SinFra Symposium 2019 on Artificial Intelligence

SinFra Symposium 2019 on Artificial Intelligence
12 & 13 December

@ InFuse, Connexis South, Level 14 & @NUS (iCube, Seminar Room1)

In line with the successful Singaporean-French Symposiums, SinFra’2019 is the 4th edition and will take place in Singapore on 12-13 December 2019. The symposium proposes outstanding talks related to the shared interest in the research field of Artificial Intelligence. SinFra welcomes researchers, industrial and collaborators under IPAL International Research Lab and those who wish to collaborate with French and Singaporean institutions.

Program SinFra 2019 (PDF: Click here)

Dec 12th (Thursday) @ InFuse, Connexis South, Level 14
1 Fusionopolis Way
Connexis Tower
Singapore 138632

0830 - 0900 Registration

0900 - 0915 Opening and Welcome Remarks by Professor Lye Kin Mun, Executive Director, Astar-I2R
0915 - 0930 Introduction to School of Computing and NUS by Mohan Kankanhalli, Dean, School of Computing, National University of Singapore
0930 - 0950 Introduction to IPAL and SinFra 2019 by Professor Christophe Jouffrais, Director of IPAL.

0950 - 1030 SESSION 1
0950 - 1010 Trust in AI Blaise Genest (CNRS, IRISA)
1010 - 1030 Trustworthy and Accountable Machine Learning: Privacy, Robustness, and Interpretability Challenges Reza Shokri (NUS, SoC)

1030 - 1100 Coffee Break

1100 - 1220 SESSION 2
1100 - 1120 Positivity Certificates for Analyzing Robustness in DNN Jean-Bernard Lasserre (CNRS, ANITI & LAAS)
1120 - 1140 Exploiting Second Order Sparsity in Big Data Optimization Kim Chuan Toh (NUS, Mathematics)
1140 - 1200 Bio-inspired Neural Networks for Robust and Understandable Neuromorphic Systems  Martial Mermillod (UGA, MIAI & LPN)
1200 - 1220 Event-based Visual Sensor for Robotic Navigation  Fabien Colonnier (A*star, I2R)

1220 - 1350 Lunch
1350 - 1450 Demos and Tour of I2R (Meet at InFuse)

1450 - 1610 SESSION 3
1450 - 1510 Introduction to MIAI Eric Gaussier, Director of MIAI

1510 - 1530 Towards Real-time Lifelong Learning Savitha Ramasamy (A*STAR - I2R)

1530 - 1550 Deep Learning Research Inspired by Biomedical Applications Hwee Kuan Lee (A*STAR - BII)

1550 – 1610 Improving Energy-efficiency for Security and AI Techniques for the IoT Compute Hierarchy Trevor Carlson (NUS, SoC)

1610 - 1640 Coffee Break

1640 - 1720 SESSION 4

1640 - 1700 Smart Buildings and Districts Benoit Delinchant (UGA, MIAI & G2E) 

1700 - 1720 The Role of AI in the Next Interaction Paradigm Shengdong Zhao (NUS, SoC)

1720 - 1830 Happy Hour / Conclude

1830 Dinner (for invited speakers)
Hosted by Professor Lye Kin Mun, Executive Director, A*Star-I2R
Yunnan Garden Restaurant
1 Fusionopolis Place, #02-02 Galaxis, Singapore 138522
Tel: (+65) 6665 8888, Fax: (+65) 62522440


Dec 13th (Friday) @NUS (iCube, Seminar Room 1)
i3 Building (I-CUBE)
Seminar Room 1 @ Level 1
21 Heng Mui Keng Terrace
Singapore 119613

0915 - 0930 Welcome by the Ambassador of France, HE Marc Abensour
0930 - 0950 Introduction to ANITI Professor Nicholas Asher, Director of ANITI
0950 - 1010 Introduction to AI.SG Ma Su Su (Head, Research Management, AI.SG)

1010 - 1050 SESSION 5
1010 - 1030 Explainability for Learning Systems using Logic Nicolas Asher (CNRS, ANITI & IRIT)
1030 - 1050  Conversational AI: Data Augmentation, Dialogue Comprehension, and Summarization Generation Nancy F. Chen (A*star, I2R)

1050 - 1120 Coffee Break

1120 - 1140 SESSION 6
1120 - 1140 Instance and Gesture Identification with Small Sized CNN for Information Access in Mobility Jean-Pierre Chevallet (UGA, MIAI & LIG)
1140 - 1200 Semantic and Sentiment Analysis for Knowledge Graph Construction Su Jian  (A*star, I2R)
1200 - 1220 Natural Language Processing for e-Health Pierre Zweigenbaum (CNRS, Limsi)
1220 - 1240 Language Translation: Technologies, Challenges and Applications Aw Ai Ti (A*star, I2R)

1240 - 1400 Lunch
1400 - 1500 SoC Tour (hosted by David Hsu, NUS, SoC) COM1 Robotics Living Lab

