deploying AI

Traffic calming measures, cycle paths, and other road safety interventions aim to save lives and promote active transportation — but to understand their impact, we must be able to compare before and after a change was made. That can be challenging because all too often, city records are inaccurate, incomplete, and out of date.

The Data Sciences Institute Catalyst Grant awarded to Professors Brice Batomen Kuimi, Dalla Lana School of Public Health and Marianne Hatzopoulou, Department of Civil and Mineral Engineering, Faculty of Applied Science and Engineering, aims to build a dataset identifying the implementation of traffic calming interventions that can be used for evaluation studies that look at the impacts of these changes. The DSI seed grant awarded to this interdisciplinary team is enabling ongoing collaboration with the City of Toronto and the development of an algorithm that can support the use of Google Street View images to identify where and when changes have occurred.

With hundreds of thousands of images to comb through, manually going through images to look at each street over time to identify when and where traffic calming interventions were implemented is challenging. So, the Eye on the Street team trained a machine learning algorithm to look at images of the same segment from one year to the next to identify when something was implemented.

The team started out with existing techniques from the literature but found that they were affected by data leakage. This refers to connections between the part of the data used for training and the part of the data used for testing that create the impression that a model is working well — but only because it’s repeating what it was trained on.

“In our case,” Prof. Batomen Kuimi explains, “because we have multiple images from the same location, you may end up having an image of the road segment in 2010 in the training and an image of 2017 in the testing. So yes, over the years things might change, but it’s basically the same image. As soon as we made sure that if an image from one location was in the training, no other image from the same location, even in another period, should be in the testing, the result was pretty bad. So we have had to do a lot of work to find other techniques.”

The DSI funding has enabled the development of a new algorithm that tackles this data challenge in a new way, narrowing down the vast number of images to an amount that is manageable for a human to check. For the Eye on the Street team, this means that the algorithm can take Toronto’s 12,000 road segments, over 10 years — more than 120,000 images — and reduce that to 5,000 images where there is a high probability of having an intervention present.

This technique can also be applied in other scenarios. With the approach now described in a published paper and the accompanying code available on GitHub, other researchers are interested in exploring its use for different types of interventions and exposures in the built environment, as well as for impact evaluations with outcomes such as noise and air pollution, where it is essential to know when and where the intervention was implemented.

Prof. Batomen Kuimi says that the algorithm can be especially helpful in cities. “The official records are not always accurate. Sometimes the year of installation in the official documents can be off by one or two years. And depending on what you are studying, it can be really problematic.”

The DSI funding enabled the team to further collaborate with the City of Toronto. As part of the training stage, where images are annotated to say whether or not traffic calming features are present, the team got input from City of Toronto and Transportation Services on how to classify images. The City of Toronto maintains maps of traffic calming measures through the Vision Zero initiative, so the team has been able to compare the model’s findings to the city records. When they compared it to the 2023 vision-zero map they were given at the start of the project, they had found a lot more that were missing from the map. But this spring, the city published a new Vision Zero map, and comparing those shows very good agreement, especially for more recent interventions.

“Toronto was already a good student in terms of keeping track on what’s going on compared to other cities. In other cities in Canada, it would be very helpful to use this type of technique.”

Gary Bader, DSI Associate Director, Research and Software adds, “DSI seed funding supported this project to solve an impactful data science challenge. It will be exciting to see its applications in road safety and its potential for helping us understand and address how a city’s built environment affects people’s lives.”

Applications are now open for the 2026 Catalyst Grants.

When a person experiencing hearing difficulties visits a hearing clinic, the assessment may involve listening to simple tones, words, or sentences and repeating them back exactly. These kinds of tests are useful, but they do not fully capture how hearing works in everyday life.

Much of the speech we listen to is continuous — conversations or stories — and we usually remember the meaning rather than the exact wording. Tests that use more naturalistic speech materials can provide a better picture of real-world comprehension difficulties but generating these materials and scoring a person’s responses is challenging and time-consuming without automation.

