Past Seminars: 2022

YouTube Video Coming Soon!

Abstract:
Many breakthroughs in one field are inspired by others. For example, neural networks are inspired by the structure and function of the human brain, and attention mechanisms borrow the idea from psychology. In this talk, we will present our recent findings in natural language processing inspired by the laws of human cognition, such as knowledge fusion, systematic generalization, and imitation learning. In the first one, we propose a plug-in framework called RoChBert to build a more robust language model explicitly for Chinese. By incorporating adversarial knowledge, we show how to fuse the necessary phonetic and glyph information into pre-trained representations to strengthen the robustness. In the next one, we investigate the extent to which neural networks can do the same as humans to generalize from old concepts to new ones systematically. We revisit this controversial topic from the perspective of meaningful learning, a concept from educational psychology. Our experimental results indicate that conventional sequence-to-sequence models can successfully one-shot generalize to novel concepts and compositions through shared semantic relationships, either inductively or deductively. In the last one, we reformulate text editing as an imitation game using behavioral cloning. Specifically, we convert standard sequence-to-sequence data into state-to-action demonstrations and propose to train an agent to mimic how humans revise texts iteratively, where the action space can be as flexible as needed. Overall, we hope this presentation will encourage and shed light on future studies at the junction of multiple fields. 

Presenter Bio:
Ning Shi is a 1st-year Ph.D. student working with Prof. Grzegorz Kondrak at the University of Alberta, associated with Alberta Machine Intelligence Institute (Amii). Previously, he was a senior algorithm engineer at Alibaba Group and a machine learning engineer at Learnable. Ning received an M.S. degree in Computer Science at the Georgia Institute of Technology, an M.S. degree in Applied Data Science at Syracuse University, an M.S. degree in Management and Systems at New York University, and a B.Mgt. degree in E-commerce at Donghua University. His current research interests primarily stand on computational linguistics and natural language processing at the junction of human cognition. 

https://www.youtube.com/watch?v=4_YmJrGsexw&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=26

Abstract:
Algorithmic fairness has received increased attention in socially sensitive domains. While rich literature on mean fairness has been established, research on quantile fairness remains sparse but vital. To fulfill great needs and advocate the significance of quantile fairness, we propose a novel framework to learn a real-valued quantile function under the fairness requirement of Demographic Parity with respect to sensitive attributes, such as race or gender, and thereby derive a reliable fair prediction interval. Using optimal transport and functional synchronization techniques, we establish theoretical guarantees of distribution-free coverage and exact fairness for the induced prediction interval constructed by fair quantiles. A hands-on pipeline is provided to incorporate flexible quantile regressions with an efficient fairness adjustment post-processing algorithm. We demonstrate the superior empirical performance of this approach on several benchmark datasets. 

Presenter Bio:
Dr. Linglong Kong is a professor in the Department of Mathematical and Statistical Sciences at the University of Alberta. He holds a Canada Research Chair in Statistical Learning, a Canada CIFAR AI Chair, and is a fellow of the Alberta Machine Intelligence Institute (AMII). His publication record includes more than 70 peer-reviewed articles in top journals such as AOS, JASA and JRSSB as well as top conferences such as NeurIPS, ICML, ICDM, AAAI, and IJCAI. Dr. Kong currently serves as associate editor of the Journal of the American Statistical Association, the Canadian Journal of Statistics, and Statistics and its Interface, as well as guest editor of Statistics and its Interface. Additionally, Dr. Kong is a member of the Executive Committee of the Western North American Region of the International Biometric Society, chair of the ASA Statistical Computing Session program, and chair of the webinar committee. He served as a guest editor of Canadian Journal of Statistics, associate editor of  International Journal of Imaging Systems and Technology, guest associate editor of Frontiers of Neurosciences, chair of the ASA Statistical Imaging Session, and member of the Statistics Society of Canada's Board of Directors. He is interested in the analysis of high-dimensional and neuroimaging data, statistical machine learning, robust statistics and quantile regression, as well as artificial intelligence for smart health. 

https://www.youtube.com/watch?v=6yCqjuuj90g&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=25

Abstract:
Building and maintaining state to learn policies and value functions is critical for deploying reinforcement learning (RL) agents in the real world. Recurrent neural networks (RNNs) have become a key point of interest for the state-building problem, and several large-scale reinforcement learning agents incorporate recurrent networks. While RNNs have become a mainstay in many RL applications, many choices are often under-reported and contain critical implementation details to improve performance. In this talk, we discuss one axis on which RNN architectures can be (and have been) modified for use in RL. Specifically, we will investigate how action information incorporates into the state update function of a recurrent cell. While action as the focus presents as an intuitive choice, several lines of research in cognitive science highlight the importance of action in perception. We will discuss several architectural choices centered on action and empirically evaluate the resulting architectures on a set of illustrative domains. This empirical evaluation includes an analysis of the learned state in a prediction problem, behavioral experiments, and performance when observations take the form of images and agent-centric sensor readings. Finally, we will discuss future work in developing and analyzing recurrent cells and key challenges needing attention in the partially observable setting. 

