International Workshop on Interpretable AI September 03-06, 2023

This workshop is held by JAIST's Interpretable AI Center, at Dalat University, on September 03-06, 2023.

happy team image

Workshop Information

This workshop is held by JAIST's Interpretable AI Center, at Dalat University (Navigation on Google Maps), on September 03-06, 2023.

Webex Information For Online Participants (Please feel free to join)

Meeting Link: https://jaist.webex.com/jaist/j.php?MTID=mda02d9a1d2e0649fd724ef66b1c7d8f8

Meeting number: 2512 219 2301
Password: cNfqPJCY332

Workshop Program

September 3, 2023

Meeting with Da Lat University

September 4, 2023

10:00-10:10: Opening section

10:10-11:00

Invited talk: Prof. Tomoko Matsui

Testing closeness of sequential data - analysis of COVID-19 evolution

Abstract: A practical algorithm for the analysis of sequential data closeness is introduced, based on the Markov chain tester. This algorithm has been applied to reported sequential data for COVID-19, enabling analysis of the disease's evolution during specific time periods (e.g., weekly, monthly).

Bio: TOMOKO MATSUI (Senior Member, IEEE) received the Ph.D. degree in computer science from the Tokyo Institute of Technology, Tokyo, Japan, in 1997. From 1988 to 2002, she was a Researcher in several NTT laboratories, focusing on speaker and speech recognition. From 1998 to 2002, she was a Senior Researcher with the Spoken Language Translation Research Laboratory, ATR, Kyoto, focusing on speech recognition. In 2001, she was an Invited Researcher with the Acoustic and Speech Research Department, Bell Laboratories, Murray Hill, NJ, USA, working on identifying effective confidence measures for verifying speech recognition results. She is currently a Professor with The Institute of Statistical Mathematics, Tokyo, working on statistical spatial–temporal modeling for various applications, including speech and image recognition. She received the Best Paper Award from the Institute of Electronics, Information, and Communication Engineers of Japan, in 1993.

11:00-11:10: Break

11:10-11:40

Presentation by Dr. Tung Le

Combining Multi-vision Embedding in Contextual Attention for Vietnamese Visual Question Answering

Abstract: Visual question answering (VQA) represents a complex task involving the integration of vision and language and has garnered significant attention from researchers in recent times. Although a majority of VQA models have been developed and optimized for English datasets, only a limited number of studies have focused on addressing the task within the Vietnamese language domain. This research gap has inspired our interest in further exploring and applying the latest advancements in VQA models to the Vietnamese context. Currently employed methods primarily rely on the extraction of regional features of objects, effectively capturing the local contextual information present in images. However, this approach can result in the loss of global context. In contrast, successful VQA systems rely on the extraction and assimilation of both global and local information in images. In this study, we introduce the Multimodal Contextual Attention method, which enables the adaptive learning of attentional embeddings from different contexts derived from both the image and the accompanying question. This mutual guidance between the image and question facilitates the exploration of downstream systems capable of harnessing the capabilities of pre-trained models for Vietnamese VQA tasks.

Bio: Tung Le is a Lecturer of Information Technology at the University of Science, VNU-HCM. He completed his B.S. degree in Computer Science in 2012 as part of the Honor Program at the University of Science, VNU-HCM. Subsequently, he pursued his M.S. degree in Information Science in 2018 at Japan Advanced Institute of Science and Technology (JAIST). In 2021, he obtained his Ph.D. degree in Information Science from the School of Information Science at JAIST. Tung Le's research interests encompass various domains, including Information Retrieval, Deep Learning, Natural Language Processing, Visual Question Answering, and Multi-modal Systems.

11:40-12:10

Presentation by Dr. Huy Nguyen

Explainable multimodal model for Youtube user engagement analysis

Abstract: As popularity of video-sharing platforms, content creators have a high demand to produce content which attracts the large amount of viewers. There are many factors related to engagement: visual, sound, transcript, title, etc. To take into account of these factors, we propose an explainable multimodal model for YouTube video engagement. The architecture allows us to be easy to adapt state-of-the-art models for a particular task or variety of modalities, then fuse them to obtain more information aim to classify better. In addition, XAI techniques are employed to show what the model learns from data and to explain how these factors affect to video engagement. This will help users understand more about the important factors in attracting the attention and engagement of audiences on YouTube.