The Data Sciences Institute Catalyst Grant awarded to Björn Herrmann, a psychologist and cognitive neuroscientist at Baycrest’s Rotman Research Institute and Karen Gordon, an audiologist in the Hospital for Sick Children’s Department of Otolaryngology aims to improve hearing assessment processes through the use of AI. The DSI seed grant awarded to this interdisciplinary team is enabling new applications of large language models (LLMs) to generate naturalistic speech materials for comprehension testing, and to create tools to automate scoring of how well the participant understood what was said.

LLMs can compare the semantic meaning of the text that was played to a participant or patient and the text of what they recalled, assigning a higher recall score when the meaning of the recalled response more closely matches the original. “We can use large language models to identify meaningful units of speech and automatically score how much a listener understood them,” Dr. Herrmann explains. “LLMs can actually be more consistent than if we asked humans to do that kind of task.”

Importantly, LLMs also enable testing participants in their first language, removing the added language-processing effort required of non-native speakers. Speech materials can be generated in the participant’s first language, and their responses can be scored in that same language using a standardized automated approach. This helps ensure that results are comparable across languages, including with English-language assessments.

 “With modern AI-based voices, we can use the same voice across many languages. So, it’s as if one voice actor could record speech materials in more than 100 languages, and a multilingual scorer evaluates comprehension consistently across them,” says Dr. Herrmann.

With the support of the DSI seed funding, the team is collecting data from older adults with hearing aids, who come in and listen to speech in background noise as would be done at a hearing clinic. They do this with and without hearing aids to see if the automated scoring approaches that the project has developed can help tell whether hearing aids help with speech comprehension or not. This will be the first step towards a clinically viable approach.

The team is also working with hearing aid companies. Manufacturers are interested in measuring how well their hearing aids work with more natural stimuli to further improve their products. The team’s findings suggest that hearing aids have more of a benefit with verbatim recognition than with the naturalistic story listening, which is consistent with many people’s experience in their everyday life. Hearing aid manufacturers want to directly test verbatim recall versus story-based recall, using these automated tools, to develop better metrics to evaluate hearing aids and how well they work.

“This is an exciting application of AI and data science to address an important health challenge. It’s amazing to see how this innovative work is attracting industry partnerships and moving towards having a clinical benefit,” says Gary Bader, DSI Associate Director, Research and Software.

Applications are now open for the 2026 Catalyst Grants.

As the AI boom sends tech firms scrambling for more data to improve their models and puts a premium on the companies that own it, data scientists say businesses and governments need to understand that the technology won’t be useful or accurate without their work.

“Data-driven discovery is becoming so impactful,” said Lisa Strug, a professor in the University of Toronto’s statistical and computer science departments and a senior scientist at SickKids [and Director of the Data Sciences Institute]. As businesses in many sectors and research fields like microbiology, astronomy and the social sciences all adopt AI, she says, they’ll need new data science methods and more practitioners.

Read the full article from The Logic with a free account

• Access through University of Toronto Libraries
• Access through Toronto Public Library

Lisa Strug, Director of the Data Sciences Institute, is a Professor in the Departments of Statistical Sciences, Computer Science and the Division of Biostatistics, and the Director of CANSSI Ontario, at the University of Toronto. She is also the Associate Director of The Centre for Applied Genomics and a Senior Scientist at The Hospital for Sick Children.

The Canadian federal government’s Sovereign AI Compute Strategy commits billions to building and maintaining high-performance computing (HPC) domestically and ushers us towards a pivotal moment in our approach to artificial intelligence.

On the one hand, investing in HPC is overdue and welcome: our capacity for compute is a critical component of modern data-driven research, innovation in industry, and global competitiveness with G7 peers. On the other hand, compute alone cannot sustain Canada’s pre-eminence in AI innovation. Expanding our GPU racks may bring more HPC under Canadian control, but it does not solve equally pressing problems: stagnation in productivity, low AI adoption among firms, outdated public-sector digital infrastructure, uneven and unequitable access, and a lack of high-quality, well-linked, trustworthy data.

