Past Seminars: 2024
June 21
How to Specify Aligned Reinforcement Learning Objectives
Dr. Brad Knox, Associate Research Professor, University of Texas at Austin
Abstract:
I will discuss how practically to specify reinforcement learning (RL) objectives through careful design of reward functions and discounting. I'll provide tools for diagnosing misalignment in RL objectives, such as finding preference mismatches between the RL objective and human judgments, and examining the indifference point between risky and safe trajectory lotteries. Common pitfalls that can lead to misalignment are highlighted, including naive reward shaping, trial-and-error reward tuning, and improper handling of discount factors. I also sketch best practices for designing interpretable, aligned RL objectives and discuss open problems that hinder the design of aligned RL objectives in practice.
Presenter Bio:
Dr. Brad Knox is a Research Associate Professor of Computer Science at the University of Texas at Austin. His research has largely focused on the human side of reinforcement learning. He is currently concerned with how humans can specify reward functions that are aligned with their interests. Brad’s dissertation, “Learning from Human-Generated Reward”, comprised early pioneering work on human-in-the-loop reinforcement learning and won the 2012 best dissertation award for the UT Austin Department of Computer Science. His postdoctoral research at the MIT Media Lab focused on creating interactive characters through machine learning on puppetry-style demonstrations of interaction. Stepping away from research during 2015–2018, Brad founded and sold his startup Bots Alive, working in the toy robotics sector. In recent years, Brad co-led the Bosch Learning Agents Lab at UT Austin and was a Senior Research Scientist at Google. He has won multiple best paper awards and was named to IEEE Intelligent System’s AI’s 10 to Watch in 2013.
June 14
Video Question-Answering in the Era of LLMs
Dr. Angela Yao, Assistant Professor, National University of Singapore
Abstract:
Video question-answering is an ideal test-bed for evaluating the reasoning capabilities of the latest multi-modal large language models (MLLMs). This talk will take a deep dive into the behaviours of multimodal LLMs for VideoQA. We show that despite their higher QA accuracy, MLLMs share many of the weaknesses of previous non-LLM techniques. For example, they are weak in temporal understanding, visual grounding, and display limited capability of multimodel reasoning. Moreover, despite featuring LLMs, the models are not robust to language variations like question rephrasing. Understanding these limitations is crucial not only for developing future VideoQA techniques but also for integrating MLLMs into everyday applications.
Presenter Bio:
Dr. Angela Yao is an associate professor in the School of Computing at the National University of Singapore. She received a PhD from ETH Zurich and a BASc from the University of Toronto. Angela leads the Computer Vision and Machine Learning group, with a special focus on vision-based human motion analysis. She is the recipient of the German Pattern Recognition (DAGM) award (2018) and Singapore’s National Research Foundation's Fellowship in Artificial Intelligence (2019).
June 12
Special Wednesday Seminar
The Power of Good Old-Fashioned AI for Urban Traffic Control
Dr. Mauro Vallati, Professor, University of Huddersfield, UK
Abstract:
AI approaches are increasingly being used in intelligent urban traffic control techniques, to improve mobility and reduce emissions. While a large number of studies leverage on machine learning techniques, the good old model-based AI is gaining traction. In this talk, we will first look into how urban traffic control is currently performed. Then, we will present a traffic signal optimization approach based on AI planning that is being trialed in the UK, and will discuss how GOFAI can provide useful tools for traffic engineers, domain experts, and practitioners.
Presenter Bio:
Dr. Mauro Vallati is Professor of AI at the University of Huddersfield, UK, where he is the director of the Research Centre on Autonomous and Intelligent Systems. He is an ACM Senior Member, and ACM Distinguished Speaker on artificial intelligence (AI) for the UK. He has extensive experience in real-world applications of AI methods and techniques, spanning from healthcare to train dispatching. Since 2016, he has led several research grants and contracts in the field of urban traffic control, leading to numerous high-impact academic publications, and patents filed in United Kingdom, China, and United States. In 2021 he was awarded a prestigious UKRI Future Leaders Fellowship for investigating AI-based autonomic urban traffic monitoring and control, with the aim of designing intelligent systems that can autonomously recognize the insurgence of traffic congestion and implement traffic light strategies to mitigate its impact on the urban traffic network.
