Machine Learning

Making Data Meaningful

Whether it be for a community project to determine which water is safe for drinking, or a hospital-based project to determine the most efficient and effective ways to treat mental illness, Trishla Shah is an expert in taking the data we collect so easily today, and making them meaningful.

As the IT research scientist at the Nova Scotia Community College and a PhD candidate in computer science at Dalhousie University, Shah focuses on designing solutions by analyzing massive amounts of data. For example, Nova Scotia’s water comes from various places, including freshwater lakes, but also possibly from contaminated lakes, she says.

“We want to educate the organizations taking care of our water resources so we can see how we can preserve the well water and how we can make sure the water from the fresh lakes is not contaminated,” Shah says, adding that many of her projects come to her after receiving funding from a national social innovation fund.

Private IT companies that want to design and test a product from scratch also often approach the research scientist.

“We are being approached by startups, by midsize companies and by large-scale companies,” Shah says. “We are being approached by organizations that take care of Indigenous communities. One project we’re working on involves artificial intelligence, Indigenous communities and Halifax’s storied Bluenose ship.”

For another project, Shah is helping small- and medium-sized enterprises identify the right clean technology investments for their business. Her lab is building a data repository to provide solutions to scale access to low-carbon emissions for businesses in the agri-food, construction and manufacturing industries. The repository is integrated from various sources using web-scraping and text-mining techniques. The parameters of the data repository will be set to allow an algorithm to forecast carbon emission reductions and analyze the return-on-investment of various interventions.

Shah and her team get involved from the beginning, first collecting the data, then building a database, preprocessing, storing and managing the data in a way that ensures they’re secure and making sure there’s no breach. From there, they can make models of the data so they don’t have to analyze them manually.

Shah hasn’t used the services of ACENET directly, but she has circulated its many programs among the information technology students at the college.

“We have a lot of academic programs that have IT in them,” she says. “And students might want extra practice, or students might need to have something like a prerequisite. With ACENET, the courses are really applicable and I like to circulate them among students who need additional practice on that topic. So that’s my interest with ACENET. I’m a fan.”

In addition, Shah is in the process of designing her own student training module that has an online portal where students can enrol themselves in training that is focused on real-time projects.

“I have proposed to the committee that’s approving my student training module and funding that ACENET’s courses be prerequisites.”

Artificial Intelligence to Improve Human-System Interaction

While debates rage over the ethics of artificial intelligence’s role in art or helping students cheat on term papers, Sid Ahmed Selouani has quietly worked with AI to develop industrial applications and improve people’s quality of life.

He and his lab at the University of Moncton, Shippagan Campus – the Human-System Interaction Research Laboratory (Laboratoire de Recherche en Interaction Humain-Système, LARIHS) – use AI to advance human-system interaction, working on subjects as diverse as emotion, pathological speech and language recognition, telepathology automation and intelligent automation for industrial robots.

Machine learning in cyber-physical systems, which integrate physical and digital components, relies on complex algorithms to make decisions and commit to actions in the real world.

“For example,” explains Selouani, “to help someone with difficulty pronouncing words, like someone who has had a stroke, and can only say bits of words, we use algorithms to understand how the person expresses themself, then replace broken words or segments to generate understandable text.” This is the objective of a company built on LARIHS’s research: to commercialize software that provides people with speech disorders an interpreter to help them communicate. The algorithm analyzes speech, replaces unclear sections with text, then generates a sound similar to a person’s voice. The software will also help doctors to monitor patients throughout their rehabilitation.

For an algorithm developed by LARIHS to discern what a person is trying to say, however, it first needs experience listening to human speech. That requires lots of data, analyses, data mining, and comparison, often involving millions of parameters. It takes top-of-the-line supercomputer systems, like those provided by ACENET and the Digital Research Alliance of Canada, to give algorithms the processing power and storage capacity to make the necessary calculations and learn how to interact with the real world. “Our computing capacity has increased at least 10-fold with ACENET,” Selouani explains.

