Session 1
Hardware Security Circuits for Reliable Key Generation and Protection
Tsung-Te Liu
Department of Electrical Engineering, National Taiwan University, Taiwan
https://sites.google.com/eecs.ee.ntu.edu.tw/eecslab
Hardware security circuits are critical components to realize future secure Internet of Things (IoT) devices. Among them, the embedded key generation circuit serves as the very fundamental element to support various cryptography algorithms and protocols. In this talk, a reliable key generation circuit based on physically unclonable function (PUF) will be presented, along with different design techniques to enhance its robustness and maintain its secrecy.
AI Security from Hardware Perspective
Naofumi Homma
Research Institute of Electrical Communication, Tohoku University, Japan
https://www.riec.tohoku.ac.jp/en/organization/section4/homma/
This talk introduces AI hardware security and some recent related results. Information security related to AI, so-called AI security, is drawing much attention in the field of security research. AI security is classified into “Security for AI” and “AI-assisted security.” The number of AI security studies also rapidly increases in the hardware security field. In this talk, we describe an overview of AI hardware security research and show some ongoing related studies.
Session 2
Augmented Robotic Content Interaction via Egocentric Spatial User Interface
Kazuki Takashima
Research Institute of Electrical Communication, Tohoku University, Japan
https://www.icd.riec.tohoku.ac.jp/en/
We consider physical robotic entities such as self-actuating furniture and drones the dominant ones to configure adaptive and intelligent workspaces. In this talk, we focus on two projects, an interactive robotic partition system and an interactive drone manipulation system using the egocentric spatial user interface. Users can easily and intuitively arrange those moving devices from their egocentric (user-centric) perspective, where all instructions can be spatially described in their coordinate system. Based on these projects, we discuss the benefits and challenges of this approach.
Bringing Human-like Cognition to Social Robots: Implementing Cognitive Robots for Human-robot Interaction
Li-Chen Fu
Department of Computer Science and Information Engineering & Electrical Engineering, National Taiwan University, Taiwan
Director, NTU Center for AI and Advanced Robotics
http://robotlab.csie.ntu.edu.tw/
Cognitive Robotics is an emerging field that combines artificial intelligence, robotics, and cognitive science to create robots that can perceive, reason, and act in complex environments. In recent years, this field has gained significant attention as researchers strive to develop social mobile robots that can interact with humans in a natural and intuitive way. A key aspect of this involves bringing human-like cognition to robots, which includes the development of spatial cognition, decision-making ability, and navigational skills, as well as the understanding of human intention through speech. This presentation will address the above aspect and highlight it by a study that focuses on creating cognitive maps, retrieving human intent, and reasoning about destinations to provide more effective and interactive services to users. An important objective of this talk is to explore the potential of AI in enhancing the capabilities of social mobile robots and improving the overall human-robot experience.
Enhancing Well-Being in the Age of AI: How Psychology Can Help
Su-Ling Yeh
Department of Psychology, National Taiwan University, Taiwan
http://epa.psy.ntu.edu.tw/SuLingYeh_eng.html
AI is closely related to psychology because AI systems are designed to work like humans and because AI’s evolution parallels the history of psychology. AI systems are created to function like the human brain, and the concepts and vocabularies used in AI models (e.g., attention and memory) are similar to those of psychology because AI was initially intended to address issues that could only be handled by human intelligence. As AI continues to evolve and revolutionize our world, the link between AI and psychology becomes clearer. In this AI era, psychology is critical for preparing for rising population aging and the associated challenges to improve well-being, such as using Maslow's hierarchy of needs as a theoretical framework to assist older people by designing AI systems that better meet their various levels of needs. Even if AI can approximate humans in many ways, phenomenal consciousness is the last frontier that AI will never cross. Understanding the psychological needs of the aging population, as well as the uniqueness of human beings, is vital for developing AI capable of improving global well-being while maintaining human dignity.
Theatrical methods for Communication skills from the perspective of embodied cognitive neuroscience
Hajime Mushiake a & Miki Mushiake b
a Department of Physiology, School of Medicine, Tohoku University, Japan
b Faculty of Education, Miyagi University of Education, Japan
http://www.neurophysiology.med.tohoku.ac.jp/
https://researchmap.jp/mkmushiake
Communication skills are important for everyone but they are not easy to acquire. From the perspective of embodied cognitive neuroscience, communication is considered as a cooperative activity through mutual interaction under uncertainty. Improvisational activities in theatrical methods enhance this kind of cooperation among participants. Based on these ideas, we introduced improvisational activities in theatrical methods for liberal arts education at Tohoku University. I will also show our findings in current projects supported by Japan Science and Technology Agency (JST).