Presenter Bio:
Matthew Schlegel  is a PhD candidate at University of Alberta in Edmonton, Alberta, Canada. I have a BS in Physics and an MS in Computer Science both from Indiana University Bloomington. My current PhD work is focused on reinforcement learning, and specifically in understanding how agents may perceive their world. I focus primarily on prediction making, but have been known to dabble in control from time-to-time. My active research interests include: predictions as a component in intelligence (both artificial and biological), off-policy prediction and policy evaluation, deep learning and resulting learned representations in the reinforcement learning context, and discovery or attention of important abstractions (described as predictions) through interaction. 

YouTube Video Coming Soon!

Abstract:
Decentralized consensus learning has been hugely successful, which minimizes a finite sum of expected objective functions over a network of nodes. However, the local communication across neighboring nodes in the network may lead to the leakage of private information. To address this challenge, we propose a general differentially private (DP) learning framework for decentralized data that applies to many non-smooth learning problems. We show that the proposed algorithm retains the performance guarantee in terms of stability, generalization, and finite sample performance. We investigate the impact of local privacy-preserving computation on the global DP guarantee. Further, we extend the discussion by adopting a new class of noise-adding DP mechanisms based on generalized Gaussian distributions to improve the utility-privacy trade-offs. Our numerical results demonstrate the effectiveness of our algorithm and its better performance over the state-of-the-art baseline methods in various decentralized settings. 

Presenter Bio:
Dr. Bei Jiang is an Associate Professor at the Department of Mathematical and Statistical Sciences of the University of Alberta and Fellow of the Alberta Machine Intelligence Institute (Amii). She received her PhD in Biostatistics in 2014 from University of Michigan. Prior to joining the University of Alberta in 2015 as an Assistant Professor, she was a postdoctoral researcher at the Department of Biostatistics at the Columbia University from 2014 to 2015. Her main research interests focus on statistical integration of multi-source and multi-modal data, and statistical disclosure control and learning methods for privacy protection. She has also worked closely with collaborators in women’s health, mental health, neurology, and industry partners to apply cutting-edge statistical learning methods to real-world applications. 

YouTube Video Coming Soon!

Abstract:
From 2012 through 2021, a computer program that I wrote called Dr.Fill participated in the American Crossword Puzzle Tournament, the premier crossword solving event in the world. In 2021, it won, thereby adding crosswords to the list of games at which computers outperform humans. This (primarily) nontechnical talk describes what this journey taught me – about search, about machine learning and large language models, about crosswords themselves, and about managing human interactions while displacing human champions at this very human pastime. And also a little bit about Jonathan Schaeffer. 

Presenter Bio:
Matt Ginsberg got a doctorate in astrophysics from Oxford when he was 24. He quickly came to his senses, however, switching to artificial intelligence and teaching at Stanford for a decade. He’s been on the front page of the New York Times and, surprisingly, was happy about it. He’s been a political columnist and published playwright, and constructs crosswords for the New York Times. He has written about a hundred technical papers, and one novel. 

https://www.youtube.com/watch?v=xGEgAIEoJgs&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=22 

Abstract:
Infectious diseases are a leading cause of morbidity and mortality worldwide. One persistent challenge in the mitigation of infectious diseases is in the ability to accurately diagnose the etiology of infections using standard clinical diagnostics. Metagenomic next-generation sequencing (mNGS) has transformed disease surveillance by enabling the rapid, unbiased detection and identification of microbes without pathogen-specific reagents, culturing, or a priori knowledge of the microbial landscape. However, mNGS data analysis requires a series of computationally intensive processing steps to accurately determine the microbial composition of a sample and downstream processing to translate the multidimensional data into actionable information. In this talk, I will present work by our team to develop tools and facilitate trainings that enable researchers around the world to analyze mNGS data, gaining further insight into the microbial composition of diverse sample types. I will then discuss how these tools have been applied alongside a variety of machine learning techniques to improve the diagnosis of infections in two distinct patient cohorts. In an initial study, these tools were applied to a cohort of 92 adults with acute respiratory failure due to infectious and non-infectious causes and the development of logistic regression models enabled improved diagnosis of lower respiratory tract infection. In a recent follow-up study, support vector machines were trained to classify patients with and without sepsis amongst a heterogeneous cohort of critically ill patients. Altogether, this talk will highlight the value of mNGS for diagnosis of infectious diseases, the tools that underlie the ability to develop, and considerations in early development of diagnostic tools. 