Such processing power opens many doors. Selouani and his team of more than two dozen post-docs, grads, undergrads and research personnel have many projects on the burner, each with their own sub-projects.

For instance, LARIHS is working on a project to teach industrial robots to sort oysters, identifying live oysters from dead ones, and grouping them based on size and overall quality. LARIHS is also working on optimizing telepathology – that is, the classification of human tissues and biological samples at a distance. By processing the images extracted from biopsy analyses, for example, we will be able to better detect abnormalities in the scanned images, thus helping doctors in their assessments.

The applications for these algorithms are as varied as their fields of study. How do you fit all that in one lab? ”Our know-how,” Selouani explains, “which we have developed over the last 20 years, is that we can take an algorithm we have developed for one application and adapt it to other ones.”

Machine Learning: Connecting the Brain to the Body

Xianta Jiang uses machine learning and wearable-sensor technology to study how the brain works in telling the body what to do and how to move. Currently, he’s using these techniques to solve a health-care problem for those who have prosthetic hands. But he doesn’t work in a hospital setting; rather, he works in a computer lab.

A professor of computer science at Memorial University, Jiang works with a team of students to study all of the different ways humans use their hands. With that information, they’ll use machine learning to help those who wear prosthetics control them.

“Controlling the prosthetic hand is a super difficult problem to solve, because when a human’s hand is amputated, the brain has a harder time communicating to the arm about how to maneuver the prosthetic limb,” Jiang says. “We are trying to help people control them as naturally as possible, and without surgery. To try to solve the problem, we use muscular sensors attached to a part of the arm to infer movement intentions from the area of the brain used to control the hand.”

However, it’s difficult to make the connections that allows the brain to send the right signals to the hand.

“For that, we attach a camera to the prosthetic and when the camera can identity the target, the hand can configure correspondingly — just like self-driving,” he says.

That’s the hands-on part of the research, but, in the end, much of the work is done by his students in a computer lab using the high-performance computing resources of the Digital Research Alliance of Canada and ACENET. They identify the grasp types the prosthesis user will need and use computer modelling to fine-tune the fit.

“It’s quite a basic question, but we need a lot of computing resources for it,” he says. “We are working to cover 95 per cent of daily life grasps with a total of 16 movements and need a lot of data to train this model, so we use high-performance computing.”

He said he couldn’t do his work without ACENET’s services.

“It’s not affordable to purchase the computing power we’d need,” he says. “We need a lot of memory.” In 2022 alone, Jiang’s group used 128 CPU years and 23 GPU years of compute power.

Another of Jiang’s projects involves monitoring human activity using wearable devices such as smartwatches. The goal is to find a way to allow rehabilitation staff or sports coaches to monitor their patients’ progress digitally. This work is in the early collaboration stage and not yet commercialized. “For this one, we’ll collaborate with industry over the next few years,” he says.

Persuasive Computing for Social Good

As she describes it, Rita Orji designs technology “to promote social good.”

Some of the applications the Dalhousie University computer science professor and Canada Research Chair and her students have produced promote physical health and wellness, while others seek to improve one’s mental health. Generally, these applications help people achieve desired changes in their behaviour, whether that be discouraging risky sexual practices or binge drinking, promoting physical activity and healthy diet, or improving mental health. Each application starts with Orji studying the people who might use them — often by collaborating with academics in other faculties — and then making users part of the design process.

“We design persuasive technologies,” Orji says, adding that they integrate modern technologies such as artificial intelligence and virtual reality into the apps they build to engage users and improve the apps’ effectiveness and user experience. “We need to keep the user in mind throughout the design process.”