Session 3
Thinking About Research on Implementing Machine Consciousness
Kazunori Yamada
Unprecedented-scale Data Analytics Center, Tohoku University
https://yamada-kd.com/
In the current artificial intelligence (AI) boom, we have successfully developed AI that has partially surpassed human capabilities in certain areas. However, we have yet to develop "strong AI," which is AI that can perform a range of cognitive processes like humans. Machine consciousness, the consciousness built into devices, has been studied from two perspectives: studying machine consciousness as a tool to elucidate human consciousness and achieving the technological goal of furthering AI research with conscious AI. Our focus is on the second perspective. In this presentation, we will introduce the current state of research on implementing machine consciousness, based on our survey paper from 2022. The field of machine consciousness research is a complete frontier, and we do not even know where to begin. However, we will present our efforts to build machine consciousness using neural networks as a first step toward achieving larger research goals in this field.
How Versatile Are Self-supervised Foundation Models?
Hung-Yi Lee
Department of Electrical Engineering, National Taiwan University, Taiwan
http://speech.ee.ntu.edu.tw/~tlkagk/
Self-supervised learning (SSL) has shown to be vital for advancing research in natural language processing (NLP), computer vision (CV), and speech processing. The paradigm pre-trains a shared foundation model on large volumes of unlabeled data and achieves state-of-the-art for various tasks with minimal adaptation. This talk first introduces the Speech processing Universal PERformance Benchmark (SUPERB), which is a leaderboard to benchmark the performance of SSL model across a wide range of speech processing tasks. The results on SUPERB demonstrate that SSL representations show competitive generalizability across speech processing tasks. This talk will also share some surprising findings that SSL models pretrained from text are helpful for non-text token sequence classification data, including amino acid, DNA, and music.
Session 4
Crossmodal Correspondences: Methodology and Applications
Yi-Chuan Chen
Department of Medicine, MacKay Medical College, Taiwan
https://www.mmc.edu.tw/medicine-eng/Teacher_Detail.asp?hidPage1=4&hidPeopleCatID=1&hidPeopleID=130
The human brain searches for the correlations between incoming sensory signals in order to better reconstruct and understand the physical world. One classic example is illustrated by the Bouba/Kiki effect where the meaningless speech sounds (i.e., “Bouba” and “Kiki”) are mapped onto rounded and angular patterns, respectively. Previously researchers often used arbitrary visual patterns to test this phenomenon, and the correspondences were verified qualitatively if group consensus was reached. By systematically manipulating the visual features, we were able to quantitatively predict people’s matching judgments either at the individual or group level. The knowledge of crossmodal correspondences and the developed methodology can be applied to investigate people’s impression and acceptance conveyed by the appearance of robots.
Sensory Nudges and Eating Behavior: A multidisciplinary approach combining the questionnaire-based social science and cognitive neuroscience.
Nobuyuki Sakai
School of Arts and Letters, the Research Institute of Electrical Communication, and the Advanced Institute of Yotta Informatics, Tohoku University, Sendai, Japan.
https://www.sal.tohoku.ac.jp/en/research/researcher/profile/---id-71.html
We have Japanese traditional cuisine, called as WASHOKU, which is regarded as healthy foods from foreign people. However, we Japanese face health problems evoked by excessive salt intake. The average of daily salt-intake of Japanese people is over 10g/day for each person. Thus, we should cut it in half. This study aimed to reveal the reason for difficulties on salt reduction in Japanese people. We found that the Japanese people have a distorted knowledge about salt in Japanese foods, have a positive attitude to salt in cuisine, and do nothing for salt reduction in our daily lives. We also found that Japanese can detect the “low-salt” food with flavor perception, which cannot be distorted by the packages or other marketing tools. We should develop the new approach to make Japanese people reduce their salt intakes.