June 10
Special Monday Seminar
Exploring in Sparse-Reward Domains
Dr. Guni Sharon, Assistant Professor, Texas A&M University
Abstract:
Sparse-reward domains pose significant challenges in reinforcement learning (RL) due to the rarity of feedback signals, making it difficult for agents to learn effective policies. This talk delves into the intricacies of exploration in such environments, balancing the need to explore new possibilities with exploiting known rewards—a dilemma known as the explore-exploit tradeoff. Attendees will gain insights into cutting-edge exploration techniques, practical challenges, and innovative solutions, equipping them with the knowledge to tackle sparse-reward problems in their own RL research and applications.
Presenter Bio:
Dr. Guni Sharon is an Assistant Professor at Texas A&M University in the Department of Computer Science & Engineering. His research interests encompass Artificial Intelligence, Intelligent Transportation Systems, Reinforcement Learning, and Combinatorial Optimization. Dr. Sharon has a strong theoretical foundation in AI, particularly in reinforcement learning, combinatorial search, multiagent route assignment, game theory, flow and convex optimization, and multiagent modeling and simulation. He has extensive experience applying these theoretical concepts to practical problems in traffic management and optimization. Committed to bridging theory and practice, Dr. Sharon aims to advance both the applicability of his research to real-world problems and the theoretical frameworks that underpin these solutions.
June 7
HonestDoor: Entrepreneurship and the Journey of Using Data Science to Disrupt Real Estate
Dan Belostotsky, Founder & CEO of HonestDoor
Co-hosted by Technology Alberta
Abstract:
HonestDoor is a technology company with a focus on data in real estate. With Canada’s largest real estate database, it uses AI/ML to estimate home values for all residential properties across Canada. I will be discussing how our company has used this to become a fast-growing site for hundreds of thousands of buyers and sellers. I’ll also discuss the challenges we’ve had both as a company and also in developing our product.
Presenter Bio:
Dan Belostotsky is the founder of the Digital Media company sold to the Pattison Group, founder of HonestDoor.com, investor in multiple tech companies and VC funds, and owner of the real estate company Otto Capital Group Inc.
May 31
NO SEMINAR - Computing Science Center Building Power Shut-down
May 30
Special Thursday Seminar
Evaluation of NLG Systems: From Reaction to Anticipation
Dr. Jackie Chi Kit Cheung, Associate Professor at McGill and Canada CIFAR AI Chair at Mila
Abstract:
Large language models (LLMs) can make factual errors or unsupported claims—sometimes called hallucinations—which can be costly or harmful to the developers, users, and other stakeholders relevant to an NLG system. This type of errors had been anticipated by the NLG research communities years in advance of the release of popular pretrained LLMs, yet they still occur in deployed LLMs. Hallucinations, however, are not the only issue faced by LLMs. In this talk, I ask: how can we anticipate and mitigate potential issues with NLG systems before they become the next embarrassing headline? First, I discuss our work in systematically surveying the recent NLP literature on automatic summarization as a study of how NLP researchers discuss and frame responsible AI issues. Overall, we find that the papers we examined typically do not discuss downstream stakeholders or imagine potential impacts or harms that the summarization system may cause. Next, I present our current efforts to encourage more structured reflection of evaluation practices in NLP, focusing in particular on benchmark design and creation. I introduce our novel framework, Evidence-Centred Benchmark Design, inspired by work in educational assessment.
Presenter Bio:
Dr. Jackie Chi Kit Cheung is an associate professor at McGill University's School of Computer Science, where he co-directs the Reasoning and Learning Lab. He is a Canada CIFAR AI Chair and an Associate Scientific Co-Director at the Mila Quebec AI Institute. His research focuses on topics in natural language generation such as automatic summarization, and on integrating diverse knowledge sources into NLP systems for pragmatic and common-sense reasoning. He also works on applications of NLP to domains such as education, health, and language revitalization. He is motivated in particular by how the structure of the world can be reflected in the structure of language processing systems. He is a consulting researcher at Microsoft Research Montreal.