One such app is a collaboration between Orji’s Persuasive Computing Lab in the faculty of computer science and the psychiatry department. Using information from a research study that aims to improve the mental well-being of youth, Orji and her colleague, Professor Sandra Meier, developed an app called PROSIT, which stands for “Predicting Risks and Outcomes of Social inTeractions.” The app passively collects information on the daily social interactions of youth, including how many calls they’re making, how many messages they’re sending and how often they’re using social media to interact with friends and family.

Another app Orji and her student, Oladapo Oyebode, developed is called TreeCare. The app simulates users’ physical activity to represent the growth of trees in a virtual garden. If they are physically active in the real world, their virtual trees will flourish. “If they are more sedentary or less physically active, their tree will start losing leaves, losing fruit and becoming unhealthy,” Oyebode says.

To be able to personalize the systems to individuals, Orji and her students collect behavioural and physiological data from their target audiences, and then analyze them to detect patterns or make predictions using AI techniques, including natural language processing and machine learning, all of which they do using the services of ACENET.

“For one project, we collected more than 47 million comments related to COVID-19,” Oyebode says. “We used ACENET’s resources to preprocess and analyze that data.” ACENET’s computing infrastructure helped the two to sift through the data, clean it up and prepare it for analysis, he says, adding that they performed keyphrase extraction and sentiment classification to uncover various issues related to the COVID-19 pandemic.

“ACENET has been really, really helpful for being able to crunch, process and analyze this big data,” Oyebode says.

Oyebode also added that they have used ACENET resources, such as the GPU-enabled clusters, to train advanced machine learning models that inform adaptive mental health interventions.

Making Generative AI Sound More Human

Uyen “Rachel” Lai is looking into differences between the way AI algorithms that generate text, such as ChatGPT, express themselves compared to humans. The goal is to improve the usefulness of generative AI models in producing content.

“For example, my preliminary work suggests that generative AI models produce text that is more lexically diverse, though possibly less focused,” Lai says. “I want to make generative AI models that produce more “human sounding,” natural or realistic text.”

Lai, a fourth-year computer science honours student is working under the supervision of Paul Sheridan, assistant professor at the School of Mathematical and Computational Sciences at the University of Prince Edward Island. Even as an undergraduate, Lai has travelled as far as Tokyo to present a paper on her work with Sheridan Stable — Sheridan’s professional wrestling-inspired name for his “plucky” group of undergraduate researchers — which focuses on natural language processing and text analysis. The team is trying to understand how the two worlds of classic statistics and deep learning are connected.

“Deep learning inherits a lot of ideas from classical statistical methods. We’re looking at those connections and then mapping them out, which we hope will lead to insights about how to use all the statistics knowledge the world has accumulated over the years to do deep learning in more interpretable and efficient ways.”

Lai is setting the standard in that team of researchers.

“Rachel is an exceptional student who’s on a good roll right now,” Sheridan says, “and there are lots more great students on the way who are hungry to make names for themselves.”

Lai uses ACENET resources to generate text using AI models, and she has also trained some of her fellow students on how to use them.

“We need a lot of text to compare lexical diversity in human-written and AI-generated text,” Lai says. “The compute clusters come in handy, helping us generate a massive amount of text. By using these resources, we can generate what we need in a few hours at most. Without them, it would take months or years to do our work.”

Sheridan says that as the program grows, the team will be generating vast volumes of text that will require hundreds of gigabytes which would be impractical to store on a regular computer.

“We need to use GPUs to generate text,” he says. “GPUs are in high-demand for deep learning applications.”

Lai lauds ACENET’s support services, which she says are also very responsive.

“If one person can’t help, they will find us someone else who has more expertise in the problem we are trying to solve,” she says.

Strategically Manipulating Bacteria for the Greater good

Lourdes Peña-Castillo is looking to understand bacteria to find ways to strategically manipulate them.

“Bacteria or, more generally, microbes are everywhere,” says Peña-Castillo, who is jointly appointed as a professor in the departments of computer science and biology at Memorial University. “They are in the house, they are in the soil, they are in the environment and they interact with everything. We know that some bacteria cause diseases, but they are a minority. All other bacteria are beneficial for plants, animals and for us.”