Presenter Bio:
Katrina Kalantar is currently a Computational Biologist at the Chan Zuckerberg Initiative. Her work focuses on development of tools for analysis of metagenomic next-generation sequencing (mNGS) in the context of infectious diseases. She is deeply involved in capacity building efforts aimed at expanding and supporting researchers’ ability to perform NGS and data analysis globally. Her previous research focused on applying mNGS to improve diagnosis of lower respiratory tract infections and she continues to be involved in research through collaboration with scientists working on translational applications of mNGS technology. 

https://www.youtube.com/watch?v=NlGW1fy_A-c&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=18 

Abstract:
Over the past decade, significant interest has developed for using machine learning in medical devices for assisting in the diagnosis and risk prediction of diseases. Many publications have demonstrated promising preliminary results although the path of deploying a predictive model in the clinic is challenging from a technological and regulatory perspective. This presentation will discuss Nanostics’ work developing ClarityDX Prostate, a medical device which predicts clinically significant prostate cancer, including the challenges of obtaining FDA approval. Training predictive models on large, diverse, clinical datasets is vital for developing trustworthy models although obtaining such datasets can be problematic. Potential solutions for working with large clinical datasets will be discussed. 

Presenter Bio:
Robert Paproski is the co-founder and Chief Technology Officer of Nanostics which develops medical devices using machine learning models. Since earning his B.Sc. in Pharmacology and Ph.D. in Oncology at the University of Alberta, Dr. Paproski has developed expertise in laboratory assay development and computational analysis. Within Nanostics, Dr. Paproski oversees machine learning experiments, software development, and regulatory compliance for software products. 

https://www.youtube.com/watch?v=wqmL3rJHWUw&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=21 

Abstract:
Modern machine learning techniques have been shown to successfully encode / decode linguistic information from brain signals. It therefore seems a natural next step to use neurolinguistic data in ML models as an additional input stream to many NLP tasks. In the domain of vision, it has been shown that by forcing models to predict neural data (as well as the learned similarity representations from brain signals), models can become more robust and make “better / more natural mistakes”. This begs the question whether this effect transfers over to the domain of language / NLP with human data. This additional input stream provides many desirable properties, as the modelling process is less susceptible to idiosyncrasies of a single input modality (e.g. covariate shift, adversarial examples and non-robustness). In this talk I will recount my journey tackling these issues during my PhD, using models to decode linguistic and semi-linguistic information from single-trial EEG data, touching on various training methods that I have shown boost performance over directly training on single-trial EEG data. This work forms the core of my recent paper at ACL 2022 entitled, “Decoding PoS from Human EEG”. A key theme present throughout this talk is the effect of confounding that arises when working with linguistic data, both in terms of the linguistic status of stimuli (that also generate strong neural responses), but also biological confounds such as eye-movements, whose interference can be even further pronounced in EEG data. I will then summarize some of the main issues I believe still face us and reflect on how these might be surmounted by utilizing recent developments in multimodal Transformer networks and prediction-based representation learning in Reinforcement Learning. These techniques afford us a bridge in which both cognitive neuroscience and machine learning can jointly benefit from advances in the intersection of both domains. 

Presenter Bio:
Alex has recently completed his PhD at the University of Birmingham (UK) in Cognitive Neuroscience & NLP. His interests span many varied topics across (neuro)linguistics, cognitive neuroscience, machine learning and natural language processing. He holds a Bachelor's degree in Linguistics (Bangor University, Wales), alongside Masters degrees in Language Technology (University of Iceland) and IT & Cognition (University of Copenhagen, Denmark). During his PhD he was a Technical Intern on the Language Team at Google Brain (London) where he worked on analyzing brain data with state-of-the-art deep learning models. He is interested in how (human) neural data can be incorporated into natural language processing applications and more widely in other domains of artificial intelligence. 

https://www.youtube.com/watch?v=aGdgKq-9V68&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=16 

Abstract:
The background for the presentation is described in this YouTube video https://youtu.be/9m5kWBayClU . The August 5th presentation substitutes for the September 24th event described in the video. The draft PowerPoint is at https://www.slideshare.net/ksolez/kim-solez-how-ai-can-improve-human-cooperation-through-suggesting-followup-action-for-modelled-newscasts3pptx  Recently there has been concern expressed about the safety of machine learning/artificial intelligence in the long run by Eliezer Yudkowsky AGI Ruin: A List of Lethalities https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities widely quoted elsewhere. The first of the 40+ bolded sections of the piece is the most significant because of the way it is equally true of an AI utopia: “AGI will not be upper-bounded by human ability or human learning speed.  Things much smarter than humans would be able to learn from less evidence than humans require” Reading the beginning of that paragraph it is true that AI has taught humans much more beautiful moves in the game of Go than we would ever have been able to design ourselves. That means that AI can teach us ways of cooperating with each other that are superior to human’s own innate ability to cooperate successfully. Therefore, increasing use of machine learning is not the beginning of the slippery slope toward humanity’s demise, it could be exactly the opposite, a transition toward a world better than anything we ever imagined. We can determine which of these two contrasting futures happens, and AI can assist with that!  An AI may also be more foresighted and have a longer temporal horizon, both of which promote cooperation. 

