present a novel constant approximation algorithm to solve the so-called Marginal Maximum-A-Posteriori (MMAP) problem for finding the optimal policy maximizing the expectation of a stochastic objective. To tackle this problem, I propose the embedding of its intractable counting subproblems as queries to NP-oracles subject to additional XOR constraints. As a result, the entire problem is encoded as a single NP-equivalent optimization. The approach outperforms state-of-the-art solvers based on variational inference as well as MCMC sampling on probabilistic inference benchmarks, deep learning applications, as well as on a novel decision-making application in network design for wildlife conservation. Lastly, I will talk about how the embeddings of phase-field modeling in an end-to-end neural network allow us to learn partial differential equations governing the dynamics of nanostructures in metallic materials under extreme heat and irradiation conditions.
Biography
Dr. Yexiang Xue is an assistant professor at the Department of Computer Science at Purdue University, USA. The goal of Dr. Xue’s research is to bridge large-scale constraint-based reasoning and optimization with state-of-the-art machine learning techniques to enable intelligent agents to make optimal decisions in high-dimensional and uncertain real-world applications. More specifically, Dr. Xue’s research focuses on scalable and accurate probabilistic reasoning techniques, statistical modeling of data, and robust decision-making under uncertainty. Dr. Xue’s work is motivated by key problems across multiple scientific domains, ranging from artificial intelligence, machine learning, renewable energy, materials science, crowdsourcing, citizen science, urban computing, ecology, to behavioral econometrics. Dr. Xue focuses on developing cross-cutting computational methods, with an emphasis on the areas of computational sustainability and scientific discovery. Dr. Xue received several NSF grants, Purdue’s seed of success award, Cornell’s Ph.D. dissertation award, and the IAAI Innovative application award. He published over 45 papers at top-tier CS conferences, and also journal articles including in Science, Nature Communications, the communications of ACM, Materials Research Society Communications, and the Artificial Intelligence magazine.
Tutorial Session III: 14:00 ~ 14:50, February 23, 2022 (Wednesday)
Prof. Dong Seog Han
Facial Emotion Recognition with Deep Learning
Abstract:
Facial emotion recognition (FER) is vital for interactive robots detecting users’ feelings. The solid performance of the FER requires a well-designed neural network and a reliable FER dataset. Managing the FER dataset is highly effective in having a solid performance than redesigning the neural network. These days, many FER researchers are more focused on designing a deep learning model without thoroughly inspecting the FER dataset samples. In addition, the FER without improper pre-processing of the FER dataset could cause degrading the deep learning model’s performance even with a well-designed neural network. Some FER datasets contain irrelevant facial images or unnecessary features, confusing a deep neural network’s training. In this tutorial, we demonstrate how properly pre-process the FER dataset to enhance the overall quality of the FER dataset and improve the performance of FER’s training.
Biography
Dong Seog Han received the B.S. degree in electronic engineering from Kyungpook National University (KNU), Daegu, Korea, in 1987, and the M.S. and Ph.D. degrees in electrical engineering from the Korea Advanced Institute of Science and Technology (KAIST), Daejon, Korea, in 1989 and 1993, respectively. From 1987 to 1996, he was with Samsung Electronics Company Ltd., where he developed the receiver chipset for HDTV. Since 1996, he has been with the School of Electronics Engineering, KNU, as a faculty and is currently a full Professor. He was a courtesy Associate Professor with the Department of Electrical and Computer Engineering, University of Florida, in 2004. He was the Director of the Center of Digital TV and Broadcasting, Institute for Information Technology Advancement (IITP), from 2006 to 2008. He is currently directing the Center for ICT & Autonomous Convergence, KNU, since 2011. His main research interests include intelligent signal processing and autonomous vehicles.