Given that, Peña-Castillo’s research, which she describes as being computational microbiology, applies machine learning to understand patterns in the genome of bacteria that signal to them how to “turn on” or express their genes.

“Right now, if we want to treat a disease, we basically take antibiotics and kill every single bacterium, the good ones along with the ones causing the disease,” says Peña-Castillo, who did her PhD in Computer Science in Germany and her postdoctoral work at the University of Toronto. “With my research, we are understanding in more detail every single bacterium, and then potentially we could actually either modify the gene expression of that bacterium or create treatments specifically designed for that bacterium. Instead of killing everything, let’s try to just control a specific part of a bacterium.”

She refers to this work as a foundational endeavour in “smart biotechnology,” noting that in industrial processes, for example, they sometimes want bacteria to help to create more of a certain substance.

Peña-Castillo says she couldn’t do her job without ACENET.

“In my lab, we work with collections of sequencing data,” she says. “Each raw uncompressed sequencing file can be tens of gigabytes (GBs). As each dataset can have several of these files, it can quickly add up to hundreds of GBs in disk space. Add to that the fact that the software used to process these data can easily require tens of GBs of random access memory (RAM), often at least 50, and most laptops only have 8 to 16 GBs of RAM, and you can see that ACENET is not only necessary but indispensable.”

She says her team could run a single experiment in a high-end computer, but in many cases, it runs dozens of experiments to optimize its models.

“Using supercomputers allows us to run these experiments in parallel,” she says, adding that in one recent project her team combined more than 20 different datasets in an effort to train its models.

“ACENET enables us to do all the computational analysis that we do,” she says. “Most of my students run their calculations on ACENET, so it basically allows us to run all the experiments and analysis in an efficient way. Without ACENET, I would have to buy a lot of very expensive computers to do my job.”

Curbing COVID with Computers

Photo byClarisse CrosetonUnsplash > When one thinks about computer science’s role in tackling COVID-19, it’s hard to imagine what it might be. Yet James Hughes is working on that very thing. The assistant professor at St. Francis Xavier University has been spending the summer of 2020 trying to figure out how authorities can most effectively deliver what is bound to start as a finite number of vaccines to have the best chance of slowing the virus. “Vaxing a population is perhaps more a programming problem than people might think,” Hughes says. “When the vaccines become available, we won’t have enough for everyone.” The World Health Organization and the Centers for Disease Control have guidelines on who to vaccinate, which they establish after considering a number of factors, including risk economics and ethics. But, Hughes is working on a more dispassionate model that represents a community as a network of connected people. “Imagine a remarkably simple network with three people,” he explains. “The middle person is connected to the other two, but the people at each end aren’t connected to each other. So if a person on the end gets COVID-19, we vaccinate the middle person, thereby saving them and the person at the other end. If that person is connected to a million other people, we just protected a million people with one vaccine.” Lots of great minds have ideas on how to curb COVID once the vaccine is available but working out the math on them will become extremely labour intensive very quickly, he says. That’s where computers come in. “We’re using artificial intelligence and machine learning to find strategies,” he says. “We’re doing simulations and the AI is generating a system of programs that tell us who to vaccinate.” “The AI isn’t biased by pre-conceived notions,” he says. After they have the strategies, humans can then apply their ethical standards and tweak the answers if necessary. He’s working with other teams —one in Guelph and one at Brock University — and is hoping to get a larger research community involved. Hughes pivoted to COVID when it hit, but his research has always used machine learning to solve real-world problems for which answers aren’t otherwise readily available. The problems can come from kinematics, geology, music, finance and the human brain. He was recently, for example, doing some modelling on people with Parkinson’s, while they were using treadmills. “We were using Compute Canada resources to come up with mathematical functions on how people walk,” he says. “I was also working towards coming up with a predictive model that will tell you what a traumatic brain-injury patient’s intercranial pressure is. The best way to test that is a dangerous neural surgery. We’re working to see if we could find another way that’s safer.” His work also always requires intensive computer resources. In 2018 alone, he used “hundreds and hundreds” of core years. If he’d done the same work on a typical desktop computer, he would have had to hit “run” and return in 200 years for any findings, he jokes.