1500 - 1600 SESSION 7
1500 - 1520 TheoremKB: Towards a Knowledge Base of Mathematical Results Pierre Senellart (ENS, DI ENS)
1520 - 1540 Nonnegative Matrix Factorisation for Data Processing Cédric Févotte ( CNRS, ANITI & IRIT)
1540 - 1600 A Ranking Model Motivated by NMF with Applications to Tennis Analytics (and other research activities in my group) Vincent Tan (NUS, Mathematics/ECE)

1600 - 1630 Coffee Break

1630 - 1710 SESSION 8
1630 - 1650 Transparency in AI and Ranked Retrieval Philippe Mulhem (CNRs, MIAI & LIG)
1650 - 1710 Detecting Fake Videos Terence Sim (NUS, SoC)
1710 - 1720 Closing Remarks
1720 - 1800 Mingling / Free Discussion Session
1800 - 1830 Transport to the Residence of the Ambassador of France
1830 Reception at the Residence of the Ambassador of France
1800 - 1830 Transport to the Residence of the Ambassador of France
1830 Reception at the Residence of the Ambassador of France

Talks information


12 December, 0950-1030, InFuse @ Fusionopolis

Trust in AI

Blaise Genest (CNRS, IRISA)


In this talk, we will describe several ways to foster trust in AI tools. Indeed, the most prominent current AI tools (Neural Networks, Reinforcement learning) are in the same time very efficient, but also very complex to be understood by humans. Recently, the new paradigm of explainable AI has emerged. We describe that while explainable AI is necessary, it is also not sufficient for complete trust. In order to complement explainable AI, we describe recent advances in Formal Methods and AI that provide guarantees about their behaviors.


Blaise Genest is a CNRS senior researcher with IRISA, Rennes. He spent 3 years in IPAL in 2010-2012. He is working at the interface of Formal Methods and AI. He organized the 2sd International Workshop on FM and AI in Rennes in 2019, which attracted around 50 researchers from 11 countries. He was a member of the Committee for French National Research, 2009-2012, in charge of evaluating researchers, labs, and hiring permanent researchers for CNRS.

Trustworthy and Accountable Machine Learning: Privacy, Robustness, and Interpretability Challenges

Reza Shokri (NUS, School of Computing)


Machine learning algorithms have shown an unprecedented predictive power for many complex learning tasks. As they are increasingly being deployed in large scale critical applications for processing various types of data, new questions related to their trustworthiness would arise. Can machine learning algorithms be trusted to have access to individuals' sensitive data? Can they be robust against noisy or adversarially perturbed data? Can we reliably interpret their learning process, and explain their predictions? In this talk, I will go over the challenges of building trustworthy and accountable machine learning algorithms in centralized and collaborative (federated) settings, and will discuss the inter-relation between privacy, robustness, and interpretability, and whether they are preserved in a fair manner.


Reza Shokri is a NUS Presidential Young Professor of Computer Science. His research focuses on trustworthy machine learning, quantitative analysis of data privacy, and design of privacy-preserving algorithms for practical applications, ranging from data synthesis to collaborative machine learning. He is an active member of the security and privacy community, and has served as a PC member of IEEE S&P, ACM CCS, Usenix Security, NDSS, and PETS. He received the Caspar Bowden Award for Outstanding Research in Privacy Enhancing Technologies in 2018, for his work on analyzing the privacy risks of machine learning models, and was a runner-up in 2012, for his work on quantifying location privacy. He also received the NUS Early Career Research Award 2019 for working on trustworthy machine learning. He obtained his PhD from EPFL.


12 December, 1100-1220, InFuse @ Fusionopolis

Positivity Certificates for Analyzing Robustness in DNN

Jean Bernard Lasserre (CNRS, ANITI)


We briefly describe how semi-algebraic techniques can help analyze robustness of DNN. We also introduce the Christoffel function for some data analysis in ML.


JB Lasserre is a senior researcher éméritus at CNRS. He is interested in global optimization in a broad sense. He has published several books in Markov control processes, Markov chains, semi-algebraic optimization, moment-sum-of-squares.

Exploiting Second Order Sparsity in Big Data Optimization

Kim Chuan Toh (NUS, Mathematics)


In this talk, we shall demonstrate how second order sparsity (SOS) in important optimization problems such as sparse optimization models in machine learning, semidefinite programming, and many others can be exploited to design highly efficient algorithms.  

The SOS property appears naturally when one applies a semismooth Newton (SSN) method to solve the subproblems in an augmented Lagrangian method (ALM) designed for certain classes of structured convex optimization problems. With in-depth analysis of the underlying generalized Jacobians and sophisticated numerical implementation, one can solve the subproblems at surprisingly low costs. For lasso problems with sparse solutions, the cost of solving a single ALM subproblem by our second order method is comparable or even lower than that in a single iteration of many first order methods.Consequently, with the fast convergence of the SSN based ALM, we are able to solve many challenging large scale convex optimization problems in big data applications efficiently and robustly. For the purpose of illustration, we present a highly efficient software called SuiteLasso for solving various well-known Lasso-type problems.This talk is based on joint work with Xudong Li (Fudan U.) and Defeng Sun (PolyU).


