University Of Waterloo, Canada
Title:
Visual SLAM for the Real World
Abstract:
We have developed a hybrid Visual Simultaneous Localization and Mapping (VSLAM) method that combines the advantages of direct and indirect SLAM and is able to function in real time using standard computer architectures. Most VSLAM methods, including ours, assume the world is static however most applications and the real world is dynamic. Dynamics are typically handled as outliers and the results are less than desirable. We will present some of our attempts on extending our hybrid approach to dynamic environments. We will also touch upon other challenges such as on the fly unstructured geometric and photometric camera calibration which is a necessity. We will also discuss the point cloud maps that are built and the challenge of scaling them to large scale operations.
Biography:
Prof. John S. Zelek is a leading expert in computer vision and robotics. In particular, his current research interests include 3D computer vision, structure from motion, SLAM (Simultaneous Localization and Mapping), infrastructure monitoring, deep learning, anomaly detection, human pose tracking and detection as well as sports analytics from video. His research has led to the spinoff of 3 startup companies, 2 patents, four best paper awards at international conferences as well as a distinction award from the Canadian Computer and Robotic vision society. He has authored and co-authored more than 250 journal and conference publications. He has been on numerous conference organization committees in addition to chairing 2 Canadian Computer and Robotic vision conferences. Prof. Zelek's students, including undergraduate and graduate students have successfully transferred technology developed as part of graduate work and design projects under Zelek's supervision and made them the catalyst for new startup companies.
Homepage: https://uwaterloo.ca/systems-design-engineering/profile/jzelek
McMaster University, Hamilton, Ontario, Canada
Title:
Multi-modal Large Language Models: Breaking Down Barriers to Human-Like AI
Abstract:
Large language models (LLMs) have made significant strides in natural language processing, but their application has been limited to unimodal datasets, such as text or speech. However, we are now seeing a rapid evolution towards multimodal data, where LLMs must process a range of inputs such as text, images, and videos to achieve a more human-like understanding of the world. In this keynote talk, we will explore the latest developments in multi-modal LLMs, which integrate visual and textual information to model real-world scenarios more effectively. We will also discuss the challenges and opportunities associated with multi-modal LLMs and their applications. Finally, we will outline the future directions of multi-modal LLMs and their potential impact on the field of AI, including breaking down barriers to achieve human-like understanding of the world.
Biography:
Dr. Hamidreza Mahyar is an Assistant Professor in Faculty of Engineering at McMaster University. Before joining McMaster University, he was a postdoctoral research fellow at Boston University and Technical University of Vienna working with Prof. Eugene Stanley. Dr. Mahyar received his Ph.D. in Computer Science from Sharif University of Technology. His research interests lie at the intersection of machine learning, network science, and natural language processing, with a current emphasis on the fast-growing subject of graph neural networks with applications in social networks, recommendation systems, and drug discovery.
In addition to his research and teaching experiences, he has been involved in various industrial projects. He was a deep learning scientist in SemI40 and iDev40, the two biggest projects in Industry 4.0 Europe, at Infineon Technologies. He was also a senior machine learning scientist at ProteinQure, working on AI drug discovery. Dr. Mahyar is now the Director of AI Systems at BrainMaven, developing machine intelligence for human flourishing.
Homepage: https://www.eng.mcmaster.ca/faculty/hamidreza-mahyar/#overview
University of Louisiana at Lafayette, United States
Title:
AI+Systems: Bridging the Infrastructural Gap between Machine Learning and Edge-to-Cloud Systems
Abstract:
Over the past decade, human beings experienced a paradigm shift from the “communication everywhere” to the “computation everywhere”–enabled via a spectrum of distributed computing systems, from IoT and Edge devices to Fog and Cloud systems. It is the cooperation of these systems across the spectrum that will empower machine intelligence. Envisioning this paradigm shift, my research endeavor encompasses two thrusts that collectively enable an intelligent world that can positively impact the citizens’ lives. The first research thrust is to investigate Machine Learning (ML) algorithm-system software co-design and make the ML-based applications “usable” across Edge-to-Cloud continuum. In this regard, our NSF-funded project, SmartSight, has been developed to enable ambient perception for blind and visually impaired people (VIP), thereby, improving their quality of life and social inclusion. SmartSight is an example of AI+Systems for social good that operates based on smart glasses and provides real-time perception for its user via bridging the infrastructural gap between the ML methods and Edge-to-Cloud computing systems. Inspired from the human brain operation, we design platforms that function across the computing tiers to perform human-like cognition, in actions such as “concurrent perception”, “concentration”, and “progressive perception” within real-time constraints. My second research thrust, funded under NSF Career, is to develop the next generation of serverless Cloud systems that is domain-specific, heterogeneous, and interoperable. Such Clouds must offer high-level abstractions that can alleviate the burden of Cloud-native programming and democratize it. In that regard, we develop a new paradigm, known as the Object as a Service (OaaS), that hides the complexity of both Cloud resource and data management from the developer’s perspective.
Biography:
Dr. Mohsen Amini Salehi is an Associate Professor and Francis Patrick Clark/BORSF Endowed Professorship holder at the School of Computing and Informatics, University of Louisiana Lafayette. He is the director of High Performance and Cloud Computing (HPCC) Laboratory. Dr. Amini is an NSF CAREER Awardee and, so far, he has had 7 research projects funded by National Science Foundation (NSF) and Board of Regents of Louisiana (totaling $3.0 M). He has received five awards and certificates from University of Louisiana at Lafayette in recognition of his innovative research. His paper was nominated for the “Best Paper Award” in 33rd International Parallel and Distributed Processing Symposium (IPDPS ’19), Rio de Janeiro, Brazil. Dr. Amini has received his Ph.D. in Computing and Information Systems from Melbourne University, Australia, in 2012, under supervision of Professor Rajkumar Buyya. He has been a postdoctoral fellow at Colorado State University (2012—2013), and at University of Miami (2013—2014). He is an active researcher in the community and regularly serves in the organizing premier conferences in Distributed and Cloud Computing, such as CCGRID and IEEE Cloud. He has more than 70 publications in top-tier venues and has filed 4 U.S. patents. His research interests are in AI+Systems, resource allocation across edge-to-cloud continuum, heterogeneous computing, trustworthy edge-cloud, and virtualization.
Homepage: https://sciences.louisiana.edu/node/138