May 24
NO SEMINAR - UPPER BOUND
May 17
Tracking Changing Probabilities via Dynamic Learners
Dr. Omid Madani, recently, Principal Engineer at Cisco Secure Workload
Abstract:
The world is always changing, and yet may be stable enough for learning to predict with probabilities. Due to change, however, the estimated probabilities need to be modified at times, possibly substantially. In the context of online multiclass probabilistic prediction via finite-memory predictors, we present two moving average techniques, one based on the exponentiated moving average (EMA) and one based on queuing a few count snapshots. We show that the combination, and in particular supporting dynamic predictand-specific learning rates, offers advantages in terms of faster change detection and convergence. In the process, we touch on a variety of topics including internal vs external non-stationarity, stability plasticity dilemma, bias and variance, probabilistic convergence, and the challenges of using log-loss for evaluation when the input stream includes unseen (possibly noise) items, for which we develop approximate propriety. We motivate this task within the framework of prediction games, an approach to self-supervised lifelong cumulative learning of a hierarchy of concepts.
Presenter Bio:
Omid is interested in all aspects of intelligence and mind, especially from a computational perspective. Most recently, he was a founding member of the Tetration Analytics division of Cisco, and led the machine learning efforts at Tetration and Cisco Secure Workload. Prior to Cisco, Omid held research and engineering positions at Google Research (the perception group), SRI, and at Yahoo! Research, developing and applying machine learning to a variety of problems such as web search, video indexing, and network security. Omid received his BS from the University of Houston, and PhD from the University of Washington, and was a Postdoctoral fellow at the University of Alberta in Edmonton, working with Russell Greiner.
May 10
Learning to Rephrase Inputs for Downstream Text Classification
Saeed Najafi, PhD student, University of Alberta
Abstract:
Recent NLP research has developed effective techniques to control or alter the behavior of Pre-trained Language Models (PLMs). However, current PLM control techniques have not considered altering the original input text to improve the performance of the PLMs. We investigate this idea by training a secondary, smaller PLM to paraphrase the original input at training and test time, thus augmenting the existing data and improving model performance. We experiment on six text classification datasets, demonstrating that incorporating paraphrase augmentation during both training and testing phases enhances the performance of discrete/soft prompt optimization and efficient tuning techniques. Finally, we discuss our future work, which aims to extend this learning framework to multi-hop question-answering datasets for complex question decomposition.
Presenter Bio:
Saeed Najafi is a fourth-year PhD student in Computing Science at the University of Alberta, working with Professor Alona Fyshe. He explores various topics, including parameter-efficient optimization, question-answering, and policy optimization within LLMs. Previously, Saeed earned a Master's degree in Computing Science from the University of Alberta and a Bachelor's degree from Amirkabir University of Technology. He has experience working at both small-scale startups and big tech companies, applying various research techniques in applied NLP projects.
May 3
DOCEO AI journey in implementing AI
Ahmad Jawad, CEO of DOCEO (DOKEO) AI, CEO of Intellimedia, Co-founder Edmonton Research Park Business Consortium, and AI champion
Co-hosted by Technology Alberta
Abstract:
As an EdTech software company, we specialize in developing solutions that collect and analyze data on K-12 students to enhance educational practices and support student learning. Currently, we are in the planning phase of integrating AI technology to offer predictive analytics and tailored recommendations. This is aimed to facilitate and augment decision-making processes in both educational planning and student support services. It is important to recognize that schools manage more than just student data; they also handle financial, HR, and feedback data, all of which play a crucial role in influencing student success.
Presenter Bio:
Ahmad Jawad is an established member of the Alberta technology community, having founded and expanded Intellimedia LP and DOCEO AI into a thriving corporation in the EdTech field across Canada that focuses on education analytics, digitized form management in school districts and leveraging AI as a data informed decision making tool to support student learning. Ahmad holds a Bachelor of Science from the University of Alberta, an Executive Master of Business Administration and completed a Management Excellence program at Harvard Business School. Ahmad is a believer in continuous learning. Ahmad speaks three languages: Arabic, English, and French.
Ahmad is a passionate community and industry steward, committing both time and resources to supporting several not-for-profit and industry organizations. Ahmad is on the board of Technology Alberta, board member of the Computer Science Industry Advisory Board at the University of Alberta and Edmonton Regional Innovation Network steering committee. Ahmad is the co-founder and director of the Edmonton Research Park Business Consortium, promoting innovation and collaboration between companies in the Edmonton research park and larger innovation community, and strongly believes in supporting a leading edge ecosystem in Alberta to make it attractive to all talents, innovators and investors throughout Canada.