Dr Toh is the Leo Tan Professor in the Department of Mathematics at the National University of Singapore (NUS). He works extensively on convex programming, particularly large-scale matrix optimization problems such as semidefinite programming, and structured convex problems arising from machine learning and statistics. He is currently an Area Editor for Mathematical Programming Computation, an Associate Editor for SIAM Journal on Optimization, Mathematical Programming, and ACM Transactions on Mathematical Software. He received the 2017 Farkas Prize awarded by the INFORMS Optimization Society and the 2018 triennial Beale-Orchard Hays Prize awarded by the Mathematical Optimization Society. He is also a Fellow of the Society for Industrial and Applied Mathematics.

Bio-inspired Neural Networks for Robust and Understandable Neuromorphic Systems

Martial Mermillod (UGA, MIAI & LPN)


In this talk, I will present the bio-inspired origins of convolutional neural networks (CNN), what part of the brain CNN is the formal analog but also some important differences with the human visual system. Then, by means of bio-inspiration from the human brain and cognitive sciences, I will present how we can improve the reliability of neural networks with regards to (1) anticipation and classification of visual events (2) incremental learning and (3) resilience to adversarial attacks.


Martial Mermillod (Full Professor, IUF, Univ. Grenoble Alpes) is a scientific researcher in the fields of neural computation, psychological sciences, cognitive neurosciences applied to the field of visual perception and visual cognition. He has published more than 80 articles in highly ranked journals (including Neural Networks, NeuroComputing, Connection Science, Cognition, Psychological Science, Behavioural and Brain Science, Nature Scientific Reports, etc.) Details are provided here: He has successfully participated in 12 multicentre research projects (LABEX, ANR, PHRC, etc.)

Event-based Visual Sensor for Robotic Navigation

Fabien Colonnier (A*STAR, I2R)


Event-based vision sensors show great promise for use in embedded applications requiring low-latency passive sensing with high power efficiency. In order to leverage the full potential of such sensor, new algorithms should be designed. A first example will be the estimation of Time-To-Contact (TTC) retrieved from event-based optic flow computation. The TTC estimation enables a fast flying quadrotor (2.5m/s) to perform obstacle avoidance. A second example will be a 6 degree of freedom pose estimation displayed with handheld and flight experiments. The pixel association with an event-by-event Extended Kalman Filter allows a real-time efficient computation.


Fabien received a MSc in automotive engineering with a specialty on embedded systems from the ESTACA engineering school in 2012 and a PhD degree in bio-robotics from the Aix-Marseille University in 2017, both in France. His research is focused on visual perception for robotic applications. He investigates bio-inspired solutions to solve localization and navigation problematics using artificial compound eyes, event-based sensors and neuromorphic computation.


12 December, 1450-1610, InFuse@Fisionopolis

Introduction to MIAI

Eric Gaussier (UGA, Scientific Director of MIAI)


MIAI @ Grenoble Alpes: MIAI, the Grenoble Multidisciplinary Institute in Artificial Intelligence, develops the new generation of AI models and systems, from hardware architectures to software systems, with a focus on three application domains related to human beings and the environment: Health, Environment & Energy, and Industry 4.0. In addition, MIAI investigates the ethical and legal issues raised by the use of AI in almost all aspects of our lives, and studies how AI is perceived by French citizens. More than 500 academic and industrial partners, from more than 60 companies, interact on these aspects to support innovation in large companies and SMEs. A dedicated startup program also aims at creating, every year, several startups. MIAI also offers attractive courses for students of all levels and professionals, who can, according to the courses followed, be granted a “core AI” or an “AI application” certificate. Lastly, MIAI develops several programs to inform and interact with citizens on all aspects of AI.


Eric Gaussier is a Professor in Computer Science at University Grenoble Alpes. He is, since 2015, director of the Computer Science Lab in Grenoble and since June 2019 director of MIAI, one of the four interdisciplinary institutes in AI recently created by the French government. His research work lies at the intersection of Artificial Intelligence (encompassing machine learning and computational linguistics) and what would now be qualified as Data Science. I am interested in the general problem of accessing, mining and learning from large (text) collections, through machine learning models and methods and work on both fundamental problems (through the development of new models that explain different characteristics of large-scale collections/networks) and applications related to computational linguistics and information retrieval.

Towards Real-time Lifelong Learning 

Savitha Ramasamy (A*STAR - I2R)


Human beings have the ability to learn dynamically from the environment, and adapt themselves with non-stationarity. However, the state of the art AI systems lack this adaptability, without catastrophic forgetting. Hence, there is a need to develop algorithms that are capable of adapting to the non-stationarity of data,. This can include learning with dynamic network architectures without catastrophic forgetting, in the context of supervised classification and unsupervised representations. This talk will introduce some of our efforts along this direction.