Presenter Bio:
Kim Solez, MD is Professor of Pathology in the Faculty of Medicine and Dentistry at the University of Alberta, and Chair of the Regenerative Medicine Community of Practice in the American Society of Transplantation. He is an American pathologist and co-founder of the Banff Classification, the first standardized international classification for renal allograft biopsies. He is also the founder of the Banff Foundation for Allograft Pathology.

Kim Solez obtained his M.D. with AOA honours from the University of Rochester School of Medicine and Dentistry, and trained in pathology at Johns Hopkins Medical Institutions in Baltimore, Maryland where he was mentored in renal pathology by Robert Heptinstall. He joined the faculty at Johns Hopkins and in 1987 became chairman of the Department of Pathology at the University of Alberta in Edmonton, Canada. In 1991, he established the Banff Classification, the first standardized, international classification for renal allograft biopsies, with Johns Hopkins pathologist Lorraine Racusen. The Banff Classification, updated in regular intervals, continues to "set standards worldwide for how biopsies from kidney and other solid organ transplants are interpreted". He is the author of over 230 journal articles. He was awarded:

·       National Kidney Foundation International Distinguished Medal, 2009. 

·       University of Alberta Faculty of Medicine and Dentistry Tier 1 Clinical Mentoring Award, 2016. 

·       Catalan Society for Transplantation Gold Medal, 2017 

·       American Society of Transplantation Fellowship, FAST designation 2020.

In 2011 Kim Solez pioneered a unique graduate level medical course Technology and the Future of Medicine at the University of Alberta with strong AI content LABMP 590. Computing Science Professors Rich Sutton, Osmar Zaiane, Patrick Pilarski, and Russ Greiner teach in the course. Doctor Kim's The Future and All That Jazz musical and poetic spinoff of the LABMP 590 course with singer Mallory Chipman had its first product release on May 15, 2022. https://www.youtube.com/watch?v=QturFFBynhA Heart Drive – My Ex Was Made of Flesh. Rich Sutton attended a 2016 Future and All That Jazz event on the same day as the great AlphaGo success! Dr. Solez and Dr. Sutton also attended the Strathearn Fall Festival together in 2015 where Kim Solez performed AI related poetry https://www.youtube.com/watch?v=jTVO74FW7dc . 

https://www.youtube.com/watch?v=oNXkseddpuQ&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=15

Abstract:
Computational lexical semantics refers to the application of knowledge about word meaning to natural language processing. In this talk, we demonstrate that tasks and resources in lexical semantics benefit from theoretical analysis. First, we argue that three tasks commonly used to evaluate semantically-aware methods and models are equivalent, enabling data and methods developed for one task to be applied to the others. Second, we explore the linguistic phenomena underlying multilingual semantic resources, and posit that lexicalized concepts are universal, and thus can be annotated cross-linguistically in parallel corpora. Third, we demonstrate the utility of this theoretical analysis, by showing that it can be applied to automatically generate multilingual data for word sense disambiguation, a key semantic task. Taken together, our work makes the case that theoretical models which humans can understand, interpret, and test, are vital to semantics research.

Presenter Bio:
Bradley Hauer is a PhD candidate at the University of Alberta. He has published more than 20 papers in refereed venues, earning the NAACL 2013 Best Student Paper Award, and a nomination for the IEEE ICSC 2019 Best Paper Award. His work covers a variety of subjects in natural language processing, including multilingual semantics, classical decipherment, and cognate identification. His present research is motivated by a strong interest in theoretical models and explainable methods.

https://www.youtube.com/watch?v=g3Ojdaq7Y3o&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=12

Abstract:
Conversational agents (CAs) such as Amazon Alexa, Google Assistant, and Apple Siri have become popular in recent years. Conversational agents are computer systems designed to have natural conversations with human users, either for casual chatting or to provide the user with relevant information for a specific task. In this project, in partnership with the Mood Disorders Society of Canada (MDSC), the Asia-Pacific Economic Cooperation (APEC), Digital Hub for Mental Health, the University of Alberta, University of Saskatchewan, Dalhousie University, and the University of Alberta’s AI4Society Signature Area, we developed a mental health conversational agent (MIRA), aiming to assist healthcare workers and their families in finding relevant mental health information and local services in the Canadian provinces of Alberta and Nova Scotia. You can find the chatbot via mymira.ca.