April 26
A Prototype of an Autonomous Control System using Sound for Reducing Anxiety in ICU Patients
Dr. Martha Steenstrup, Independent
Abstract:
Patients in critical care commonly experience anxiety which can negatively affect both recovery and longer-term health. Sound has proven effective in reducing anxiety and is, in many respects, preferable to conventional pharmacological treatment. In this talk, I describe the design and implementation of and initial experimentation with a prototype of an autonomous control system that employs sound for anxiety reduction and reinforcement learning in patient-appropriate sound selection.
Presenter Bio:
Volunteer dogsbody and wannabe polymath who strives to live lightly on the earth.
April 19
Revisiting Overestimation in Value-based Deep Reinforcement Learning
Prabhat Nagarajan, PhD student, University of Alberta
Abstract:
Since the introduction of Deep Q-networks (DQN), deep reinforcement learning has had several back-to-back advances. One such early advance is Double DQN, which was meant to address the well-known overestimation phenomenon in algorithms such as Q-learning and, consequently, DQN. In light of recent insights, we revisit both DQN and Double DQN. We observe that the optimizer-loss function combination can substantially impact overestimation. Moreover, we explore alternatives to Double DQN where two Q-functions are trained and find that they further reduce overestimation upon Double DQN.
Presenter Bio:
Prabhat Nagarajan is currently a PhD student working with Marlos C. Machado and Martha White. Prior to his PhD, he completed a Master’s in Computer Science at UT Austin, where he researched reproducibility in deep reinforcement learning and preference-based inverse reinforcement learning. His industry research experience includes 3 years at Preferred Networks where he worked on open source deep reinforcement learning and reinforcement learning for robotics. He has also interned at Sony AI, researching efficient methods for reward shaping in reinforcement learning.
April 12
From Low-Rank Adaptation to High-Rank Updates: Training Large Models with Limited Memory
Yongchang Hao, PhD student, University of Alberta
Abstract:
As pre-trained models grow larger, fine-tuning them becomes challenging due to memory constraints. Low-rank adaptation (LoRA) reduces parameters needing fine-tuning but limits update expressiveness. In this presentation, we will analyze the dynamics of LoRA and show it can be approximated by random projections. Based on this observation, we propose our method Flora, which is able to achieve high-rank updates by resampling the projection matrix. Our method has the same space complexity as LoRA while demonstrating superior performance on several benchmarks.
Presenter Bio:
Yongchang Hao is currently a first-year PhD student at the University of Alberta working with Professor Lili Mou. His research interests are broad in machine learning, including language models, reinforcement learning, and optimization. He has published relevant papers at internationally renowned conferences such as NeurIPS, ICLR, ACL, NAACL. Previously, Yongchang obtained a Master's degree from the University of Alberta and a Bachelor's degree from Soochow University. He previously interned at Tencent AI Lab and Google. His long-term research goal is to build compute-efficient large-scale intelligence systems.
April 5
Business Case For AI/ML in Health and Wellness Industry
Mehadi Sayed, President & CEO, Clinisys EMR Inc.
Co-hosted by Technology Alberta
Abstract:
The impact of AI/ML in health and wellness industry has become significantly prominent in the last few years. In this presentation, I will share our company’s journey over the past decade and demonstrate a variety of business cases where the use of AI/ML has made the most impact. I will discuss the challenges, successes, impediments, and even failures encountered by a company such as Clinisys. Finally, I will explore the notable shifts that we are witnessing in the health and wellness domain, drawing upon examples from some of our current projects.
Presenter Bio:
Mehadi Sayed is the founding President and CEO of Clinisys EMR Inc. The company has been operational for over 12 years with its head office located at the Edmonton Research Park. Under Mehadi Sayed's guidance, Clinisys data scientists have worked on numerous projects in the health and wellness domain covering a wide variety of topics, including but not limited to patient re-admission, cancer care resource optimization, and the use of ML in predictive modelling for athlete data.