Savitha Ramasamy is a Principal Investigator and Lab Head at the Institute for Infocomm Research (I2R), A*STAR, Singapore, leading projects for AME and healthcare applications. Prior to that, she obtained her PhD from the Nanyang Technological University in 2011. Her PhD thesis was nominated for the IEEE CIS best thesis award in 2012. Between Oct 2010 and Mar 2013, she was a Postdoctoral research fellow at NTU, Singapore. She has more than 60 papers published in several international conferences and journals.

Deep Learning Research Inspired by Biomedical Applications

Hwee Kuan Lee (A*STAR - BII)


Deep learning has achieved superb performance over many benchmark datasets. indeed deep learning can be useful in real life scenarios. however, when solving clinical problems, one starts to realise that many mainstream deep learning paradigms no longer fit into very specialized demands of clinical applications. in this presentation, we developed some non-mainstream deep learning basic research that are all inspired by the need to solve problems faced by clinicians in their everyday work.


Hwee Kuan is a Senior Principal Investigator of the Imaging Informatics division in Bioinformatics Institute. His current research work involves developing computer vision algorithms for clinical and biological studies. Hwee Kuan obtained his Ph.D. in 2001 in Theoretical Physics from Carnegie Mellon University with a thesis on liquid-liquid phase transitions and quasicrystals. He then held a joint postdoctoral position with Oak Ridge National Laboratory (USA) and University of Georgia where he worked on developing advanced Monte Carlo methods and nano-magnetism. In 2003, with an award from the Japan Society for Promotion of Science, Hwee Kuan moved to Tokyo Metropolitan University where he developed solutions to extremely long time scaled problems and a reweighting method for nonequilibrium systems. In 2005 he returned home to join Data Storage Institute, investigating novel recording methods such as hard disk recording via magnetic resonance. In 2006, he joined Bioinformatics Institute as a Principle Investigator in the Imaging Informatics Division.

Improving Energy-efficiency for Security and AI Techniques for the IoT Compute Hierarchy

Trevor Carlson (NUS, School of Computing)


Processor efficiency continues to be a main concern for processor architects across domains, from smart IoT systems to high-end datacenter servers. But, recent demands have compelled designers for more than area/power efficiency and high-performance. New demands for hardware security and always-on machine learning have created additional challenges for architects.

In this talk, we discuss these two research areas, and their implications for future processing systems. In hardware security, many challenges remain to protect future systems from intrusion, monitoring and hardware trojans, while doing so at an extremely low cost. With the future of IoT pointing to city-wide monitoring and information systems, physically securing this future infrastructure becomes a significant challenge. While naive methods to protect hardware is possible, they are far too expensive for the low-cost IoT devices they intend to protect. Our work aims to move to a higher level, above the circuit-design level, to the architecture level, to allow us to physically protect systems without the expensive overheads currently seen.

For the future of accelerators and always-on machine learning, many designers have been producing innovative solutions to apply deep neural networks in a fast and energy efficient manner. What has been overlooked recently, is the ability to transform modern networks, while still maintaining accuracy, into lighter-weight solutions that change the formula with respect to processing demands currently seen. Future designs, such as Binary and Ternary Neural Networks (BNN/TNN) and Neuromorphic computing (or Spiking Neural Networks (SNN) in this case) show promise to reduce latency while improving overall efficiency, allowing us to target future highly efficient system design goals.


Trevor E. Carlson is currently an assistant professor at the National University of Singapore (NUS), where he has worked since July, 2017. He received his B.S. and M.S. degrees from Carnegie Mellon University in 2002 and 2003, his Ph.D. from Ghent University in 2014, and has worked for 3 years as a postdoctoral researcher at Uppsala University in Sweden until 2017. He has also spent a number of years working in industry, at IBM in Poughkeepsie, NY from 2003 to 2007, at the imec research lab in Leuven, Belgium, from 2007 to 2009, and collaborated with the Intel ExaScience Lab in Belgium from 2009 to 2014. Overall, he has over 16 years of computer systems and architecture experience in both industry and academia.

Trevor Carlson’s research interests are in computer architecture targeting highly-efficient microarchitectures, secure processor designs, hardware/software co-design for energy efficiency, performance modeling and fast and scalable simulation methodologies. His goal is to improve the performance, efficiency and security of next-generation processors, covering applications that target edge (IoT) and cloud-scale applications. Additional topics of interest include processor and system security, as well as efficient accelerator design for next-generation machine learning systems. While a staff engineer at IBM, he helped to author 4 issued patents. During his PhD, in collaboration with the Intel ExaScience Lab, he co-developed the Sniper Multi-core Simulator which is being used by hundreds of researchers to evaluate the performance and power-efficiency of next generation systems, and continues to be used to explore next-generation processor design at Intel today. Starting as a researcher at imec, and as a postdoctoral researcher at Uppsala University, and continuing to his current work at NUS, he investigates processor architectures to more efficiently handle long-latency memory accesses. Dr. Carlson’s research has been published at leading journals and conferences in computer architecture and simulation such as the International Symposium on Computer Architecture (ISCA), the International Symposium on Microarchitecture (MICRO) and the International Symposium on High Performance Computer Architecture (HPCA).