Presenter Bio:
I am an M.Sc. student at the University of Alberta in computing science under the supervision of Osmar R. Zaiane. Also, I am an ML Engineer, Data Scientist, NLP researcher, and member of Amii (Alberta Machine Intelligence Institute) with 3+ years of research and industrial experience in linguistics, sentiment analysis, intent detection, entity extraction, and text classification. Moreover, I developed and implemented a mental health chatbot (mymira.ca) through my M.Sc. thesis using ML and NLP techniques to support healthcare workers affected by mental health issues. Through developing the mental health chatbot, I showed strong teamwork skills by collaborating with healthcare workers and computer science students. Also, I have advanced knowledge of software engineering with 3+ years of experience in developing different websites.v

https://www.youtube.com/watch?v=hzURHdMxf-Y&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=11

Abstract:
Dustin's thesis develops foundations for the development of dependable, scalable reinforcement learning algorithms with strong connections to game theory. A key contribution is a proposal for a rationality objective for reinforcement learning that is grounded in the learner's experience and is connected with the rationality concepts of optimality and equilibrium. This notion of "hindsight rationality" is based on regret, a well-known concept for evaluating a sequence of decisions with unilateral deviations, and it demands resiliency to uncertainty, environmental changes, and adversarial pressures. In this talk, he will describe how particular natural sets of deviations can be constructed specifically for sequential decision-making settings to overcome computational challenges. Dustin shows how the strategic strength of special low-complexity deviation sets can be elevated with observable sequential rationality. He then presents a unifying algorithm: extensive-form regret minimization (EFR), which achieves observable sequential hindsight rationality for a broad and natural class of deviations. The EFR algorithm often performs better in practice when it uses stronger deviation types and EFR inherits the extensibility of the counterfactual regret minimization (CFR) algorithm. This talk outlines how his thesis provides the conceptual, theoretical, and algorithmic bases for practical research directions toward the advancement of both single and multi-agent reinforcement learning.

Presenter Bio:
Dustin is a research scientist at Sony AI and a Ph.D. candidate at the University of Alberta and the Alberta Machine Intelligence Institute (Amii), co-supervised by Professor Michael Bowling and Professor Amy Greenwald of Brown University. He works on multi-agent reinforcement learning and scaleable, dependable learning algorithms. He is currently working on the GT Sophy project at Sony AI, extending and improving algorithms that learn to outpace expert e-sports racers in a realistic racing simulator. He is also a coauthor of DeepStack, an expert-level player of heads-up, no-limit, Texas hold'em poker, and he created a public match interface for Cepheus, a near solution to heads-up, limit, Texas hold'em. He completed a B.Sc. and M.Sc. in computing science at the University of Alberta, also supervised by Michael Bowling. As an undergraduate, he worked with the Computer Poker Research Group at the University of Alberta to create an open-source web interface to play against poker bots and to develop the 1st-place 3-player Kuhn poker entry in the 2014 Annual Computer Poker Competition.

https://www.youtube.com/watch?v=KbgZ5_R1Sb0&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=9

Abstract:
Gradient descent is slow to converge for ill-conditioned problems and non-convex problems. An important technique for acceleration is step-size adaptation. In this talk, I'll review the efforts in attacking this problem in decades. Interested algorithms covered are Polyak step-size and IDBD from classical optimization and the first wave of neural networks, and Adam, Hypergradient descent, L4 and LossGrad from recent deep learning. We will also review the connection of step-size adaptation to meta learning. We will discuss the projection power of diagonal-matrix step-size, and show using negative step-sizes can lead to faster convergence even for deterministic gradient descent. In the end, we will discuss the possibility of applying the idea of Dyna-style planning for step-adaptation, for which a new algorithm solves the famous Rosenbrock function in under 500 gradient evaluations with zero error, while gradient descent needs 10000+ evaluations to reach an accuracy of 10^{-3}.

Presenter Bio:
Hengshuai Yao's research interest is model-based reinforcement learning with a recent focus on gradient descent especially step-size adaptation. He has brought a unique connection between the two seemingly unrelated topics.

https://www.youtube.com/watch?v=zUwZ1vMn7Qw&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=7

Abstract:
Humans continually learn and adapt to new knowledge and environments throughout their lifetimes. Rarely does learning new information cause humans to catastrophically forget previous knowledge. While deep neural networks (DNNs) now rival human performance on several supervised machine perception tasks, when updated on changing data distributions, they catastrophically forget previous knowledge. Enabling DNNs to learn new information over time opens the door for new applications such as self-driving cars that adapt to seasonal changes or smartphones that adapt to changing user preferences. In this talk, we propose new methods and experimental paradigms for efficiently training continual DNNs without forgetting. We then apply these methods to several visual perception tasks.