April 4
Special Thursday Seminar
Explainable Planning and Decision Making
Dr. William Yeoh, Washington University in St.Louis
Abstract:
With increasing proliferation and integration of planning and decision-making systems in our daily life, there is a surge of interest in systems that can explain the reasons for their decisions. In this talk, I will describe such explainable systems through the lens of model reconciliation, which is the problem of reconciling the mental models of the AI system and the human user receiving the explanation. I will also provide an overview of our work in this space, including how logic- and argumentation-based approaches can be used.
Presenter Bio:
William Yeoh is an associate professor in the Department of Computer Science and Engineering and the director of the Division of Computational and Data Sciences at Washington University in St. Louis. His research focus is on AI with the goal of enabling effective human-AI collaboration. He builds on his primary expertise in distributed constraint optimization as well as automated planning and scheduling to develop optimization algorithms for human-aware agent-based systems. He is an NSF CAREER recipient and was named in IEEE's AI's 10-to-Watch list in 2015. He currently serves on the editorial board of the Artificial Intelligence journal, on the IFAAMAS Board of Directors, on the ICAPS Council, and on the SoCS Council.
March 29
NO SEMINAR - UNIVERSITY CLOSED FOR GOOD FRIDAY
March 22
Directions of Curvature as an Explanation for Loss of Plasticity
Alex Lewandowski, PhD student, University of Alberta
Abstract:
I will argue that the Hessian rank, which counts the number of curvature directions, determines the learning ability of a neural network. While randomly initialized neural networks have a high Hessian rank, current learning algorithms reduce the Hessian rank over time. Thus, loss of curvature directions explains loss of plasticity. Lastly, we show that regularizers which prevent loss of plasticity also prevent loss of curvature, motivating a simple distributional regularizer that proves to be effective across the problem settings considered.
Presenter Bio:
Alex Lewandowski is a PhD student supervised by Dale Schuurmans and Marlos C. Machado at the University of Alberta. His research interest is in understanding scalable continual learning algorithms, with the goal of developing algorithms capable of learning autonomously at every scale.
March 15
The AI Crewmate: A Crush, A Crutch, or A Confidant?
Dr. Rajagopalan Srinivasan, University of Alberta & Indian Institute of Technology Madras
Abstract:
Modern-day industry is replete with large socio-technical systems, and people have to interact closely with complex technologies. Traditionally, it has been assumed that operators are the weak links who cause system failure. In this talk, we will explore how AI-based sensors such as eye tracking can offer insights into the operators’ cognition. These enable science-based approaches to study the dynamics in Human-AI teams with numerous applications, such as the design of human-aware interfaces and decision-support systems as well as training protocols.
Presenter Bio:
Dr. Rajagopalan Srinivasan is a Professor of Chemical Engineering and the Head of the American Express Lab for Data Analytics, Risk & Technology (DART Lab) at IIT Madras. He is currently a Visiting Professor at University of Alberta. Raj received his B.Tech from the Indian Institute of Technology Madras and PhD from Purdue University. Raj’s research program is targeted towards developing systems engineering approaches for the design and resilient operation of complex systems. He is a consultant to over 20 well-known companies such as ABB, ExxonMobil, Honeywell, and Shell. His research has been recognized by Best Paper Awards from several journals & conferences. He is an Associate Editor of several journals including PLOS One, Frontiers in Energy Research, and Process Safety and Environmental Protection.
March 8
Target Networks and Over-parameterization Stabilize Off-policy Bootstrapping with Function Approximation
Fengdi Che, PhD student, University of Alberta
Abstract:
Off-policy value prediction plays a vital role in reinforcement learning. It is crucial for real-world applications and enhances data efficiency during online learning. But divergence caused by the deadly triad, including bootstrapping, function approximation and off-policy learning, has bothered the field for a long time. Our work proves that combining a target network and over-parameterized linear function approximation stabilizes the learning, even with off-policy data. That is, the deadly triad can be eliminated by two simple, common and popular augmentations.
Presenter Bio:
Fengdi Che is a fourth-year Ph.D. student supervised by Rupam Mahmood. Her research interests focus on off-policy reinforcement learning and improvements in policy gradient algorithms. Recently, she has been reading about the learning ability of transformers and welcomes all kinds of discussions.