12 December, 1640 - 1720, InFuse@Fusionopolis

Smart Building & Districts

Benoit Delinchant (UGA, MIAI & G2E)


Energy in Building and Districts is of great interest in urban areas due to the growth of consumption and the introduction of decentralized renewable production. Our research is about energy system design and energy management, including human in the loop.

3 axes of research are conducted:

  • Axis 1: Energy management and flexibility (residential / tertiary / industry)

  • Axis 2: Decision support tools

  • Axis 3: Data Production for AI


Benoit Delinchant is senior researcher in Electrical Engineering. Associate Professor at the University of Grenoble Alps since 2004 and works at G2ELab (Grenoble Electrical Engineering Laboratory). His research regards new methodologies such as AI, for the modeling and optimization during design and operation of energy systems such as smart grid and smart buildings.

The Role of AI in the Next Interaction Paradigm

Shengdong Zhao (NUS, School of Computing)


Interaction paradigms (the style of interaction between human and computers) can significantly change the way we work and live. However, as much as we are empowered by interaction paradigms, we are also significantly constrained by them. Existing interaction paradigms limits our movements and activities, which can negatively affect our overall well-being. Desktop computing, described as “sitting at a desk, interpreting and manipulating symbols”, isolates human beings from interacting with other human beings and nature. Mobile computing, although free us from the office environment, demands constant eye-and-hand engagement, leading to the notorious phenomenon called “smartphone zombies”. We need a new style of interaction that can better support human activities in nature and with other people as well as reducing cognitive load by blending reactive operations with appropriately designed proactive initiatives that can offer just-in-time assistance. This new interaction paradigm cannot work well without the help of AI. In this talk, I will discuss how AI may work with wearable computing, sensors, multimodal I/O, and distributed networking, etc. to enable the new synergistic & proactive interaction. The AI-enabled interaction paradigm can allow humans to interact with information while engaging in a variety of activities; therefore, facilitates a holistic lifestyle where essential human needs in work and life can be more seamlessly blended and fulfilled.

Dr. Shengdong Zhao is an Associate Professor in the Department of Computer Science, National University of Singapore. He established the NUS-HCI research lab. Dr. Zhao completes his Ph.D. degree in Computer Science at the University of Toronto. He also holds a Master’s degree in Information Management & Systems from the University of California at Berkeley. Dr. Zhao has a wealth of experience in developing new interface tools and applications (i.e., Draco, won best iPad App of the year in 2016), and publishes regularly in top HCI conferences and journals. He also works closely with the industry, and currently serves as a senior consultant with the Huawei Consumer Business Group. Dr. Zhao frequently serves in program committees of top HCI conferences, and will work as the paper chair for the ACM SIGCHI 2019, and 2020 conferences. More information about Dr. Zhao and the NUS-HCI lab can be found at and


13 November, 1010 - 1050, I3, Seminar Room 1, NUS

Explainability for Learning Systems using Logic

Nicholas Asher (CNRS, ANITI & IRIT)


I will explore how to use tools from logic to extract explanations of program behavior from learning systems.


Nicholas Asher is a CNRS director of research with IRIT in Toulouse. He is the scientific director of ANITI, the Artificial and Natural Intelligence Toulouse Institute

Conversational AI: Data Augmentation, Dialogue Comprehension, and Summarization Generation

Nancy F. Chen (A*STAR, I2)


Despite the avid interest and development of dialogue systems, there are still many challenges to address before conversational AI technology can be adopted widely with ease. In this talk, I will discuss three technical challenges. First, while there has been much work in natural language processing (NLP) on developing computational models to process documents, such approaches might not always be directly portable to dialogue processing, as the linguistic characteristics of documents and spoken conversations are intrinsically distinct: documents are one-way communications between the writer and the reader, whereas spoken conversations are spontaneous, dynamic information exchanges between at least two speakers. Second, the scarcity of large-scale well-annotated linguistic resources for modeling spoken conversations makes it difficult to develop scalable solutions that are readily extendible or generalizable to different scenario setups and domains. Third, the implicit assumption that “the data is always correct” might need to be challenged, as language data reflects the imperfections of human behavior. In this work, we demonstrate how linguistically strategic data augmentation frameworks can enable end-to-end neural modeling of dialogue comprehension. In addition, we show how proposed neural architectures that consider dialogue turns and topic segments in spoken conversations can enhance dialogue comprehension and summarization capabilities, enabling targeted applications in healthcare, journalism, and education


Nancy F. Chen received her Ph.D. from MIT and Harvard in 2011. She worked at MIT Lincoln Laboratory on her Ph.D. research in multilingual speech processing. She is currently leading research efforts in conversational AI and natural language generation with applications to healthcare, education, journalism, and finance at the Institute for Infocomm Research (I2R), A*STAR. Dr. Chen led a cross-continent team for low-resource spoken language processing, which was one of the top performers in the NIST Open Keyword Search Evaluations (2013-2016), funded by the IARPA Babel program. Dr. Chen is a senior IEEE member, an elected member of the IEEE Speech and Language Technical Committee (2016-2018, 2019-2021), associate editor of IEEE Signal Processing Letters (2019-2021) and the guest editor for the special issue of “End-to-End Speech and Language Processing” in the IEEE Journal of Selected Topics in Signal Processing (2017). Dr. Chen has received numerous awards, including Best Paper at APSIPA ASC (2016), the Singapore MOE Outstanding Mentor Award (2012), the Microsoft-sponsored IEEE Spoken Language Processing Grant (2011), and the NIH Ruth L. Kirschstein National Research Award (2004-2008). In addition to her academic endeavors, Dr. Chen has also consulted for various companies ranging from startups to multinational corporations in the areas of emotional intelligence (Cogito Health), speech recognition (Vlingo, acquired by Nuance), and defense and aerospace (BAE Systems).