Presenter Bio:
Tyler Hayes recently defended her PhD in the Chester F. Carlson Center for Imaging Science at the Rochester Institute of Technology (RIT) in Rochester, NY. During her PhD, she worked with Dr. Christopher Kanan to advance online continual learning. Her current research interests include lifelong machine learning, computer vision, and computational mathematics. Previously, she earned a BS in Applied Mathematics from RIT in 2014 and an MS in Applied and Computational Mathematics from RIT in 2017. She has published in a wide range of venues in AI, including CVPR, ICRA, AAAI, ECCV, BMVC, and more. Her website can be found here: https://tyler-hayes.github.io/

https://www.youtube.com/watch?v=koBeei2T350&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=4

Abstract:
Trust Science ® is a FinTech SaaS committed to using cutting-edge AI and ML technology to advance alternative credit underwriting for a fairer, more inclusive, financial future. FinTech for Good. We are a talented team of self-starters that thrive in a dynamic environment and are passionate about making the world more financially inclusive, one credit score at a time.

The presentation will discuss:

• Types of Modeling:

• Types of Models

• Example

• Engineering at Trust Science:

• Engineering Teams

• A Day in Engineering

• AI Job Types

Presenter Bios:
Mat Lavoie - VP Data Engineering - Mat is a University of Alberta Computer Science graduate, who has worked at a number of Edmonton based companies, growing from coding to executive roles including Founder and CTO. From an AI/ML perspective, Mat has worked to predict sporting event outcomes, learning systems (k-12) and with Trust Science is responsible for Data Engineering. As VP of Data Engineering, Mat is responsible for modeling, data science, initial analysis and data sourcing.

Lenora Thomas - VP of Customer Operations - Lenora is an accomplished professional with over 19 years of experience in various executive positions in high-technology/Software, with strategic leadership roles in customer success, support and operations.  As VP of Customer Operations at Trust Science, Lenora thrives in the challenges that come with planning, organizing, implementing and supporting customers as they embrace advanced solutions that fundamentally improve their businesses’ workflow and profitability.

https://www.youtube.com/watch?v=rvc7k208ssw&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=2

Abstract:Trust and reputation systems constitute an active branch of research in multi-agent systems. In various application domains, agents interact with one another to collect information, goods, or services that help complete a set task. For such interactions to be largely successful, agents try to estimate how trustworthy other individual agents are. To evaluate the trustworthiness of an agent in a multi-agent system, one often combines two types of trust information: direct trust information derived from one's interactions with that agent, and indirect trust information based on advice from other agents.

In this work, we first focus on the trust established through indirect trust information. This is a non-trivial problem, since advice may not always be reliable; it may come from a deceptive agent whose goal is to mislead the truster. We propose a new and easy-to-implement method for computing indirect trust, based on a simple prediction from expert advice strategy as is often used in online learning. This method either competes with or outperforms all tested systems in most of the simulated settings while scaling substantially better. We also provide the first systematic study on when it is beneficial to combine the two types of trust as opposed to relying on only one of them. Our large-scale experimental study shows that strong methods for computing indirect trust make direct trust redundant in a surprisingly wide variety of scenarios. Further, a new method for the combination of the two trust types is proposed that, in the remaining scenarios, outperforms the ones known from the literature. Moreover, we propose a method based on the Page-Hinkley statistics to handle the dynamic behaviour of agents in a multi-agent system.

Presenter Bios: Elham Parhizkar is a postdoctoral fellow at the University of Alberta, where she is working under the supervision of Dr. Levi Lelis. Currently, she is focused on developing algorithms for synthesizing effective and interpretable strategies for solving problems and playing games.

She completed her Ph.D. from the University of Regina under the supervision of Dr. Sandra Zilles in 2021. Her Ph.D. research focused on establishing trust between agents in multi-agent systems. 

https://www.youtube.com/watch?v=sn-sJppDlw4&list=PLKlhhkvvU8-bFpoIPFTCVsYmqFVpt7OTz&index=1

Abstract:
We design a simple reinforcement learning (RL) agent that implements an optimistic version of Q-learning and establish through regret analysis that this agent can operate with some level of competence in any environment. While we leverage concepts from the literature on provably efficient RL, we consider a general agent-environment interface and provide a novel agent design and analysis. This level of generality positions our results to inform the design of future agents for operation in complex real environments. We establish that, as time progresses, our agent performs competitively relative to policies that require longer times to evaluate. The time it takes to approach asymptotic performance is polynomial in the complexity of the agent's state representation and the time required to evaluate the best policy that the agent can represent. Notably, there is no dependence on the complexity of the environment. The ultimate per-period performance loss of the agent is bounded by a constant multiple of a measure of distortion introduced by the agent's state representation. This work is the first to establish that an algorithm approaches this asymptotic condition within a tractable time frame.
Presenter Bios: 