A myopic focus on hardware that insists the bottleneck for Canadian initiative is merely compute risks obfuscating the reality that multiple international bodies like the Organization for Economic Co-operation and Development (OECD), International Monetary Fund (IMF), and UNESCO have emphasized: AI performance and national competitiveness depend far more on human expertise in the data sciences and data quality than on raw compute. Canada has this human capital, and at the Data Sciences Institute (DSI) at the University of Toronto, we wonder why the strategy does not do more to capitalize on Canada’s existing strengths.

AI is a data science, but also it cannot exist without the data sciences, which provides the prerequisite know-how to collect, prepare and link the data that trains AI models and the skills to measure uncertainty and bias in the complex, unstructured datasets that AI is used to parse. Canada’s true strength is in its extraordinary reservoir of this kind of data science expertise across multiple fields: experts who know how to collect, clean, link, govern, interpret, and generate derived and missing variables from data.

These experts work in health, public policy, environment, social science, energy, Indigenous governance, academia, and industry. They build (synthetic-)data pipelines, design privacy-preserving systems, operationalize equity frameworks, extract insights, and shape the governance of data flows. At the DSI, we believe that this expertise is paramount to effective AI adoption and that individuals with data sense and skills need to be positioned across sectors for AI uptake to matter. We support and deliver training programs for both graduate students and prerequisite-agnostic reskilling programs for the public at large precisely because we need to increase data science literacy to make AI useful.

Across international benchmarks, Canada ranks near the top in AI-related human capital. According to Deloitte, Canada has the highest AI talent concentration in the G7 and that its AI workforce has grown faster than that of peer countries. The Canadian Institute for Advanced Research (CIFAR) asserts that the Pan-Canadian AI Strategy has helped build “one of the fastest-growing and most skilled AI communities worldwide.”

Canada’s public sector also hosts world-leading data-linkage capacity (e.g., ICES, Statistics Canada linkage environments, and Health Data Research Network Canada). This ecosystem is broader than “AI engineers”: it includes data engineers, statisticians, stewards, auditors, evaluators, and domain experts who make AI usable, trustworthy, and socially valuable.

Given these facts, we find it unhelpful that data scientists are often elided within the compute-first hyperscaler agenda that defines Canada’s strategy. While “data scientists” are mentioned, this strategy mostly seems to advance a much narrower definition of “talent”: deep-learning researchers, model engineers, and infrastructure specialists who build AI models. This group constitutes a tiny slice of the actual AI workforce. The Sovereign AI Compute Strategy only faintly gestures towards “data scientists,” despite OECD research that demonstrates that the highest productivity gains come from data-rich sectors and require cross-disciplinary data expertise.

To be clear: both kinds of expertise are critical, but Canada already has a deep backbench of data scientists, including its statistical sciences community, who should be at the forefront of these conversations, and not reductively relegated to the last item in a comma-separated list of “talent.” Compute investment is necessary, no question, but Canada must reassert leadership in the places where it has a genuine global edge: training, stewarding, and empowering data-science talent, and building the data ecosystems that make compute useful. Our policies should focus not only on large-scale model training but on the breadth of data-work that underpins real-world applications of AI that have the potential to benefit Canada’s leadership in AI development.

Narratives that insist that compute is the bottleneck matters for policy design. If talent is defined narrowly, investment follows narrowly, and compute infrastructure risks becoming under-used, unevenly accessed, or dominated by actors already positioned to train large models. The IMF’s 2024 review of AI readiness reinforces this point: the strongest national differentiators are human capital and data governance, not compute alone. In Canada, data constraints remain acute: many firms lack staff to prepare data for AI; public-sector data remains siloed and difficult to access; Indigenous data governance is under-funded despite commitments to sovereignty; and synthetic-data pipelines remain underdeveloped.

The federal government’s 2025 AI strategy correctly identifies research and talent and education and skills as central pillars alongside enabling infrastructure. The compute strategy is a necessary foundation, but it is not, by itself, a competitiveness strategy. The durable solution is expanding Canada’s data-and-AI workforce and institutions: education and upskilling, data readiness, governance capacity, and the applied expertise needed to deploy AI across hospitals, classrooms, municipalities, organizations, and environmental systems.

Compute matters, but it is talent and skills that determine whether that compute can enable broad-based productivity and international competitiveness.