13 November,  1120 - 1240, I3, Seminar Room 1, NUS

Instance and Gesture Identification with Small Sized CNN for Information Access in Mobility

Jean-Pierre Chevallet (UGA, MIAI & LIG)


Information access in mobility exploits mobile devices like smartphones. In the case of the GUIMUTEIC project, we have explored the usage of mobile assistance to museum-guided tours. We have used the camera of phone to localize the user by analyzing the museum artifact in front of him and interpret their gesture in order to interact with the recorded guided tour. In this talk we present dedicated CNN for the instances and gesture recognition. We emphasis the Neural Network size reduction, in order to reduce memory footprint of the phone.


Dr. Chevallet Jean-Pierre is Associate Professor at the University of Grenoble Alpes (UGA) since 1993. From 2003, to 2005 he has join the CNRS as Director of IPAL (Image Processing and Application Lab). Following this first lab, he has founded a new International Mixed Unit (UMI 2955) between I2R (Institute for Infocomm Research), NUS (National University of Singapore) and CNRS based in Singapore called Image Perception Access and Language (IPAL), that he directed from 2006 to 2008. He has now joined the Laboratoire d’Informatique de Grenoble (LIG) at UGA. He is the co-founder of the French Association and Conference for Information Retrieval (ARIA and CORIA). He directed one of the first ICT Asia project (ISERE 2005-06). He has participated to several European projects and working group, to major Information Retrieval competitions including TREC, AMARYLLIS, and CLEF. He has produced more that 90 scientific papers, and was the supervisor of 12 Phd Students.

Semantic and Sentiment Analysis for Knowledge Graph Construction

Su Jian (A*STAR, I2R)


I will give a brief overview of the work at NLP unit on text based Knowledge Graph construction in this talk. It covers various technologies which we have been advancing on  information extraction, sentiment analysis for the real world applications. I’ll also introduce some large scale deployments of these technologies on search, personal assistant, information gathering and standard enforcement for commercial entities as well as government organisations.


Dr. Jian Su is Unit Leader of NLP, Co-Director of Baidu I2R Research Center / Collaboration at A*STAR I2R in Singapore. She has worked intensively on information extraction, sentiment analysis, discourse analysis, text mining and NLP application. She has been the Principal Investigator in various large scale technology deployments for Search, Dialogue, Information Gathering and Standard Enforcement. She is an officer (VP-Elect 2017, VP 2018, President 2019) and Advisory Board Member (2015 - ) of SIGDAT, organiser of EMNLP,  the largest and one of the oldest Special Interest Group (SIG) of ACL, the premier association of Computational Linguistics. She is also a founding executive board member (2018-2020) of Asia Pacific Chapter of The ACL (AACL). She is a past executive committee member of ACL (2012-2014) as well. She has also served as General / Program Chair of tier-1 NLP conferences, including EMNLP 2016 / 2015 as well as ACL 2009. She has been the founding Action / Associate Editor for the two flagship journals, Transactions of ACL (TACL) and ACM transaction of Information System(TIST)

Natural Language Processing for e-Health

Pierre Zweigenbaum (CNRS, LIMSI)


From the patient record to scientific publications through on-line user comments, natural language text plays an important role in electronic health. I will illustrate how Natural Language Processing helps to unlock information and knowledge from text found in patient records, in patient forums, in the scientific biomedical literature, and in doctor-patient dialogues. Current natural language processing methods heavily rely on self-trained word representations, known as word embeddings, whose quality depends on the availability of very large text corpora: in a specialized domain however, text corpora are necessarily smaller than in unrestricted domains. I will therefore present methods that aim to increase the quality of word embeddings in a specialized domain by exploiting large out-of-domain corpora and a priori domain knowledge.


Pierre Zweigenbaum is a Senior Researcher at LIMSI (Orsay, France), a laboratory of the French National Research Council (CNRS) at Université Paris-Saclay, where he leads the ILES Natural Language Processing group. Before CNRS he was a researcher at Paris Public Hospitals in an Inserm team for twenty years. He also was a part-time professor at the National Institute for Oriental Languages and Civilizations during ten years. His research focus is Natural Language Processing, with medicine as a main application domain. He is the author of over 250 peer-reviewed conference and journal publications. He was elected fellow of the American College of Medical Informatics in 2014 and fellow of the International Academy of Health Sciences Information in 2019.