Curbing Sea Lice in Salmon Farming

Gregor McEwan wants to help salmon farmers manage sea lice on their farms, and he’s using supercomputing to do it. It seems odd to think of research into aquatic life involving high-powered computing, but that’s what McEwan, a research scientist, is doing. He works under the supervision of Professor Crawford Revie in Revie’s lab in the Atlantic Veterinary College’s Department of Health Management at the University of Prince Edward Island. Salmon farming is an industry worth more than $16 billion worldwide, and Canada is the industry’s fourth-largest producer after Norway, Chile and Scotland. Farmed salmon is raised from eggs on land facilities and, when large enough, moved to cages in the sea. The fish are kept in the cages until they reach harvest weight — typically three to five kilograms. After that, they’re sent to processing plants to eventually become fillets available through retail. But often, in the sea cages, sea lice, among the most pervasive problems facing salmon farmers, invade. They attach to the salmon and feed off the fish’s mucus, flesh and blood, causing discomfort and reduced immunity for the salmon. “Usually, [due to immunity issues], they then get sick from something else, but enough sea lice will also kill the salmon,” McEwan says. “It’s a major issue for salmon farmers and they’re especially interested in using treatment methods they have — whether chemicals or mechanical treatments such as warm water baths — to solve it.” McEwan is trying to determine how they can use existing treatments most effectively and in what kind of patterns they should use them. “I build computer simulations to try out different strategies,” McEwan says. “It’s expensive to try them on the salmon farm in real time, so it makes sense to simulate them on computers.” To make the simulation work, he must set parameters — such as ‘how many lice are flowing on to the farm?’ — which is tricky. Such information is unknown because it’s not feasible to count lice in the open sea around each farm. “I take all the historical information I can get about the environment and the farmers’ weekly count records of sea lice on the salmon,” McEwan says. “I train the model with that information and extract the relevant parameters. I use machine-learning algorithms to get the correct parameters for the model. I can use the trend model to create predictions for the future. We can run scenarios where we say, ‘okay, what if this happens, or what if we use this kind of strategy?’” One of the things McEwan must consider is the ability of the lice to evolve a genetic resistance to the chemical treatments. For example, past research investigated the repercussions of wild salmon coming near the farm cages. Wild salmon carry lice, but McEwan’s discovered that, because their lice are genetically naïve to the chemical treatments, they bring what he calls a “genetic cleaner” to the equation. “Yes, you end up with a few more lice in the short term, but whatever treatments you’re using are more effective and work a lot longer because the genetic resistance is being cleaned out of the population,” he says. “It’s actually a net positive.” For McEwan’s simulations, ACENET’s resources are vital. His lab has powerful computers, but nothing on the scale of ACENET. He says without ACENET, everything he’s doing would take months instead of weeks. “We’d never be finished.”