Shi Dong is currently a Ph.D. candidate in the Department of Electrical Engineering at Stanford University, where he is advised by Prof. Benjamin Van Roy.  He is interested in using theoretical tools to understand the essential elements in practical reinforcement learning (RL) agent design, and to help bring the benefits of RL to real life.  He received his bachelor’s degree from Tsinghua University, and master’s degree from the Department of Statistics at Stanford University.  His industrial experiences include assisting ByteDance in improving their news and video recommendation system, and research internships at Google, DeepMind, and Microsoft.  One of his works was awarded winner in the 2021 INFORMS George Nicholson Student Paper Competition.

https://www.youtube.com/watch?v=V_Oyl5cx4Mc

Abstract: This seminar will highlight many of the interesting applications of Machine Learning in use and being explored at Electronic Arts, a leading gaming publisher. Just as machine learning is transforming many industries, the gaming industry is finding exciting ways to address many of the unique challenges that game software development presents, and EA is making great investments into the space. This seminar will also provide insight into EA’s commitment to attracting talent and should answer many questions about how one might join the organization as an AI/ML developer or scientist.

Presenter Bios:

William Gordon (Bill) is a senior software engineer with Electronic Arts (EA), working in AI and rendering. He currently works on a DTS research and applications team conducting practical machine learning applications for advancing gameplay AI and game validation for all EA games. Bill’s currently is working on ML projects that include: Object-Detection, Object tracking, Image Classification, Image segmentation, Imitation Learning (IL), Reinforcement Learning (RL) and Natural Language Processing (NLP). Before joining EA, Bill worked at Disney and other mobile start-ups in a variety of low-level embedded software and mobile app development roles. Bill has a B.S.E.E. from Purdue University. He is a Disney Inventor and has patents pending with EA. In his free time, Bill loves spending time at the beach with his wife and kids in sunny Florida.

Alex Lucas is the Director of Quality Verification for BioWare Studios. Although a Calgary native, he’s been with BioWare in Edmonton for the past 18 years, first as a software engineer and later as both a technical and department director. Alex has for many years been an advocate for leveraging relationships with academia and research (most notably with the University of Alberta and the NRCC) and has led many partnership engagements in the ML space. He has a BSc in Computer Science from the University of Calgary and several patents with EA. He is a new, first-time dad, and very tired.

No recording available

Abstract: We take lots of pictures everyday with our mobile phones where the images are 2D arrays storing the color values. But we live in a 3D world and the ability to reason about the 3D properties is the basic technique for accomplishing tasks such as navigation, object manipulation and scene understanding, etc. In this talk, she will present some of her recent works on designing low-cost and convenient 3D modeling systems to build high quality 3D models such as human avatars, by exploiting machine learning techniques. And stepping forward from the modeling problem, she will also share some interesting works on model animation.

Presenter Bios: Xinxin Zuo is currently a Postdoctoral Fellow at the Vision and Learning Lab, University of Alberta. She received her Ph.D degree from the Department of Computer Sciences, University of Kentucky, in 2019. Before that, she received B.Eng. and M.Eng. degrees from the School of Computer Science, Northwestern Polytechnical University.  Her research interests include Machine Learning, Computer Vision and Computer Graphics, in particular in 3D vision. She has published 20+ papers on related areas, including top journals such as TPAMI/IJCV/TMM and top conferences like CVPR/ICCV/ECCV.

https://www.youtube.com/watch?v=L8BgNLzhPjE&list=PLKlhhkvvU8-ZYAernGzP2ZKMz1RILmXd6&index=3

Abstract: In recent years, a growing number of deep model-based reinforcement learning (RL) methods have been introduced. The interest in deep model-based RL is not surprising, given their many potential benefits, such as higher sample efficiency and the potential for fast adaption to changes in the environment. However, we demonstrate, using an improved version of the recently introduced Local Change Adaptation (LoCA) setup, that the well-known model-based methods such as PlaNet and DreamerV2 perform poorly in their ability to adapt to local environmental changes.  Combined with prior work that made a similar observation about the other popular model-based method, MuZero, a trend appears to emerge suggesting that current deep model-based methods have serious limitations. We dive deeper into the causes of this poor performance, by identifying elements that hurt adaptive behavior and linking these to underlying techniques frequently used in deep model-based RL. We empirically validate these insights in the case of linear function approximation by demonstrating that a modified version of linear Dyna achieves effective adaptation to local changes. Furthermore, we provide detailed insights into the challenges of building an adaptive non-linear model-based method, by experimenting with a non-linear version of Dyna.