Language Translation: Technologies, Challenges and Applications.

Aw Ai Ti (A*STAR, I2R)


Machine Translation is one of the most important applications of Natural Language Processing (NLP). Nowadays, Neural Machine Translation (NMT) systems achieve state-of-the-art translation performance. However, their effectiveness strongly relies on the availability of large amounts of parallel sentences, which hinders their applicability to different language pairs. Further, language varieties such as linguistic, lexical and grammar divergences among different languages have impact on their translations. In this work, we investigate machine translation for local languages with language varieties and under different conditions of language resources and demonstrate the customised machine translation engines that we have developed with benchmark results as compared to Google and Microsoft.


Ms Aw Ai Ti is the Head of Aural & Language Intelligence Department at the Institute for Infocomm Research (I²R). She leads the department in the development and implementation of A*STAR’s Audio, Speech and Language R&D strategies.

Prior to her current appointment, Ms Aw spent most of her career spearheading the research and development of Southeast Asian Language Processing and Machine Translation capabilities in Singapore and has successfully established the team as a locally renowned Machine Translation technology provider in Southeast Asian languages. By integrating language linguistic knowledge into data statistics, the team overcame challenges of insufficient data and delivered multiple translational projects. The developed technologies were licensed to MNC, Government Agency, and Start-ups. Ms Aw also led the team to successfully deliver a locally relevant translation engine to support translation needs on Singapore official languages (Chinese, Malay).

 For her contribution and leadership in machine translation and Southeast Asian language processing, she was awarded the ASEAN Outstanding Engineering Achievement Team Award (2015), IES Prestigious Engineering Achievement Team Award (2015), President’s Technology Team Award (2013), I2R Achiever Award, Research (2009) and TEC Innovator Award (2003).

TOUR: School of Computing Robotics Living Lab
Towards Human-Centered Robot Autonomy

David Hsu (NUS, School of Computing)


David Hsu is a professor of computer science at the National University of Singapore (NUS) and the Director of AI Laboratory.  He is an IEEE Fellow. He received BSc in Computer Science & Mathematics from the University of British Columbia and PhD in computer science from Stanford University. At NUS, he co-founded NUS Advanced Robotics Center in 2013. He founded the AI Laboratory in 2019 and is serving as the director.   His research spans robotics, AI, and computational structural biology. In recent years, he has been working on robot planning and learning under uncertainty for human-centered robots. He has chaired or co-chaired several major international robotics conferences, including International Workshop on the Algorithmic Foundation of Robotics (WAFR) 2004 and 2010,  Robotics: Science & Systems (RSS) 2015, and IEEE International Conference on Robotics & Automation (ICRA) 2016. He currently serves on the editorial boards of International Journal of Robotics Research (IJRR) and Journal of Artificial Intelligence Research (JAIR).


13 November, 1500 - 1540, I3, Seminar Room 1, NUS

TheoremKB: Towards a Knowledge Base of Mathematical Results

Pierre Senellart (ENS)


In today's academic search engines and databases, the basic unit of information is a scientific article (usually a PDF document). But the actual unit of information of use by scientists in mathematical sciences is not the scientific article per se, but the mathematical results (theorems, lemmas, etc.) it contains: their statements, proofs, and possible other metadata. In this talk, we present a vision and preliminary work on TheoremKB, a project to turn the scientific literature in these fields from a collection of PDF articles to an open knowledge base of theorems where mathematical results are the object of interest, that can be explored in new ways. We intend to rely on a broad combination of techniques from information extraction, document engineering, machine learning, knowledge representation.


Pierre Senellart is a Professor in the Computer Science Department at the École normale supérieure (ENS, PSL University) in Paris, France, and an Adjunct Professor at Télécom Paris. He is an alumnus of ENS and obtained his M.Sc. (2003) and Ph.D. (2007) in computer science from Université Paris-Sud, studying under the supervision of Serge Abiteboul. He was awarded an Habilitation à diriger les recherches in 2012 from Université Pierre et Marie Curie. Before joining ENS, he was an Associate Professor (2008–2013) then a Professor (2013–2016) at Télécom Paris. He also held secondary appointments as Lecturer at the University of Hong Kong in 2012–2013, and as Senior Research Fellow at the National University of Singapore from 2014 to 2016.

Nonnegative Matrix Factorisation for Data Processing

Cédric Févotte (CNRS, IRIT, Université de Toulouse)

Data is often available in matrix form, in which columns are samples, and processing of such data often entails finding an approximate factorisation of the matrix in two factors. The first factor (the “dictionary”) yields recurring patterns characteristic of the data. The second factor (“the activation matrix”) describes in which proportions each data sample is made of these patterns. Nonnegative matrix factorisation (NMF) has become a popular technique for analysing data with nonnegative values, with applications in many areas such as in text information retrieval, user recommendation, hyperspectral imaging or audio signal processing. The presentation will give an overview of NMF and some ongoing works in my group [1] and with Vincent Tan from NUS (follow-up talk by Vincent in the next slot) [2,3].