March 1
Monitored Markov Decision Processes (Mon-MDPs)
Montaser Mohammedalamen, PhD student, University of Alberta
Abstract:
In reinforcement learning (RL), an agent learns to perform a task by interacting with an environment and receiving feedback (a numerical reward) for its actions. However, the assumption that rewards are always observable is often not applicable in real-world problems. For example, the agent may need to ask a human to supervise its actions or activate a monitoring system to receive feedback. There may even be a period of time before rewards become observable, or a period of time after which rewards are no longer given. In other words, there are cases where the environment generates rewards in response to the agent's actions but the agent cannot observe them. In this paper, we formalize a novel but general RL framework --- Monitored MDPs --- where the agent cannot always observe rewards. We discuss the theoretical and practical consequences of this setting, show challenges raised even in toy environments, and propose algorithms to begin to tackle this novel setting. This paper introduces a powerful new formalism that encompasses both new and existing problems.
Presenter Bio:
Montaser (Monta) is a 2nd-year PhD student working with Prof Michael Bowling. Monta's research areas are around how to make agents learn with reward unavailability and how to make agents behave safely and cautiously when facing novel observations. Before joining the University of Alberta, Monta was a Reinforcement Learning (RL) engineer at SonyAI working on Designing a Multi-agent high dynamic environment in a physics simulator and transferring learned policies to the real world.
February 23
NO SEMINAR - U OF A READING WEEK
February 16
An experimentalist’s journey into RL theory: two successes and a failure
Abhishek Naik, PhD Student, University of Alberta
Abstract:
In this talk, I will discuss two new multi-step algorithms for the average-reward formulation. Multi-step methods can be much more sample-efficient than the more prevalent one-step methods. I will begin with a high-level overview of the “ODE approach” to proving the convergence of RL algorithms. I will then outline how I used those theoretical tools to prove the convergence of two algorithms—and to show that a third algorithm of mine can diverge even in the tabular setting. Finally, I will discuss simple experiments that validate the theory.
Presenter Bio:
Abhishek Naik is a Ph.D. candidate working with Rich Sutton. He will soon defend his dissertation on reinforcement learning in continuing (non-episodic) problems using average reward. After his Ph.D., he wants to continue performing AI research in the space industry. Among other things, he loves reading, playing hockey, and watching Formula 1.
February 9
Policy-Guided Heuristic Search
Dr. Levi Lelis, University of Alberta
Abstract:
In this talk, I will present Levin Tree Search (LTS) and some of its variants. LTS is a state-space search algorithm that uses a policy to guide its search. Variants of LTS use both a policy and a heuristic function to guide the search. Algorithms from the Levin family offer guarantees regarding the number of expansions required before finding a solution to search problems. These guarantees depend on the quality of both the policy and the heuristic function guiding the search. The guarantees are important because they offer a loss function---the Levin loss---that allows us to learn policies that minimize the size of the resulting search tree. We will explore learning schemas using neural networks and context models, with the latter offering additional guarantees. Specifically, the parameters of context models optimizing the Levin loss can be derived by solving a convex optimization problem. I will also present some empirical results. In particular, I will show what may be the fastest policy, learned from data, that is able to solve random instances of the Rubik's Cube.
I aim to make this presentation highly accessible, comparable to a lecture from a 300-level undergraduate course in our department. The talk will summarize the findings of three papers, which are listed below, including the paper that received the Distinguished Paper Award at IJCAI'23.
Single-Agent Policy Tree Search with Guarantees. Laurent Orseau, Levi Lelis, Tor Lattimore, and Theophane Weber. In the Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2018.
Policy-Guided Heuristic Search with Guarantees. Laurent Orseau and Levi Lelis. In the Proceedings of the Conference on Artificial Intelligence (AAAI), 2021.
Levin Tree Search with Context Models. Laurent Orseau, Marcus Hutter, and Levi Lelis. In the Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2023.
Presenter Bio:
Dr. Levi Lelis is an Assistant Professor at the University of Alberta, an Amii Fellow, and a CIFAR AI Chair.
February 2
Health Gauge: Innovating in Applied AI/ML Solutions in Digital Health
Randy Duguay, CEO of Health Gauge Inc.
Co-hosted by Technology Alberta
Abstract:
Randy Duguay presents his company and his experience as a founder of an AI/ML & digital health platform software solution provider. Health Gauge has hired a number of U of A grads from the Computing Science Department, and has a number of key connections with the U of A via ST Innovations and the Faculties of Engineering and Medicine. It also has key business partnerships in the United States, through MedWatch Technologies, and in France, with Tech2Heal, and have been continuing to innovate by tapping into the excellent resources at the U of A and NAIT.