Bio: Huy Tien Nguyen is currently a lecturer of Computer Science at the University of Science, Ho Chi Minh City, Vietnam (HCMUS). He received the B.S degree in Software Technology (2010) and the M.S degree in Computer Science (2015) from HCMUS. He received his Ph.D. degree in Information Science from School of Information Science, Japan Advanced Institute of Science and Technology (JAIST) in 2019. His current research interests include information retrieval, multimodal learning, natural language processing.

12:10-14:00: Lunch

14:00-14:50

Invited talk by Prof. Satoshi Tojo

From Quantum Computing To Machine Learning

Abstract: We have tackled a quantum representation of knowledge, to clarify that (i) knowledge is a superposition of multiple meanings, (ii) the recipient of information must selectively observe the received message, and (iii) the retrieved information depends on its inherent probability. The semantics of quantum logic employs Hilbert space, in which an element is a vector, and between them an inner product is defined. This implies that the semantics possesses quite high affinity with machine learning, where every unit of information is a vector, and there exists a notion of cosine similarity. In this talk, we propose a new direction of future information science, to combine quantum computing and probabilistic behavior, considering other physical analogy such as cross entropy and KL-divergence.

Bio: Satoshi Tojo received B.E., M.E. and Ph.D. from the University of Tokyo. After he worked at Mitsubishi Research Institute, Inc. in 1983-1995, he joined Japan Advanced Institute of Science and Technology (JAIST) as an associate professor of School of Information Science from 1995, and has been a full professor from 2000 to 2023. From 2023, He currently is a professor of the department of data science, Asia University. His interests include formal semantics of natural language, logic of knowledge and belief in agent communication. His recent works also concern laguage evolution, as well as linguistic analysis of music.

14:50-15:20

Presentation by Dr. Vu Tran

Towards Enhancing Information Extraction via Public Discussions on Reddit about COVID-19 Research

Abstract: In the scene where an enormous number of COVID-19-related research papers have been published, we study public discussions and sentiments on Reddit, a popular social media platform, about COVID-19 research by analyzing the related Reddit posts. From a preliminary analysis of Reddit, it could be helpful to use social media to enhance information extraction, which, in turn, can help improve the interpretation of research results and the conveying of research messages.

Bio: Dr. Vu Tran (The Institute of Statistical Mathematics) Project Assistant Professor, The Institute of Statistical Mathematics, Tokyo, Japan His major is machine learning, particularly deep learning and natural language processing. He received his Ph.D. degree in information science from Japan Advanced Institute of Science and Technology (JAIST) in 2019, with his work on deep learning methodologies for the legal domain. After that, he continued to work at JAIST on the same topic. In 2021, he moved to the Institute of Statistical Mathematics in his current position. He has been working on utilizing deep learning and other statistical methods with social media data for tackling COVID-19-related problems, for example, predicting the number of COVID-19 cases.

15:20-15:50

Invited Talk by Prof. Bui Thu Lam

Artificial Intelligence and Applications in Cyber Security

Abstract: Cyber security is a fascinating field of research with a wide range of problems involved with data analysis and prediction. That is why the applications of artificial intelligence in this field has been a hot topic. In this talk, I will cover the recent development in using AI for cyber security applications, such as deep learning and adversarial machine learning. Furthermore, the relationship between machine learning and cryptography will also be covered in this talk and how to protect machine learning models with detailed examples.

Bio: Dr. Lam Thu BUI received the Ph.D. degree in computer science from the University of New South Wales (UNSW), Australia, in 2007. He did postdoctoral training at UNSW from 2007 until 2009. He has been involved with academics including teaching and research since 1998. Currently, he is an Associate Professor in Computer Science, Academy of Cryptography Techniques (ACT), Hanoi, Vietnam. He is doing research in the field of AI, specialized with natural computation including neural computation and evolutionary multiobjective optimization. He is the co-editor of the book Multiobjective Optimization in Computational Intelligence: Theory and Practice (IGI Global Information Science Reference Series); and the General Chair of the Ninth International Conference on Simulated Evolution and Learning – SEAL2012. Dr. Bui was EiC of Journal of Science and Technology: Section on Information and Communication Technology (LQDTU-JICT), a member of the Editorial Board, International Journal of Computational Intelligence and Applications (IJCIA), and was the Vice-Chair of the Evolutionary Computation Technical Committee (ECTC), IEEE Computational Intelligence Society. He has been a member of the program committees of several conferences and workshops Artificial Intelligence.