Using Implicit Association Test to Explore Crossmodal Correspondences
Pi-Chun Huang
Department of Psychology, National Cheng Kung University, Taiwan
https://sites.google.com/site/pchvisionlab/main
We adopted implicit association test (IAT) that measures the strength of associations between two pairs of items or concepts to investigate the crossmodal correspondences. We investigated how vowels and lexical tones drive sound–shape (rounded or angular) mappings among native Mandarin Chinese speakers. Our results demonstrated that both vowels and lexical tones play prominent roles in crossmodal correspondences but lexical tone did not modulate the strength of congruency effect in IAT, indicating different underlying mechanisms responsible for tones and vowels. We also adopted IAT to explore taste-shape (rounded or angular) mapping. Our results showed the taste of the sweet/bitter pair showed no taste-shape association but the sweet/sour pair did. IAT provides a useful paradigm to probe the crossmodal correspondences.
Session 5
Detecting important information from scene images via texts
Shinichiro Omachi
Graduate School of Engineering, Tohoku University, Japan
http://www.iic.ecei.tohoku.ac.jp/~machi/index-e.html
When transmitting images and videos over a narrow-band network, or getting their overview, detecting important information can be an effective approach. Visual saliency has often been used for detecting salient objects, which can be regarded as important objects. However, this approach only considers visual appearances to judge the importance. In this talk, our attempts to detect important information from scene images using text information are presented. One is a method of detecting important information by detecting important texts in the scene images. Another one is to detect important by generating texts that represent the scene images.
Recent Advances in Vision & Language
Yu-Chiang Wang
Department of Electrical Engineering, National Taiwan University, Taiwan
http://vllab.ee.ntu.edu.tw/ycwang.html
Traditionally AI systems are unimodal. For example, given a single-modality input such as an image or speech, one aims to analyze the corresponding information such as semantics or identity. In real-world applications, one expects to observe vital contextual information from cross-modality data. Thus, multimodal AI is developed to combine data such as image, text, speech, and audio for achieving further improved performances. In this talk, we share our recent works related to visual and language models. The first part of the talk will focus on visual captioning, particularly the task of Novel Object Captioning. As for the second part of the talk, we will share our recent development of latent diffusion model for image synthesis. These works are recently published at NeurIPS 2022 and AAAI 2023.
Session 6
Smart Healthcare: How can AI Revolutionize the Healthcare Ecosystem
Che Lin
Senior Member, IEEE
Department of Electrical Engineering & Graduate Institute of Communication Engineering & Center for Computational and Systems Biology & Center for Biotechnology
Director, TIGP Program on AIoT
Director, Computer and Information Networking Center of Electrical Engineering
Vice Director, Smart Medicine and Health Informatics Program
National Taiwan University (NTU)
https://www.idssp.ee.ntu.edu.tw/che-lin
Artificial intelligence (AI) can be applied to a wide range of areas. Notably, the fast-growing Smart Healthcare area has created tremendous opportunities recently. It also creates an ideal environment for AI-Biomedical interdisciplinary specialists to make considerable contributions and significantly impact the healthcare ecosystem. Although there are millions of promising opportunities to be explored in this field, we simply have too many different modalities of data of very diverse natures describing a person's health state. It is hence important to consider the fundamental data issues for different medical data sources. In this talk, I would like to provide my perspectives on these fundamental data issues and how we can design cutting-edge deep learning algorithms to tackle them. I will introduce how we can utilize state-of-the-art AI technologies for many smart healthcare applications, such as accurate disease risk prediction, essential predictors selection, and even drug discovery. I will use several real case studies from my lab to illustrate how researchers from different disciplines can work together and achieve tremendous advancement to revolutionize the healthcare ecosystem via AI.
Recommendation systems with network structure and big data
Tsukasa Ishigaki
Graduate School of Economics and Management, Tohoku University
http://www2.econ.tohoku.ac.jp/~isgk/
Recommendation system supports to provide some valued items for each user from huge number of alternatives and is used in a variety of industries, including retail, restaurant, news, trip, music, movie and media. In this talk, I will talk about some novel recommendation methods using network structures and big data. First, I will provide
some introductory topics about motivations, properties of data, problem settings and basic methods on recommendation system. Then, some topics that we have proposed will be discussed. The methods have been realized using deep learning with implicit feedback or knowledge graph. The results show that the proposed methods have a high performance in terms of accuracy or novel recommendation in some experiments using big data.