[1] ERC project FACTORY,

[2] V. Y. F. Tan & C. Févotte. ``Automatic relevance determination in nonnegative matrix factorization with the beta-divergence’’, IEEE PAMI 2013.

[3] R. Xia, V. Y .F. Tan, L. Filstroff & C. Févotte. ``A ranking model motivated by nonnegative matrix factorization with applications to tennis tournaments’’, Proc. ECML-PKDD 2019.


Cédric Févotte is a CNRS senior researcher with Institut de Recherche en Informatique de Toulouse (IRIT). Previously, he has been a CNRS researcher at Laboratoire Lagrange (Nice, 2013-2016) & Télécom ParisTech (2007-2013), a research engineer at Mist-Technologies (the startup that became Audionamix, 2006-2007) and a postdoc at University of Cambridge (2003-2006). He holds MEng and PhD degrees in EECS from École Centrale de Nantes. His research interests concern statistical signal processing and machine learning. He is the principal investigator of the European Research Council (ERC) project FACTORY (New paradigms for latent factor estimation, 2016-2021).

A Ranking Model Motivated by NMF with Applications to Tennis Analytics (and other research activities in my group)

Vincent Tan (NUS, Mathematics)


In this talk I will present past and ongoing collaborations with Cédric Févotte on the use of NMF for the ranking of professional tennis players from tournament performance data [1]. I will also talk briefly about some other activities in my research group at NUS. In particular, I will introduce a new framework--based on sum-of-squares optimization---for unifying and systematizing the performance analysis of first-order black-box optimization algorithms for unconstrained convex minimization [2].


Vincent Tan is a Dean's Chair Associate Professor in the Department of Electrical and Computer Engineering and the Department of Mathematics at NUS.


Transparency in AI and Ranked Retrieval

Philippe Mulhem (UGA, MIAI Grenoble & G2E)


I will present some works achieved in the MRIM team of grenoble regarding explainabiity in image classification using CNNs and works toward transparency in ranked retrieval (Web and job search). The main idea behind this work is to go toward explaining why a learning algorithm gives one results, or at least how large are the impact of some input elements on ranking systems.


Philippe Mulhem has published more than 120 papers in national and international conferences and journals. He was during 1998 and 2003 co-director of the IPAL Singaporean/French joint laboratory located in Singapore. He was head of the Grenoble research work for the AVEIR ANR 2006 project, head of the Regional projects CLICIDE (2013) and RespIR (2014-2017). He participated the PENG (FP6) european project, and to the ANR EIRAP project (2007-2009). He is one of the heads of the Translago french platform (, dedicated to organize and promote transparency for information retrieval and learning algorithms. He supervised 9 PhD thesis students. His focus of interest are related to the models and evaluation of information retrieval. He was one of the organizers of the Trec Microblog Search evaluation task for CLEF (Conference and Labs of the Evaluation Forum) between 2014 and 2017. The MRIM research group MRIM (Modeling of Multimedia Information Retrieval) in which P. Mulhem focuses is one of leading research groups in France in the areas of Information Retrieval and Information Access. The MRIM Team studies models and algorithms for efficient access to information contained in large, multimedia, multilingual and semi-structured data collections. Its activities focus mainly on the modeling of multimedia data (text/image/video) for Information Retrieval and Information Filtering, personalization, machine learning models and methods for information retrieval, system experimentation and evaluation. For system evaluation, MRIM is actively involved in the evaluation of our systems and prototypes. This includes a continuous analysis of system quality, usability studies, and a participation to standard IR evaluation campaigns (TREC, INEX, CLEF). He participated to specific challenges (VideOlympics, Star Challenge).

Detecting Fake Videos

Terence Sim (NUS, School of Computing)


Fake videos, a type of fake news, are commonly spread via social media. They are increasingly difficult to spot by the naked eye, because modern video editing tools that leverage AI have made them more realistic. Fake news and fake videos are the weapons of modern cyber warfare. They spread disinformation, and are used to subvert governments, undermine democracy, create fear and confusion in the military and civilian population, and instigate violence. Detecting and containing such videos become a vital part of the defense strategy of any modern nation. Recently, we started tackling this problem. We are developing novel methods, based on biometrics, to detect fake videos. Unlike many existing methods, we do not detect artifacts (imperfections of the faking), because such artifacts will disappear as fake quality improves. Biometric signatures, however, will still be detectable, because they are the inherent movements of the person.


Terence actively conducts research in biometrics and computer vision, employing AI and Machine Learning to create powerful new algorithms that analyze and synthesize images. He is also a Principal Investigator at NCRiPT, a strategic research center at NUS that develops privacy-preserving technologies. His current projects include understanding the privacy of gait biometrics, protecting facial images from machine spying, as well as detecting fake videos.

Terence has published over 100 papers in top international journals and conferences. He obtained his SB from the Massachusetts Institute of Technology, MSCS from Stanford University, and PhD from Carnegie Mellon University. Terence is also an award-winning teacher, and an engaging speaker.

Housing: Hotel Meritus Mandarin, Orchard Road.