15:50-16:20

Invited talk by Senior Lecturer RACHARAK Teeradaj

Interpretable Decision Tree Ensemble Learning with Abstract Argumentation for Binary Classification

Abstract: We marry two powerful ideas: decision tree ensemble for rule induction and abstract argumentation for aggregating inferences from diverse decision trees to produce better predictive performance and intrinsically interpretable than state-of-the-art ensemble models. Our approach, called Arguing Tree Ensemble, is a self-explainable model that first learns a group of decision trees from a given dataset. It then treats all decision trees as knowledgeable agents and lets them argue with each other for concluding a prediction. Unlike conventional ensemble methods, this proposal offers full transparency to the prediction process. Therefore, AI users are able to interpret and diagnose the prediction’s output.

Bio: Teeradaj Racharak received his M.Eng. in Computer Science (with specialization in Software Engineering) from Asian Institute of Technology, Ph.D. in Information Science from JAIST (majoring Description Logic), and Ph.D. in Engineering and Technology from Thammasat University (majoring Formal Argumentation). He has joined JAIST as an Assistant Professor since April 2019 and becomes a Senior Lecturer in October 2022. Before JAIST, he was a software and DevOps engineer, involved in the development of large-scale web applications including MangaMagazine.net and Inkblazers. Prior to that, he was a researcher at Institute for Information Technology Innovation, Kasetsart University. He is broadly interested in mathematical modeling and implementation of AI systems. His research interests span across: Knowledge Representation and Reasoning (KRR), Machine Learning (ML), and Integration of KRR and ML for Trustworthy AI, as well as AI applications for social good.

16:20-16:30: Break

16:30-17:20

Invited talk by Prof. Yuji Matsumoto

How to live with Large Language Models

Abstract: Recent rapid progress of Large Language Models (LLMs) is remarkable. They are said to have emergent abilities, which enable them to perform various natural language related tasks with very high performance. Such high performance is obtained without fine-tuning but with in-context learning with few-shot or even with zero-shot settings. This talk roughly surveys recent progress of LLMs, issues and weaknesses of LLMs, and discusses how to live with LLMs, especially of the topic related with sophisticated prompt engineering such as Chain-of-Thought and its extensions.

Bio: Yuji Matsumoto received his M.S. and Ph.D. degrees in information science from Kyoto University in 1979 and in 1990, respectively. He joined the Machine Inference Section of Electrotechnical Laboratory in 1979. He has been an academic visitor at Imperial College of Science and Technology, the deputy chief of First Laboratory at ICOT, an associate professor at Kyoto University, and a professor at Nara Institute of Science and Technology. He is currently a team leader at RIKEN Center for Advanced Intelligence Project. His main research interests are natural language understanding and machine learning. He is an ACL fellow and a fellow of the Information Processing Society of Japan.

September 5, 2023 (Morning)

9:30-10:10

Invited talk by Prof. Shinobu Hasegawa

Generative AIs in EduTech

Abstract: In this talk, we will discuss the recent practical aspects of generative AIs in the field of educational technology (EduTech), especially in terms of its application (pros and cons) for education and research for univiersity / graduate students.

Bio: Shinobu Hasegawa is currently a professor at Center for Innovative Distance Education and Research, JAIST. He received his B.S., M.S., and Ph.D. degrees in systems science from Osaka University in 1998, 2000, and 2002, respectively. The primary goal of his research is to facilitate “Human Learning and Computer mediated Interaction” in a distributed environment. His research fieldis mainly learning technology, EduTech, and AI in education, which include support for Web-based learning, game-based learning, cognitive skill learning, affective learning, distance learning system, and community based learning.

10:10-10:50

Presentation by Prof. Naoya Inoue (JAIST)

Towards Complex Reasoning with Large Language Models

Abstract: Large Language Models are now essential core technologies in both Natural Language Processing and Vision-Language tasks. However, our recent studies, in line with other related studies, suggest that these models still fall short in complex reasoning tasks such as textual multi-step reasoning and visual logical reasoning. In this talk, I will give an overview of our ongoing efforts in analyzing the reasoning ability of LLMs and multi-modal LLMs. I then discuss our approach to enhance these models towards more robust, accurate, and trustworthy complex reasoning.