Presenter Bio:
Randy Duguay has a B.Sc (Engineering), an M.Eng (Management) degree, as well as completed executive business development courses with Rotman Business School (University of Toronto), and Sloan Business School (MIT). Randy has over 25 years of experience in telecommunications engineering, new technology commercialization, and senior management positions, from large scale to startups. Randy’s experience includes 25 years of experience with TELUS Communications and TELUS Health Solutions, including senior management roles. For over 20 years, he has worked in health solutions and new technology development and in this time has worked with some of the world’s leading technology, software and solutions companies. Current roles and affiliations include: Director, and former CEO, AI/ML Innovations Inc.; Business Advisor, ST Innovations (UofA); Trustee, Glenrose Rehabilitation Hospital Foundation; Director, My Fertility Labs Inc.
January 26
Exploring Methods for Generating and Evaluating Skill Targeted Reading Comprehension Questions
Spencer von der Ohe, PhD student, University of Alberta
Abstract:
There have been many methods proposed to generate reading comprehension questions. However, few allow specific skills to be targeted and there is still a gap in quality between automatically generated and manually created questions. To narrow this gap, we present SoftSkillQG, a new soft-prompt based language model for generating skill targeted reading comprehension questions that does not require manual effort to target new skills. In this presentation, we explore the strengths of SoftSkillQG and how its weaknesses can be addressed.
Presenter Bio:
Spencer von der Ohe is a Computing Science master’s student at the University of Alberta. He is interested in natural language processing and how it can be applied to real world problems. Spencer is supervised by Dr. Alona Fyshe.
January 19
Exploring Creative Paths in DLC Generation
Johor Jara Gonzalez, PhD student, University of Alberta
Abstract:
In the context of video games, Downloadable Content (DLC) refers to additional content or features that can be incorporated into a game after its initial release. In the fields of Automated Game Design (AGD) and Procedural Content Generation (PCG), various approaches have been explored to augment game content. However, no work has attempted to automatically generate DLCs. This talk will cover our concept of how AGD and PCG can be employed to develop compelling DLCs. We’ll specifically focus on extending the existing level structure through level inpainting and generating additional mechanics through a reinforcement learning-based approach.
Presenter Bio:
Johor Jara Gonzalez is a PhD student at the University of Alberta and Alberta Machine Intelligence Institute (Amii). His research interests are focused on computational creativity, and more precisely DLC Generation and Mechanics in Video Games. He has presented at AAAI’s Artificial Intelligence and Interactive Digital Entertainment conference and reviewed for IEEE's Transactions on Games.
January 12
Advances in Data Driven Teleoperation: A synthesis of previous works & our own
Michael Przystupa, PhD student, University of Alberta
Abstract:
Data-driven teleoperation research investigates the application of machine learning models for mapping interactions between control interfaces (e.g., a joystick) and robotic arms to simplify human-robotic interactions. There is a real-world utility for such AI systems in assistive robotics, where people living with a disability rely on robots to complete daily tasks (e.g., opening doors, brushing teeth, etc.). Existing mappings involve cycling between controlling subsets of a robot’s pose. These can be tedious because they can require switching the control mode anywhere from 30 to 60 times to accomplish daily tasks. Data-driven mappings simplify teleoperation by mapping a user's inputs directly to desired robot motions, eliminating the need to switch the control mode. This presentation covers previous works investigating data aspects of such systems, including data collection procedures (supervised vs unsupervised) and choice of data modalities (robotic proprioceptive information, images, text, and belief). The latter half of the presentation discusses our own work explicitly enforcing teleoperation properties and briefly includes incorporating such models with shared autonomy systems. Overall, attendees can expect to learn how to build data-driven teleoperation systems and an introduction to existing research in this application of machine learning.
Presenter Bio:
Michael Przystupa is a 5th year PhD student at the University of Alberta working at the intersection of machine learning and robotics. His research has spanned the applications of deep learning in robotics topics including visual servoing, motor primitives and teleoperation. He is co-supervised by Dr. Martin Jagersand and Dr. Matthew Taylor.
January 5
NO SEMINAR - U OF A WINTER CLOSURE