Creating a Non-linear Model of our Non-linear Brains

James Hughes likes to use applied machine learning to solve real-world problems for which answers aren’t otherwise readily available. An assistant professor in St. Francis Xavier’s computer science department, his research can apply to kinematics, geology, music, finance and the human brain. He was recently, for example, doing some machine-learning modelling on people with Parkinson’s, studying them while they were using treadmills. To do his work, he uses the high-powered computing resources of Compute Canada and ACENET. “We were using Compute Canada resources to come up with mathematical functions on how people walk,” he says. “I was also working towards coming up with a predictive model for traumatic brain injury patients that will tell you what a subject’s intercranial pressure is. The best way to test that is a very dangerous neural surgery. I was working to see if we could find another way that’s safer.” He calls the human brain a non-linear computational system. But the types of algorithms and math currently used when trying to understand the human brain — those typically used in functional magnetic resonance imaging studies, for example — are done with strictly linear tools. “This means the models we end up with at best can be linear approximations of a non-linear system,” Hughes says. “This is not to say that all of this stuff done up to date has to be thrown out. We’ve been able to make a lot of big contributions. However, one can’t help but wonder what we could do and learn if we had a non-linear model of the brain that more accurately represents the underlying system.” To do his non-linear modelling, he uses “heavy-duty” machine learning, artificial intelligence and what he calls “evolutionary computation.” Think of it as designing a car, he says. “What happens if I come up with 100 random car designs?” he asks. “They will all be pretty bad, but what if I take two of the less bad ones? I’ll combine them and hopefully make a better car. If I keep doing this millions and millions of times, I will typically end up with a pretty good car design. It’s the same idea. You pick the ones that are relatively better.” He was once studying how a brain interpreted or dealt with a specific language task. He asked his subjects to figure out how words related to one another. “If you use the traditional modelling technique, it would say ‘OK, so the brain regions X, Y, Z are related to each other,’” he says. “But with non-linear models, we found that region X, Y, Z and A, B, and C are related because during the study, they started to react in a non-linear way.” “But the number of possible explanations for all of these studies are just astronomically large,” he says. And that’s the reason, he requires intensive computer resources. In 2018 alone, he used “hundreds and hundreds” of core years. If he’d done the same work on a typical desktop computer, he would have had to hit “run” on the computer and return in 200 years for any findings, he jokes.

Stimulating Startups While Educating Students

Pawan Lingras takes massive amounts of data and forms intelligent conclusions from them. He and his students at Saint Mary’s University in Halifax work with several companies — as many as 15 over the last three years — to help them take the data they collect and improve efficiency and or draw important conclusions. With Green Power Laboratories Inc, for example, the goal is energy conservation for the company’s client’s work sites, which can be anything from an office building, a manufacturing plant or a shopping mall. “The goal is to ensure comfort of the occupants, but once you ensure that, you want to optimize the energy consumption,” Lingras says. “Doing physical testing on a building with the heating and cooling plant can be expensive because you need to get data for all kinds of combinations to figure out the ultimate energy conservation.” Instead, he uses the services of ACENET, a regional partner of Compute Canada with fourteen Atlantic Canada member institutions that provides large-scale high-performance computing facilities, to simulate engineering plants to see how they behave under different conditions. “We have to use predictive analytics to come up with mathematical model to capture power plants’ behaviour,” he said. “This is all done with high-performance computing.” Another company he’s working with is called Hanatech Solutions, and he helps the company manage a large stream of data within its many applications and installations. “Storing all that data is a big problem,” he says. “Sometimes these sites are very remote and moving the data is very costly. Our goal is to reduce the costs and ultimately save money. Another company — HomeExcept — makes non-intrusive sensors to monitor activity in a house, for example, to keep an eye on an elderly person who lives alone. The systems are less intrusive than cameras. They will monitor, for example, whether someone is in the bathroom too long, someone has fallen, or they haven’t been to the kitchen for a few days. “This company received multiple awards, including one from the American Association of Retired Persons,” Lingras says. In this case, his research group is building the app and collecting and analyzing the data the company collects. His program — Saint Mary’s University’s M.Sc. in computing and data analytics — attracts a large number of students with experience in global IT and the students do most of the R&D work. The program has dual advantages. First, it gives students real-world experience to help develop them into highly qualified personnel (HQP), and second, it supports startups from a variety of different fields. Those startups in turn help stimulate the local and national economies. “These startups don’t need this expertise year-round, necessarily, so the availability of it on demand works for them, and it provides cutting-edge experience for our students,” Lingras says. The team uses ACENET’s services to simulate the conditions they need and also to analyze the data their clients are collecting.