Bio: Naoya Inoue received his MS degree in engineering from Nara Institute of Science and Technology in 2010 and his Ph.D. in Information Science from Tohoku University in 2013. He joined DENSO Corporation as a researcher in 2013. He was an assistant professor at Tohoku University (2015-2020) and a postdoctoral associate at Stony Brook University (2020-2022). Since 2022, he has been an associate professor at the Japan Advanced Institute of Science and Technology. He has also been a visiting researcher at RIKEN Center for Advanced Intelligence Project since 2018. His research interests include reasoning, explainability, and argumentation.

10:50-11:30

Invited talk by Prof. Mizuhito Ogawa

Formal semantics extraction from manuals of instruction sets

Abstract: Malware includes many obfuscation techniques and the most powerful tool for de-obfuscation is believed to be a symbolic Execution. FOr developing such tools, we first need to clarify the formal semantics of binary instructions. However, the number of such instructions are more than thousands. For instance, x86 (ech for 32 bits and 64 bits) has about 1500-2000 instructions, and ARM has 30-40 chip set variations in which each has 100-200 instructions. Hence, if everything is done manually, huge human effort is required. Instead, we aim to automatically extract from their manuals. We will overview how Natural Language Processing (NLP) techniques are applied for this purpose by combining software enginerring methods.

Bio: Mizuhito Ogawa graduated the master course of University of Tokyo, majoring Mathematics. He worked NTT laboratories on functional programs, dataflow machine, and dataflow analysis till 2001. Then, he was JST PRESTO fellow until 2003, and stayed at University of Tokyo as a visiting researcher. Since 2003 till now, he has been in JAIST as a professor in Schools of Information Science. Since 2020, he also joined Interpretable AI center in JAIST. His research interest covers from theory to practical tool implementation. For instance, the confluence of rewriting systems, the computational content of classical proofs, and the decidability of the reachability of infinite state transition systems are such examples in theory. Tool implementations are mostly performed under collaboration with Vietnamese universities and students. Such examples are, (1) SMT solver, raSAT, on polynomial constraints over reals, which was ranked second in QFNRA category of SMT-COMP 2016 and 2017, (2) dynamic symbolic execution for binary code, e.g., x86-32, ARM Cortex-A/M (32 bits), MIPS (32bits), targeting on malware. The tool implementation are BE-PUM (x86), Corana (ARM), and SyMIPS (MIPS). Currently, 64bits extensions are under collaboration with LQDTU, especially based on formal semantics extraction from instruction manuals in English.

11:30-11:40: Closing section

September 5, 2023 (Afternoon)

Special Section on AI and Generative AI

14:00-14:30: Introduction to Large Language Models (by Dr. Vu Tran)

14:30-15:00

Presentation by Dr. Hung Nguyen (IoIT Vietnam)

Deep Reinforcement Learning for NLP

Bio: Hung The Nguyen completed his PhD in Computer Science from the University of New South Wales, Canberra, Australia in 2023. He received his B.E degree in Information Technology from Military Technical Academy, Vietnam in 2010 and his MSc in Computer Science from University of New South Wales, Canberra, Australia in 2018. He is a reviewer for over ten journals and conferences. He has published more than fifteen works in the area of artificial intelligence. His current research focuses on apprenticeship bootstrapping, deep learning, imitation learning, reinforcement learning, inverse reinforcement learning, autonomous vehicles, natural language processing and human-machine teaming. He is working on appealing projects including trusted artificial intelligence, generative chatbots, cyber security, and face recognition.

15:00-15:30: Hallucination Problem in Large Language Models (by Prof. Minh Nguyen)

15:30-17:00: The application of LLM (chaired by Prof. Minh Nguyen)

September 6, 2023

9:00-10:00: Panel discussion on LLM and its application (chaired by Prof. Minh Nguyen)

Workshop Chairs

Prof. Nguyen Le Minh - Japan Advanced Institute of Science and Technology

JAIST NguyenLab

© 2023 NguyenLab