Presenter Bio:
Yi Wan is a fifth-year Ph.D. candidate in Computing Science at the University of Alberta, focusing on reinforcement learning, which he believes is the most promising way to artificial general intelligence. His Ph.D. supervisor is Professor Rich Sutton. His long-term research goal is to build simple, general, and scalable learning and planning algorithms for reinforcement learning problems. He is particularly interested in designing these algorithms 1) for the average reward problem setting, 2) with function approximation, and 3) with temporal abstractions. Previously, he earned a Bachelor degree in Electrical and Computer Engineering (ECE) from Shanghai Jiao Tong University (SJTU), where he worked in SJTU Speech Lab, advised by Professor Kai Yu.  After that, he obtained a master degree, also in ECE, from University of Michigan. During his master, he worked in Intelligent Robotics Lab, advised by Professor Ben Kuipers. During his master and Ph.D., he interned as a researcher in Mila, Huawei, and Tusimple and also interned as an engineer in Yitu.

https://www.youtube.com/watch?v=r6o05gOOtpg

Abstract: We receive information about the world through our sensors and influence the world through our effectors. Such low-level experiential data has gradually come to play a greater role in AI during its 70-year history. I see this as occurring in four steps, two of which are mostly past and two of which are in progress or yet to come. The first step was to view AI as the design of agents which interact with the world and thereby have sensorimotor experience; this viewpoint became prominent in the 1980s and 1990s. The second step was to view the goal of intelligence in terms of experience, as in the reward signal of optimal control and reinforcement learning. The reward formulation of goals is now widely used but rarely loved. Many would prefer to express goals in non-experiential terms, such as reaching a destination or benefiting humanity, but settle for reward because, as an experiential signal, reward is directly available to the agent without human assistance or interpretation. This is the pattern that we see in all four steps. Initially a non-experiential approach seems more intuitive, is preferred and tried, but ultimately proves a limitation on scaling; the experiential approach is more suited to learning and scaling with computational resources. The third step in the increasing role of experience in AI concerns the agent’s representation of the world’s state. Classically, the state of the world is represented in objective terms external to the agent, such as “the grass is wet” and “the car is ten meters in front of me”, or with probability distributions over world states such as in POMDPs and other Bayesian approaches. Alternatively, the state of the world can be represented experientially in terms of summaries of past experience (e.g., the last four Atari video frames input to DQN) or predictions of future experience (e.g., successor representations). The fourth step is potentially the biggest: world knowledge. Classically, world knowledge has always been expressed in terms far from experience, and this has limited its ability to be learned and maintained. Today we are seeing more calls for knowledge to be predictive and grounded in experience. After reviewing the history and prospects of the four steps, I propose a minimal architecture for an intelligent agent that is entirely grounded in experience.

Presenter Bio:
Richard S. Sutton is a distinguished research scientist at DeepMind, a professor in the Department of Computing Science at the University of Alberta, and a fellow of the Royal Society (UK), the Royal Society of Canada, the Association for the Advancement of Artificial Intelligence, the Alberta Machine Intelligence Institute (Amii), and CIFAR. Sutton received a PhD in computer science from the University of Massachusetts in 1984 and a BA in psychology from Stanford University in 1978. Prior to joining the University of Alberta in 2003, he worked in industry at AT&T Labs and GTE Labs, and in academia at the University of Massachusetts.  In Alberta, Sutton founded the Reinforcement Learning and Artificial Intelligence Lab, which now consists of ten principal investigators and about 100 people altogether. He joined DeepMind in 2017 to co-found their first satellite research lab, in Alberta. Sutton is co-author of the textbook Reinforcement Learning: An Introduction from MIT Press. His research interests center on the learning problems facing a decision-maker interacting with its environment, which he sees as central to intelligence. He has additional interests in animal learning psychology, in connectionist networks, and generally in systems that continually improve their representations and models of the world. His scientific publications have been cited more than 100,000 times. He is also a libertarian, a chess player, and a cancer survivor.

No recording available

Abstract: Samdesk is an Edmonton based start up that uses AI to detect and monitor global crises events in real time. Our system ingests a range of social media and traditional web sources, and the scope of problems that our AI tackles includes text classification, source trustworthiness estimation, input geo-localization, text similarity, and clustering. In this talk, we present how Samdesk helps our customers act smarter and faster, we examine some of the main ML and AI systems within samdesk, and we discuss some practical lessons that we have picked up in terms of gathering and processing data and applying AI in production.

Presenter Bios:
Sky Sun finished her MA in Digital Humanities at the U of A and is working full time at samdesk as Data Intel Team Lead. Her role at samdesk is to enrich and manage data to support AI performance. She is passionate about taking a humanistic and interdisciplinary approach to solving technical problems.
Adam St Arnaud received his MSc in Computing Science at the U of A in 2017 under the supervision of Greg Kondrak. Adam is a Lead Machine Learning Engineer at samdesk, where he has been working since 2017.